beenull/ladybird
1
0
Fork 0
mirror of https://github.com/LadybirdBrowser/ladybird.git synced 2024-11-22 07:30:19 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 6
ladybird/Userland/Libraries/LibRegex/RegexOptimizer.cpp

939 lines
38 KiB
C++
Raw Normal View History

LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
/*
* Copyright (c) 2021, Ali Mohammad Pur <mpfard@serenityos.org>
*
* SPDX-License-Identifier: BSD-2-Clause
*/
#include <AK/QuickSort.h>
#include <AK/RedBlackTree.h>
#include <AK/Stack.h>
#include <LibRegex/Regex.h>
#include <LibRegex/RegexBytecodeStreamOptimizer.h>
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
#if REGEX_DEBUG
# include <AK/ScopeGuard.h>
# include <AK/ScopeLogger.h>
#endif
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
namespace regex {
using Detail::Block;
template<typename Parser>
void Regex<Parser>::run_optimization_passes()
{
LibRegex: Flatten bytecode before performing optimizations This avoids doing DisjointChunks traversal for every bytecode access, significantly reducing startup time for large regular expressions.
2021-09-29 16:43:58 +00:00
parser_result.bytecode.flatten();
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
// Rewrite fork loops as atomic groups
// e.g. a*b -> (ATOMIC a*)b
LibRegex: Merge alternations based on blocks and not instructions The instructions can have dependencies (e.g. Repeat), so only unify equal blocks instead of consecutive instructions. Fixes #11247. Also adds the minimal test case(s) from that issue.
2021-12-15 05:22:07 +00:00
attempt_rewrite_loops_as_atomic_groups(split_basic_blocks(parser_result.bytecode));
LibRegex: Avoid excessive Vector copy when compiling regexps Previously we would've copied the bytecode instead of moving the chunks around, use the fancy new DisjointChunks<T> abstraction to make that happen automagically. This decreases vector copies and uses of memmove() by nearly 10x :^)
2021-09-13 19:15:22 +00:00
parser_result.bytecode.flatten();
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
}
template<typename Parser>
LibRegex: Merge alternations based on blocks and not instructions The instructions can have dependencies (e.g. Repeat), so only unify equal blocks instead of consecutive instructions. Fixes #11247. Also adds the minimal test case(s) from that issue.
2021-12-15 05:22:07 +00:00
typename Regex<Parser>::BasicBlockList Regex<Parser>::split_basic_blocks(ByteCode const& bytecode)
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
{
BasicBlockList block_boundaries;
size_t end_of_last_block = 0;
LibRegex: Make append_alternation() significantly faster ...by flattening the underlying bytecode chunks first. Also avoid calling DisjointChunks::size() inside a loop. This is a very significant improvement in performance, making the compilation of a large regex with lots of alternatives take only ~100ms instead of many minutes (I ran out of patience waiting for it) :^)
2021-12-21 14:38:15 +00:00
auto bytecode_size = bytecode.size();
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
MatchState state;
state.instruction_position = 0;
auto check_jump = [&]<typename T>(OpCode const& opcode) {
auto& op = static_cast<T const&>(opcode);
ssize_t jump_offset = op.size() + op.offset();
if (jump_offset >= 0) {
block_boundaries.append({ end_of_last_block, state.instruction_position });
end_of_last_block = state.instruction_position + opcode.size();
} else {
// This op jumps back, see if that's within this "block".
if (jump_offset + state.instruction_position > end_of_last_block) {
// Split the block!
block_boundaries.append({ end_of_last_block, jump_offset + state.instruction_position });
block_boundaries.append({ jump_offset + state.instruction_position, state.instruction_position });
end_of_last_block = state.instruction_position + opcode.size();
} else {
// Nope, it's just a jump to another block
block_boundaries.append({ end_of_last_block, state.instruction_position });
end_of_last_block = state.instruction_position + opcode.size();
}
}
};
for (;;) {
auto& opcode = bytecode.get_opcode(state);
switch (opcode.opcode_id()) {
case OpCodeId::Jump:
check_jump.template operator()<OpCode_Jump>(opcode);
break;
case OpCodeId::JumpNonEmpty:
check_jump.template operator()<OpCode_JumpNonEmpty>(opcode);
break;
case OpCodeId::ForkJump:
check_jump.template operator()<OpCode_ForkJump>(opcode);
break;
case OpCodeId::ForkStay:
check_jump.template operator()<OpCode_ForkStay>(opcode);
break;
case OpCodeId::FailForks:
block_boundaries.append({ end_of_last_block, state.instruction_position });
end_of_last_block = state.instruction_position + opcode.size();
break;
case OpCodeId::Repeat: {
// Repeat produces two blocks, one containing its repeated expr, and one after that.
auto repeat_start = state.instruction_position - static_cast<OpCode_Repeat const&>(opcode).offset();
if (repeat_start > end_of_last_block)
block_boundaries.append({ end_of_last_block, repeat_start });
block_boundaries.append({ repeat_start, state.instruction_position });
end_of_last_block = state.instruction_position + opcode.size();
break;
}
default:
break;
}
auto next_ip = state.instruction_position + opcode.size();
LibRegex: Make append_alternation() significantly faster ...by flattening the underlying bytecode chunks first. Also avoid calling DisjointChunks::size() inside a loop. This is a very significant improvement in performance, making the compilation of a large regex with lots of alternatives take only ~100ms instead of many minutes (I ran out of patience waiting for it) :^)
2021-12-21 14:38:15 +00:00
if (next_ip < bytecode_size)
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
state.instruction_position = next_ip;
else
break;
}
LibRegex: Make append_alternation() significantly faster ...by flattening the underlying bytecode chunks first. Also avoid calling DisjointChunks::size() inside a loop. This is a very significant improvement in performance, making the compilation of a large regex with lots of alternatives take only ~100ms instead of many minutes (I ran out of patience waiting for it) :^)
2021-12-21 14:38:15 +00:00
if (end_of_last_block < bytecode_size)
block_boundaries.append({ end_of_last_block, bytecode_size });
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
quick_sort(block_boundaries, [](auto& a, auto& b) { return a.start < b.start; });
return block_boundaries;
}
LibRegex: Fully interpret the Compare Op when looking for overlaps We had a really naive and simplistic implementation, which lead to various issues where the optimiser incorrectly rewrote the regex to use atomic groups; this commit fixes that.
2022-07-04 16:41:15 +00:00
static bool has_overlap(Vector<CompareTypeAndValuePair> const& lhs, Vector<CompareTypeAndValuePair> const& rhs)
{
// We have to fully interpret the two sequences to determine if they overlap (that is, keep track of inversion state and what ranges they cover).
bool inverse { false };
bool temporary_inverse { false };
bool reset_temporary_inverse { false };
auto current_lhs_inversion_state = [&]() -> bool { return temporary_inverse ^ inverse; };
RedBlackTree<u32, u32> lhs_ranges;
RedBlackTree<u32, u32> lhs_negated_ranges;
HashTable<CharClass> lhs_char_classes;
HashTable<CharClass> lhs_negated_char_classes;
auto range_contains = [&]<typename T>(T& value) -> bool {
u32 start;
u32 end;
if constexpr (IsSame<T, CharRange>) {
start = value.from;
end = value.to;
} else {
start = value;
end = value;
}
auto* max = lhs_ranges.find_smallest_not_below(start);
return max && *max <= end;
};
auto char_class_contains = [&](CharClass const& value) -> bool {
if (lhs_char_classes.contains(value))
return true;
if (lhs_negated_char_classes.contains(value))
return false;
// This char class might match something in the ranges we have, and checking that is far too expensive, so just bail out.
return true;
};
for (auto const& pair : lhs) {
if (reset_temporary_inverse) {
reset_temporary_inverse = false;
temporary_inverse = false;
} else {
reset_temporary_inverse = true;
}
switch (pair.type) {
case CharacterCompareType::Inverse:
inverse = !inverse;
break;
case CharacterCompareType::TemporaryInverse:
temporary_inverse = !temporary_inverse;
break;
case CharacterCompareType::AnyChar:
// Special case: if not inverted, AnyChar is always in the range.
if (!current_lhs_inversion_state())
return true;
break;
case CharacterCompareType::Char:
if (!current_lhs_inversion_state())
lhs_ranges.insert(pair.value, pair.value);
else
lhs_negated_ranges.insert(pair.value, pair.value);
break;
case CharacterCompareType::String:
// FIXME: We just need to look at the last character of this string, but we only have the first character here.
// Just bail out to avoid false positives.
return true;
case CharacterCompareType::CharClass:
if (!current_lhs_inversion_state())
lhs_char_classes.set(static_cast<CharClass>(pair.value));
else
lhs_negated_char_classes.set(static_cast<CharClass>(pair.value));
break;
case CharacterCompareType::CharRange: {
LibRegex: Use proper CharRange constructor instead of bit_casting Otherwise the range order would be inverted.
2022-07-05 00:26:21 +00:00
auto range = CharRange(pair.value);
LibRegex: Fully interpret the Compare Op when looking for overlaps We had a really naive and simplistic implementation, which lead to various issues where the optimiser incorrectly rewrote the regex to use atomic groups; this commit fixes that.
2022-07-04 16:41:15 +00:00
if (!current_lhs_inversion_state())
lhs_ranges.insert(range.from, range.to);
else
lhs_negated_ranges.insert(range.from, range.to);
break;
}
case CharacterCompareType::LookupTable:
// We've transformed this into a series of ranges in flat_compares(), so bail out if we see it.
return true;
case CharacterCompareType::Reference:
// We've handled this before coming here.
break;
case CharacterCompareType::Property:
case CharacterCompareType::GeneralCategory:
case CharacterCompareType::Script:
case CharacterCompareType::ScriptExtension:
// FIXME: These are too difficult to handle, so bail out.
return true;
case CharacterCompareType::Undefined:
case CharacterCompareType::RangeExpressionDummy:
// These do not occur in valid bytecode.
VERIFY_NOT_REACHED();
}
}
if constexpr (REGEX_DEBUG) {
dbgln("lhs ranges:");
for (auto it = lhs_ranges.begin(); it != lhs_ranges.end(); ++it)
dbgln(" {}..{}", it.key(), *it);
dbgln("lhs negated ranges:");
for (auto it = lhs_negated_ranges.begin(); it != lhs_negated_ranges.end(); ++it)
dbgln(" {}..{}", it.key(), *it);
}
for (auto const& pair : rhs) {
if (reset_temporary_inverse) {
reset_temporary_inverse = false;
temporary_inverse = false;
} else {
reset_temporary_inverse = true;
}
dbgln_if(REGEX_DEBUG, "check {} ({})...", character_compare_type_name(pair.type), pair.value);
switch (pair.type) {
case CharacterCompareType::Inverse:
inverse = !inverse;
break;
case CharacterCompareType::TemporaryInverse:
temporary_inverse = !temporary_inverse;
break;
case CharacterCompareType::AnyChar:
// Special case: if not inverted, AnyChar is always in the range.
if (!current_lhs_inversion_state())
return true;
break;
case CharacterCompareType::Char:
if (!current_lhs_inversion_state() && range_contains(pair.value))
return true;
break;
case CharacterCompareType::String:
// FIXME: We just need to look at the last character of this string, but we only have the first character here.
// Just bail out to avoid false positives.
return true;
case CharacterCompareType::CharClass:
if (!current_lhs_inversion_state() && char_class_contains(static_cast<CharClass>(pair.value)))
return true;
break;
case CharacterCompareType::CharRange: {
LibRegex: Use proper CharRange constructor instead of bit_casting Otherwise the range order would be inverted.
2022-07-05 00:26:21 +00:00
auto range = CharRange(pair.value);
LibRegex: Fully interpret the Compare Op when looking for overlaps We had a really naive and simplistic implementation, which lead to various issues where the optimiser incorrectly rewrote the regex to use atomic groups; this commit fixes that.
2022-07-04 16:41:15 +00:00
if (!current_lhs_inversion_state() && range_contains(range))
return true;
break;
}
case CharacterCompareType::LookupTable:
// We've transformed this into a series of ranges in flat_compares(), so bail out if we see it.
return true;
case CharacterCompareType::Reference:
// We've handled this before coming here.
break;
case CharacterCompareType::Property:
case CharacterCompareType::GeneralCategory:
case CharacterCompareType::Script:
case CharacterCompareType::ScriptExtension:
// FIXME: These are too difficult to handle, so bail out.
return true;
case CharacterCompareType::Undefined:
case CharacterCompareType::RangeExpressionDummy:
// These do not occur in valid bytecode.
VERIFY_NOT_REACHED();
}
}
return false;
}
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
enum class AtomicRewritePreconditionResult {
SatisfiedWithProperHeader,
SatisfiedWithEmptyHeader,
NotSatisfied,
};
static AtomicRewritePreconditionResult block_satisfies_atomic_rewrite_precondition(ByteCode const& bytecode, Block const& repeated_block, Block const& following_block)
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
{
Vector<Vector<CompareTypeAndValuePair>> repeated_values;
LibRegex: Make the optimizer understand references and capture groups Otherwise the fork in patterns like `(1+)\1` would be (incorrectly) optimized away.
2021-09-15 10:01:55 +00:00
HashTable<size_t> active_capture_groups;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
MatchState state;
for (state.instruction_position = repeated_block.start; state.instruction_position < repeated_block.end;) {
auto& opcode = bytecode.get_opcode(state);
switch (opcode.opcode_id()) {
case OpCodeId::Compare: {
auto compares = static_cast<OpCode_Compare const&>(opcode).flat_compares();
if (repeated_values.is_empty() && any_of(compares, [](auto& compare) { return compare.type == CharacterCompareType::AnyChar; }))
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
return AtomicRewritePreconditionResult::NotSatisfied;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
repeated_values.append(move(compares));
break;
}
case OpCodeId::CheckBegin:
case OpCodeId::CheckEnd:
if (repeated_values.is_empty())
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
return AtomicRewritePreconditionResult::SatisfiedWithProperHeader;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
break;
case OpCodeId::CheckBoundary:
// FIXME: What should we do with these? for now, let's fail.
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
return AtomicRewritePreconditionResult::NotSatisfied;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
case OpCodeId::Restore:
case OpCodeId::GoBack:
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
return AtomicRewritePreconditionResult::NotSatisfied;
LibRegex: Make the optimizer understand references and capture groups Otherwise the fork in patterns like `(1+)\1` would be (incorrectly) optimized away.
2021-09-15 10:01:55 +00:00
case OpCodeId::SaveRightCaptureGroup:
active_capture_groups.set(static_cast<OpCode_SaveRightCaptureGroup const&>(opcode).id());
break;
case OpCodeId::SaveLeftCaptureGroup:
active_capture_groups.set(static_cast<OpCode_SaveLeftCaptureGroup const&>(opcode).id());
break;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
default:
break;
}
state.instruction_position += opcode.size();
}
dbgln_if(REGEX_DEBUG, "Found {} entries in reference", repeated_values.size());
LibRegex: Make the optimizer understand references and capture groups Otherwise the fork in patterns like `(1+)\1` would be (incorrectly) optimized away.
2021-09-15 10:01:55 +00:00
dbgln_if(REGEX_DEBUG, "Found {} active capture groups", active_capture_groups.size());
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
bool following_block_has_at_least_one_compare = false;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
// Find the first compare in the following block, it must NOT match any of the values in `repeated_values'.
for (state.instruction_position = following_block.start; state.instruction_position < following_block.end;) {
auto& opcode = bytecode.get_opcode(state);
switch (opcode.opcode_id()) {
LibRegex: Make the optimizer understand references and capture groups Otherwise the fork in patterns like `(1+)\1` would be (incorrectly) optimized away.
2021-09-15 10:01:55 +00:00
// Note: These have to exist since we're effectively repeating the following block as well
case OpCodeId::SaveRightCaptureGroup:
active_capture_groups.set(static_cast<OpCode_SaveRightCaptureGroup const&>(opcode).id());
break;
case OpCodeId::SaveLeftCaptureGroup:
active_capture_groups.set(static_cast<OpCode_SaveLeftCaptureGroup const&>(opcode).id());
break;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
case OpCodeId::Compare: {
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
following_block_has_at_least_one_compare = true;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
// We found a compare, let's see what it has.
auto compares = static_cast<OpCode_Compare const&>(opcode).flat_compares();
if (compares.is_empty())
break;
LibRegex: Make the optimizer understand references and capture groups Otherwise the fork in patterns like `(1+)\1` would be (incorrectly) optimized away.
2021-09-15 10:01:55 +00:00
if (any_of(compares, [&](auto& compare) {
return compare.type == CharacterCompareType::AnyChar
|| (compare.type == CharacterCompareType::Reference && active_capture_groups.contains(compare.value));
}))
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
return AtomicRewritePreconditionResult::NotSatisfied;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
LibRegex: Fully interpret the Compare Op when looking for overlaps We had a really naive and simplistic implementation, which lead to various issues where the optimiser incorrectly rewrote the regex to use atomic groups; this commit fixes that.
2022-07-04 16:41:15 +00:00
if (any_of(repeated_values, [&](auto& repeated_value) { return has_overlap(compares, repeated_value); }))
return AtomicRewritePreconditionResult::NotSatisfied;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
return AtomicRewritePreconditionResult::SatisfiedWithProperHeader;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
}
case OpCodeId::CheckBegin:
case OpCodeId::CheckEnd:
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
return AtomicRewritePreconditionResult::SatisfiedWithProperHeader; // Nothing can match the end!
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
case OpCodeId::CheckBoundary:
// FIXME: What should we do with these? For now, consider them a failure.
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
return AtomicRewritePreconditionResult::NotSatisfied;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
default:
break;
}
state.instruction_position += opcode.size();
}
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
if (following_block_has_at_least_one_compare)
return AtomicRewritePreconditionResult::SatisfiedWithProperHeader;
return AtomicRewritePreconditionResult::SatisfiedWithEmptyHeader;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
}
template<typename Parser>
void Regex<Parser>::attempt_rewrite_loops_as_atomic_groups(BasicBlockList const& basic_blocks)
{
auto& bytecode = parser_result.bytecode;
if constexpr (REGEX_DEBUG) {
RegexDebug dbg;
dbg.print_bytecode(*this);
LibRegex: Add some implied auto qualifiers
2021-12-21 16:38:51 +00:00
for (auto const& block : basic_blocks)
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
dbgln("block from {} to {}", block.start, block.end);
}
// A pattern such as:
// bb0 | RE0
// | ForkX bb0
// -------------------------
// bb1 | RE1
// can be rewritten as:
// -------------------------
// bb0 | RE0
// | ForkReplaceX bb0
// -------------------------
// bb1 | RE1
// provided that first(RE1) not-in end(RE0), which is to say
// that RE1 cannot start with whatever RE0 has matched (ever).
//
// Alternatively, a second form of this pattern can also occur:
// bb0 | *
// | ForkX bb2
// ------------------------
// bb1 | RE0
// | Jump bb0
// ------------------------
// bb2 | RE1
// which can be transformed (with the same preconditions) to:
// bb0 | *
// | ForkReplaceX bb2
// ------------------------
// bb1 | RE0
// | Jump bb0
// ------------------------
// bb2 | RE1
enum class AlternateForm {
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
DirectLoopWithoutHeader, // loop without proper header, a block forking to itself. i.e. the first form.
DirectLoopWithoutHeaderAndEmptyFollow, // loop without proper header, a block forking to itself. i.e. the first form but with RE1 being empty.
DirectLoopWithHeader, // loop with proper header, i.e. the second form.
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
};
struct CandidateBlock {
Block forking_block;
Optional<Block> new_target_block;
AlternateForm form;
};
Vector<CandidateBlock> candidate_blocks;
auto is_an_eligible_jump = [](OpCode const& opcode, size_t ip, size_t block_start, AlternateForm alternate_form) {
switch (opcode.opcode_id()) {
case OpCodeId::JumpNonEmpty: {
LibRegex: Add some implied auto qualifiers
2021-12-21 16:38:51 +00:00
auto const& op = static_cast<OpCode_JumpNonEmpty const&>(opcode);
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
auto form = op.form();
if (form != OpCodeId::Jump && alternate_form == AlternateForm::DirectLoopWithHeader)
return false;
if (form != OpCodeId::ForkJump && form != OpCodeId::ForkStay && alternate_form == AlternateForm::DirectLoopWithoutHeader)
return false;
return op.offset() + ip + opcode.size() == block_start;
}
case OpCodeId::ForkJump:
if (alternate_form == AlternateForm::DirectLoopWithHeader)
return false;
return static_cast<OpCode_ForkJump const&>(opcode).offset() + ip + opcode.size() == block_start;
case OpCodeId::ForkStay:
if (alternate_form == AlternateForm::DirectLoopWithHeader)
return false;
return static_cast<OpCode_ForkStay const&>(opcode).offset() + ip + opcode.size() == block_start;
case OpCodeId::Jump:
// Infinite loop does *not* produce forks.
if (alternate_form == AlternateForm::DirectLoopWithoutHeader)
return false;
if (alternate_form == AlternateForm::DirectLoopWithHeader)
return static_cast<OpCode_Jump const&>(opcode).offset() + ip + opcode.size() == block_start;
VERIFY_NOT_REACHED();
default:
return false;
}
};
for (size_t i = 0; i < basic_blocks.size(); ++i) {
auto forking_block = basic_blocks[i];
Optional<Block> fork_fallback_block;
if (i + 1 < basic_blocks.size())
fork_fallback_block = basic_blocks[i + 1];
MatchState state;
// Check if the last instruction in this block is a jump to the block itself:
{
state.instruction_position = forking_block.end;
auto& opcode = bytecode.get_opcode(state);
if (is_an_eligible_jump(opcode, state.instruction_position, forking_block.start, AlternateForm::DirectLoopWithoutHeader)) {
// We've found RE0 (and RE1 is just the following block, if any), let's see if the precondition applies.
// if RE1 is empty, there's no first(RE1), so this is an automatic pass.
if (!fork_fallback_block.has_value() || fork_fallback_block->end == fork_fallback_block->start) {
candidate_blocks.append({ forking_block, fork_fallback_block, AlternateForm::DirectLoopWithoutHeader });
break;
}
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
auto precondition = block_satisfies_atomic_rewrite_precondition(bytecode, forking_block, *fork_fallback_block);
if (precondition == AtomicRewritePreconditionResult::SatisfiedWithProperHeader) {
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
candidate_blocks.append({ forking_block, fork_fallback_block, AlternateForm::DirectLoopWithoutHeader });
break;
}
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
if (precondition == AtomicRewritePreconditionResult::SatisfiedWithEmptyHeader) {
candidate_blocks.append({ forking_block, fork_fallback_block, AlternateForm::DirectLoopWithoutHeaderAndEmptyFollow });
break;
}
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
}
}
// Check if the last instruction in the last block is a direct jump to this block
if (fork_fallback_block.has_value()) {
state.instruction_position = fork_fallback_block->end;
auto& opcode = bytecode.get_opcode(state);
if (is_an_eligible_jump(opcode, state.instruction_position, forking_block.start, AlternateForm::DirectLoopWithHeader)) {
// We've found bb1 and bb0, let's just make sure that bb0 forks to bb2.
state.instruction_position = forking_block.end;
auto& opcode = bytecode.get_opcode(state);
if (opcode.opcode_id() == OpCodeId::ForkJump || opcode.opcode_id() == OpCodeId::ForkStay) {
Optional<Block> block_following_fork_fallback;
if (i + 2 < basic_blocks.size())
block_following_fork_fallback = basic_blocks[i + 2];
LibRegex: Avoid rewriting `a+` as `a*` as part of atomic rewriting The initial `ForkStay` is only needed if the looping block has a following block, if there's no following block or the following block does not attempt to match anything, we should not insert the ForkStay, otherwise we would be rewriting `a+` as `a*` by allowing the 'end' to be executed. Fixes #10952.
2021-11-18 04:30:29 +00:00
if (!block_following_fork_fallback.has_value()
|| block_satisfies_atomic_rewrite_precondition(bytecode, *fork_fallback_block, *block_following_fork_fallback) != AtomicRewritePreconditionResult::NotSatisfied) {
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
candidate_blocks.append({ forking_block, {}, AlternateForm::DirectLoopWithHeader });
break;
}
}
}
}
}
dbgln_if(REGEX_DEBUG, "Found {} candidate blocks", candidate_blocks.size());
if (candidate_blocks.is_empty()) {
dbgln_if(REGEX_DEBUG, "Failed to find anything for {}", pattern_value);
return;
}
RedBlackTree<size_t, size_t> needed_patches;
// Reverse the blocks, so we can patch the bytecode without messing with the latter patches.
quick_sort(candidate_blocks, [](auto& a, auto& b) { return b.forking_block.start > a.forking_block.start; });
for (auto& candidate : candidate_blocks) {
// Note that both forms share a ForkReplace patch in forking_block.
// Patch the ForkX in forking_block to be a ForkReplaceX instead.
auto& opcode_id = bytecode[candidate.forking_block.end];
if (opcode_id == (ByteCodeValueType)OpCodeId::ForkStay) {
opcode_id = (ByteCodeValueType)OpCodeId::ForkReplaceStay;
} else if (opcode_id == (ByteCodeValueType)OpCodeId::ForkJump) {
opcode_id = (ByteCodeValueType)OpCodeId::ForkReplaceJump;
} else if (opcode_id == (ByteCodeValueType)OpCodeId::JumpNonEmpty) {
auto& jump_opcode_id = bytecode[candidate.forking_block.end + 3];
if (jump_opcode_id == (ByteCodeValueType)OpCodeId::ForkStay)
jump_opcode_id = (ByteCodeValueType)OpCodeId::ForkReplaceStay;
else if (jump_opcode_id == (ByteCodeValueType)OpCodeId::ForkJump)
jump_opcode_id = (ByteCodeValueType)OpCodeId::ForkReplaceJump;
else
VERIFY_NOT_REACHED();
} else {
VERIFY_NOT_REACHED();
}
}
if (!needed_patches.is_empty()) {
MatchState state;
LibRegex: Make append_alternation() significantly faster ...by flattening the underlying bytecode chunks first. Also avoid calling DisjointChunks::size() inside a loop. This is a very significant improvement in performance, making the compilation of a large regex with lots of alternatives take only ~100ms instead of many minutes (I ran out of patience waiting for it) :^)
2021-12-21 14:38:15 +00:00
auto bytecode_size = bytecode.size();
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
state.instruction_position = 0;
struct Patch {
ssize_t value;
size_t offset;
bool should_negate { false };
};
for (;;) {
LibRegex: Make append_alternation() significantly faster ...by flattening the underlying bytecode chunks first. Also avoid calling DisjointChunks::size() inside a loop. This is a very significant improvement in performance, making the compilation of a large regex with lots of alternatives take only ~100ms instead of many minutes (I ran out of patience waiting for it) :^)
2021-12-21 14:38:15 +00:00
if (state.instruction_position >= bytecode_size)
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
break;
auto& opcode = bytecode.get_opcode(state);
Stack<Patch, 2> patch_points;
switch (opcode.opcode_id()) {
case OpCodeId::Jump:
patch_points.push({ static_cast<OpCode_Jump const&>(opcode).offset(), state.instruction_position + 1 });
break;
case OpCodeId::JumpNonEmpty:
patch_points.push({ static_cast<OpCode_JumpNonEmpty const&>(opcode).offset(), state.instruction_position + 1 });
patch_points.push({ static_cast<OpCode_JumpNonEmpty const&>(opcode).checkpoint(), state.instruction_position + 2 });
break;
case OpCodeId::ForkJump:
patch_points.push({ static_cast<OpCode_ForkJump const&>(opcode).offset(), state.instruction_position + 1 });
break;
case OpCodeId::ForkStay:
patch_points.push({ static_cast<OpCode_ForkStay const&>(opcode).offset(), state.instruction_position + 1 });
break;
case OpCodeId::Repeat:
patch_points.push({ -(ssize_t) static_cast<OpCode_Repeat const&>(opcode).offset(), state.instruction_position + 1, true });
break;
default:
break;
}
while (!patch_points.is_empty()) {
auto& patch_point = patch_points.top();
auto target_offset = patch_point.value + state.instruction_position + opcode.size();
constexpr auto do_patch = [](auto& patch_it, auto& patch_point, auto& target_offset, auto& bytecode, auto ip) {
if (patch_it.key() == ip)
return;
LibRegex: Correct jump offset to the start of the loop block Previously we were jumping to the new end of the previous block (created by the newly inserted ForkStay), correct the offset to jump to the correct block as shown in the comments. Fixes #12033.
2022-01-21 09:03:49 +00:00
if (patch_point.value < 0 && target_offset <= patch_it.key() && ip > patch_it.key())
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
bytecode[patch_point.offset] += (patch_point.should_negate ? 1 : -1) * (*patch_it);
LibRegex: Correct jump offset to the start of the loop block Previously we were jumping to the new end of the previous block (created by the newly inserted ForkStay), correct the offset to jump to the correct block as shown in the comments. Fixes #12033.
2022-01-21 09:03:49 +00:00
else if (patch_point.value > 0 && target_offset >= patch_it.key() && ip < patch_it.key())
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
bytecode[patch_point.offset] += (patch_point.should_negate ? -1 : 1) * (*patch_it);
};
if (auto patch_it = needed_patches.find_largest_not_above_iterator(target_offset); !patch_it.is_end())
do_patch(patch_it, patch_point, target_offset, bytecode, state.instruction_position);
else if (auto patch_it = needed_patches.find_largest_not_above_iterator(state.instruction_position); !patch_it.is_end())
do_patch(patch_it, patch_point, target_offset, bytecode, state.instruction_position);
patch_points.pop();
}
state.instruction_position += opcode.size();
}
}
if constexpr (REGEX_DEBUG) {
warnln("Transformed to:");
RegexDebug dbg;
dbg.print_bytecode(*this);
}
}
LibRegex: Avoid excessive Vector copy when compiling regexps Previously we would've copied the bytecode instead of moving the chunks around, use the fancy new DisjointChunks<T> abstraction to make that happen automagically. This decreases vector copies and uses of memmove() by nearly 10x :^)
2021-09-13 19:15:22 +00:00
void Optimizer::append_alternation(ByteCode& target, ByteCode&& left, ByteCode&& right)
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
{
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
Array<ByteCode, 2> alternatives;
alternatives[0] = move(left);
alternatives[1] = move(right);
append_alternation(target, alternatives);
}
void Optimizer::append_alternation(ByteCode& target, Span<ByteCode> alternatives)
{
if (alternatives.size() == 0)
return;
if (alternatives.size() == 1)
return target.extend(move(alternatives[0]));
if (all_of(alternatives, [](auto& x) { return x.is_empty(); }))
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
return;
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
for (auto& entry : alternatives)
entry.flatten();
#if REGEX_DEBUG
ScopeLogger<true> log;
warnln("Alternations:");
RegexDebug dbg;
for (auto& entry : alternatives) {
warnln("----------");
dbg.print_bytecode(entry);
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
}
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
ScopeGuard print_at_end {
[&] {
warnln("======================");
RegexDebug dbg;
dbg.print_bytecode(target);
}
};
#endif
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
Vector<Vector<Detail::Block>> basic_blocks;
basic_blocks.ensure_capacity(alternatives.size());
LibRegex: Make append_alternation() significantly faster ...by flattening the underlying bytecode chunks first. Also avoid calling DisjointChunks::size() inside a loop. This is a very significant improvement in performance, making the compilation of a large regex with lots of alternatives take only ~100ms instead of many minutes (I ran out of patience waiting for it) :^)
2021-12-21 14:38:15 +00:00
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
for (auto& entry : alternatives)
basic_blocks.append(Regex<PosixBasicParser>::split_basic_blocks(entry));
LibRegex: Merge alternations based on blocks and not instructions The instructions can have dependencies (e.g. Repeat), so only unify equal blocks instead of consecutive instructions. Fixes #11247. Also adds the minimal test case(s) from that issue.
2021-12-15 05:22:07 +00:00
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
size_t left_skip = 0;
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
size_t shared_block_count = basic_blocks.first().size();
for (auto& entry : basic_blocks)
shared_block_count = min(shared_block_count, entry.size());
LibRegex: Merge alternations based on blocks and not instructions The instructions can have dependencies (e.g. Repeat), so only unify equal blocks instead of consecutive instructions. Fixes #11247. Also adds the minimal test case(s) from that issue.
2021-12-15 05:22:07 +00:00
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
MatchState state;
for (size_t block_index = 0; block_index < shared_block_count; block_index++) {
auto& left_block = basic_blocks.first()[block_index];
auto left_end = block_index + 1 == basic_blocks.first().size() ? left_block.end : basic_blocks.first()[block_index + 1].start;
auto can_continue = true;
for (size_t i = 1; i < alternatives.size(); ++i) {
auto& right_blocks = basic_blocks[i];
auto& right_block = right_blocks[block_index];
auto right_end = block_index + 1 == right_blocks.size() ? right_block.end : right_blocks[block_index + 1].start;
if (left_end - left_block.start != right_end - right_block.start) {
can_continue = false;
break;
}
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
if (alternatives[0].spans().slice(left_block.start, left_end - left_block.start) != alternatives[i].spans().slice(right_block.start, right_end - right_block.start)) {
can_continue = false;
break;
}
}
if (!can_continue)
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
break;
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
size_t i = 0;
for (auto& entry : alternatives) {
auto& blocks = basic_blocks[i];
auto& block = blocks[block_index];
auto end = block_index + 1 == blocks.size() ? block.end : blocks[block_index + 1].start;
state.instruction_position = block.start;
size_t skip = 0;
while (state.instruction_position < end) {
auto& opcode = entry.get_opcode(state);
state.instruction_position += opcode.size();
skip = state.instruction_position;
}
left_skip = min(skip, left_skip);
LibRegex: Only skip full instructions when optimizing alternations It makes no sense to skip half of an instruction, so make sure to skip only full instructions!
2022-02-09 20:13:09 +00:00
}
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
}
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
dbgln_if(REGEX_DEBUG, "Skipping {}/{} bytecode entries from {}", left_skip, 0, alternatives[0].size());
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
LibRegex: Merge alternations based on blocks and not instructions The instructions can have dependencies (e.g. Repeat), so only unify equal blocks instead of consecutive instructions. Fixes #11247. Also adds the minimal test case(s) from that issue.
2021-12-15 05:22:07 +00:00
if (left_skip > 0) {
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
target.extend(alternatives[0].release_slice(basic_blocks.first().first().start, left_skip));
auto first = true;
for (auto& entry : alternatives) {
if (first) {
first = false;
continue;
}
entry = entry.release_slice(left_skip);
}
LibRegex: Avoid excessive Vector copy when compiling regexps Previously we would've copied the bytecode instead of moving the chunks around, use the fancy new DisjointChunks<T> abstraction to make that happen automagically. This decreases vector copies and uses of memmove() by nearly 10x :^)
2021-09-13 19:15:22 +00:00
}
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
if (all_of(alternatives, [](auto& entry) { return entry.is_empty(); }))
return;
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
size_t patch_start = target.size();
for (size_t i = 1; i < alternatives.size(); ++i) {
target.empend(static_cast<ByteCodeValueType>(OpCodeId::ForkJump));
target.empend(0u); // To be filled later.
}
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
size_t size_to_jump = 0;
bool seen_one_empty = false;
for (size_t i = alternatives.size(); i > 0; --i) {
auto& entry = alternatives[i - 1];
if (entry.is_empty()) {
if (seen_one_empty)
continue;
seen_one_empty = true;
}
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
auto is_first = i == 1;
auto instruction_size = entry.size() + (is_first ? 0 : 2); // Jump; -> +2
size_to_jump += instruction_size;
if (!is_first)
target[patch_start + (i - 2) * 2 + 1] = size_to_jump + (alternatives.size() - i) * 2;
dbgln_if(REGEX_DEBUG, "{} size = {}, cum={}", i - 1, instruction_size, size_to_jump);
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
}
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
seen_one_empty = false;
for (size_t i = alternatives.size(); i > 0; --i) {
auto& chunk = alternatives[i - 1];
if (chunk.is_empty()) {
if (seen_one_empty)
continue;
seen_one_empty = true;
}
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
ByteCode* previous_chunk = nullptr;
size_t j = i - 1;
auto seen_one_empty_before = chunk.is_empty();
while (j >= 1) {
--j;
auto& candidate_chunk = alternatives[j];
if (candidate_chunk.is_empty()) {
if (seen_one_empty_before)
continue;
}
previous_chunk = &candidate_chunk;
break;
}
size_to_jump -= chunk.size() + (previous_chunk ? 2 : 0);
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
LibRegex: Make codegen+optimisation for alternatives much faster Just a little thinking outside the box, and we can now parse and optimise a million copies of "a|" chained together in just a second :^)
2022-02-19 23:55:21 +00:00
target.extend(move(chunk));
target.empend(static_cast<ByteCodeValueType>(OpCodeId::Jump));
target.empend(size_to_jump); // Jump to the _END label
}
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
}
LibRegex: Use a match table for character classes Generate a sorted, compressed series of ranges in a match table for character classes, and use a binary search to find the matches. This is about a 3-4x speedup for character class match performance. :^)
2021-10-03 15:31:25 +00:00
enum class LookupTableInsertionOutcome {
Successful,
ReplaceWithAnyChar,
TemporaryInversionNeeded,
PermanentInversionNeeded,
CannotPlaceInTable,
};
static LookupTableInsertionOutcome insert_into_lookup_table(RedBlackTree<ByteCodeValueType, CharRange>& table, CompareTypeAndValuePair pair)
{
switch (pair.type) {
case CharacterCompareType::Inverse:
return LookupTableInsertionOutcome::PermanentInversionNeeded;
case CharacterCompareType::TemporaryInverse:
return LookupTableInsertionOutcome::TemporaryInversionNeeded;
case CharacterCompareType::AnyChar:
return LookupTableInsertionOutcome::ReplaceWithAnyChar;
case CharacterCompareType::CharClass:
return LookupTableInsertionOutcome::CannotPlaceInTable;
case CharacterCompareType::Char:
table.insert(pair.value, { (u32)pair.value, (u32)pair.value });
break;
case CharacterCompareType::CharRange: {
CharRange range { pair.value };
table.insert(range.from, range);
break;
}
case CharacterCompareType::Reference:
case CharacterCompareType::Property:
case CharacterCompareType::GeneralCategory:
case CharacterCompareType::Script:
case CharacterCompareType::ScriptExtension:
return LookupTableInsertionOutcome::CannotPlaceInTable;
case CharacterCompareType::Undefined:
case CharacterCompareType::RangeExpressionDummy:
case CharacterCompareType::String:
case CharacterCompareType::LookupTable:
VERIFY_NOT_REACHED();
}
return LookupTableInsertionOutcome::Successful;
}
void Optimizer::append_character_class(ByteCode& target, Vector<CompareTypeAndValuePair>&& pairs)
{
ByteCode arguments;
size_t argument_count = 0;
if (pairs.size() <= 1) {
for (auto& pair : pairs) {
arguments.append(to_underlying(pair.type));
if (pair.type != CharacterCompareType::AnyChar && pair.type != CharacterCompareType::TemporaryInverse && pair.type != CharacterCompareType::Inverse)
arguments.append(pair.value);
++argument_count;
}
} else {
RedBlackTree<ByteCodeValueType, CharRange> table;
RedBlackTree<ByteCodeValueType, CharRange> inverted_table;
auto* current_table = &table;
auto* current_inverted_table = &inverted_table;
bool invert_for_next_iteration = false;
bool is_currently_inverted = false;
LibRegex: Flush compare tables before entering a permanent inverse state
2022-07-10 07:12:31 +00:00
auto flush_tables = [&] {
auto append_table = [&](auto& table) {
++argument_count;
arguments.append(to_underlying(CharacterCompareType::LookupTable));
auto size_index = arguments.size();
arguments.append(0);
Optional<CharRange> active_range;
size_t range_count = 0;
for (auto& range : table) {
if (!active_range.has_value()) {
active_range = range;
continue;
}
if (range.from <= active_range->to + 1 && range.to + 1 >= active_range->from) {
active_range = CharRange { min(range.from, active_range->from), max(range.to, active_range->to) };
} else {
++range_count;
arguments.append(active_range.release_value());
active_range = range;
}
}
if (active_range.has_value()) {
++range_count;
arguments.append(active_range.release_value());
}
arguments[size_index] = range_count;
};
auto contains_regular_table = !table.is_empty();
auto contains_inverted_table = !inverted_table.is_empty();
if (contains_regular_table)
append_table(table);
if (contains_inverted_table) {
++argument_count;
arguments.append(to_underlying(CharacterCompareType::TemporaryInverse));
append_table(inverted_table);
}
};
LibRegex: Use a match table for character classes Generate a sorted, compressed series of ranges in a match table for character classes, and use a binary search to find the matches. This is about a 3-4x speedup for character class match performance. :^)
2021-10-03 15:31:25 +00:00
for (auto& value : pairs) {
auto should_invert_after_this_iteration = invert_for_next_iteration;
invert_for_next_iteration = false;
auto insertion_result = insert_into_lookup_table(*current_table, value);
switch (insertion_result) {
case LookupTableInsertionOutcome::Successful:
break;
case LookupTableInsertionOutcome::ReplaceWithAnyChar: {
table.clear();
inverted_table.clear();
arguments.append(to_underlying(CharacterCompareType::AnyChar));
++argument_count;
break;
}
case LookupTableInsertionOutcome::TemporaryInversionNeeded:
swap(current_table, current_inverted_table);
invert_for_next_iteration = true;
is_currently_inverted = !is_currently_inverted;
break;
case LookupTableInsertionOutcome::PermanentInversionNeeded:
LibRegex: Flush compare tables before entering a permanent inverse state
2022-07-10 07:12:31 +00:00
flush_tables();
arguments.append(to_underlying(CharacterCompareType::Inverse));
++argument_count;
LibRegex: Use a match table for character classes Generate a sorted, compressed series of ranges in a match table for character classes, and use a binary search to find the matches. This is about a 3-4x speedup for character class match performance. :^)
2021-10-03 15:31:25 +00:00
break;
case LookupTableInsertionOutcome::CannotPlaceInTable:
if (is_currently_inverted) {
arguments.append(to_underlying(CharacterCompareType::TemporaryInverse));
++argument_count;
}
arguments.append(to_underlying(value.type));
arguments.append(value.value);
++argument_count;
break;
}
if (should_invert_after_this_iteration) {
swap(current_table, current_inverted_table);
is_currently_inverted = !is_currently_inverted;
}
}
LibRegex: Flush compare tables before entering a permanent inverse state
2022-07-10 07:12:31 +00:00
flush_tables();
LibRegex: Use a match table for character classes Generate a sorted, compressed series of ranges in a match table for character classes, and use a binary search to find the matches. This is about a 3-4x speedup for character class match performance. :^)
2021-10-03 15:31:25 +00:00
}
target.empend(static_cast<ByteCodeValueType>(OpCodeId::Compare));
target.empend(argument_count); // number of arguments
target.empend(arguments.size()); // size of arguments
target.extend(move(arguments));
}
LibRegex: Add a basic optimization pass This currently tries to convert forking loops to atomic groups, and unify the left side of alternations.
2021-09-12 13:00:27 +00:00
template void Regex<PosixBasicParser>::run_optimization_passes();
template void Regex<PosixExtendedParser>::run_optimization_passes();
template void Regex<ECMA262Parser>::run_optimization_passes();
}
Reference in a new issue Copy permalink