Browse Source

Merge pull request #1396 from chriseth/sol_cseOpt2

Common subexpression elimination, including simplifications.
cl-refactor
chriseth 10 years ago
parent
commit
d10db6f7d1
  1. 184
      libevmcore/Assembly.cpp
  2. 55
      libevmcore/Assembly.h
  3. 135
      libevmcore/AssemblyItem.cpp
  4. 92
      libevmcore/AssemblyItem.h
  5. 213
      libevmcore/CommonSubexpressionEliminator.cpp
  6. 60
      libevmcore/CommonSubexpressionEliminator.h
  7. 1
      libevmcore/Exceptions.h
  8. 371
      libevmcore/ExpressionClasses.cpp
  9. 150
      libevmcore/ExpressionClasses.h
  10. 115
      test/SolidityOptimizer.cpp

184
libevmcore/Assembly.cpp

@ -28,122 +28,6 @@ using namespace std;
using namespace dev;
using namespace dev::eth;
unsigned AssemblyItem::bytesRequired(unsigned _addressLength) const
{
switch (m_type)
{
case Operation:
case Tag: // 1 byte for the JUMPDEST
return 1;
case PushString:
return 33;
case Push:
return 1 + max<unsigned>(1, dev::bytesRequired(m_data));
case PushSubSize:
case PushProgramSize:
return 4; // worst case: a 16MB program
case PushTag:
case PushData:
case PushSub:
return 1 + _addressLength;
default:
break;
}
BOOST_THROW_EXCEPTION(InvalidOpcode());
}
int AssemblyItem::deposit() const
{
switch (m_type)
{
case Operation:
return instructionInfo(instruction()).ret - instructionInfo(instruction()).args;
case Push:
case PushString:
case PushTag:
case PushData:
case PushSub:
case PushSubSize:
case PushProgramSize:
return 1;
case Tag:
return 0;
default:;
}
return 0;
}
string AssemblyItem::getJumpTypeAsString() const
{
switch (m_jumpType)
{
case JumpType::IntoFunction:
return "[in]";
case JumpType::OutOfFunction:
return "[out]";
case JumpType::Ordinary:
default:
return "";
}
}
ostream& dev::eth::operator<<(ostream& _out, AssemblyItem const& _item)
{
switch (_item.type())
{
case Operation:
_out << " " << instructionInfo(_item.instruction()).name;
if (_item.instruction() == eth::Instruction::JUMP || _item.instruction() == eth::Instruction::JUMPI)
_out << "\t" << _item.getJumpTypeAsString();
break;
case Push:
_out << " PUSH " << hex << _item.data();
break;
case PushString:
_out << " PushString" << hex << (unsigned)_item.data();
break;
case PushTag:
_out << " PushTag " << _item.data();
break;
case Tag:
_out << " Tag " << _item.data();
break;
case PushData:
_out << " PushData " << hex << (unsigned)_item.data();
break;
case PushSub:
_out << " PushSub " << hex << h256(_item.data()).abridged();
break;
case PushSubSize:
_out << " PushSubSize " << hex << h256(_item.data()).abridged();
break;
case PushProgramSize:
_out << " PushProgramSize";
break;
case UndefinedItem:
_out << " ???";
break;
default:
BOOST_THROW_EXCEPTION(InvalidOpcode());
}
return _out;
}
unsigned Assembly::bytesRequired() const
{
for (unsigned br = 1;; ++br)
{
unsigned ret = 1;
for (auto const& i: m_data)
ret += i.second.size();
for (AssemblyItem const& i: m_items)
ret += i.bytesRequired(br);
if (dev::bytesRequired(ret) <= br)
return ret;
}
}
void Assembly::append(Assembly const& _a)
{
auto newDeposit = m_deposit + _a.deposit();
@ -180,11 +64,19 @@ void Assembly::append(Assembly const& _a, int _deposit)
}
}
ostream& dev::eth::operator<<(ostream& _out, AssemblyItemsConstRef _i)
unsigned Assembly::bytesRequired() const
{
for (AssemblyItem const& i: _i)
_out << i;
return _out;
for (unsigned br = 1;; ++br)
{
unsigned ret = 1;
for (auto const& i: m_data)
ret += i.second.size();
for (AssemblyItem const& i: m_items)
ret += i.bytesRequired(br);
if (dev::bytesRequired(ret) <= br)
return ret;
}
}
string Assembly::getLocationFromSources(StringMap const& _sourceCodes, SourceLocation const& _location) const
@ -288,18 +180,6 @@ inline bool matches(AssemblyItemsConstRef _a, AssemblyItemsConstRef _b)
return true;
}
//@todo this has to move to a special optimizer class soon
template<class Iterator>
unsigned bytesRequiredBySlice(Iterator _begin, Iterator _end)
{
// this is only used in the optimizer, so we can provide a guess for the address length
unsigned addressLength = 4;
unsigned size = 0;
for (; _begin != _end; ++_begin)
size += _begin->bytesRequired(addressLength);
return size;
}
struct OptimiserChannel: public LogChannel { static const char* name() { return "OPT"; } static const int verbosity = 12; };
#define copt dev::LogOutputStream<OptimiserChannel, true>()
@ -307,16 +187,6 @@ Assembly& Assembly::optimise(bool _enable)
{
if (!_enable)
return *this;
map<Instruction, function<u256(u256, u256)>> const c_associative =
{
{ Instruction::ADD, [](u256 a, u256 b)->u256{return a + b;} },
{ Instruction::MUL, [](u256 a, u256 b)->u256{return a * b;} },
{ Instruction::AND, [](u256 a, u256 b)->u256{return a & b;} },
{ Instruction::OR, [](u256 a, u256 b)->u256{return a | b;} },
{ Instruction::XOR, [](u256 a, u256 b)->u256{return a ^ b;} },
};
std::vector<pair<AssemblyItem, u256>> const c_identities =
{ { Instruction::ADD, 0}, { Instruction::MUL, 1}, { Instruction::MOD, 0}, { Instruction::OR, 0}, { Instruction::XOR, 0} };
std::vector<pair<AssemblyItems, function<AssemblyItems(AssemblyItemsConstRef)>>> rules =
{
{ { Push, Instruction::POP }, [](AssemblyItemsConstRef) -> AssemblyItems { return {}; } },
@ -329,22 +199,13 @@ Assembly& Assembly::optimise(bool _enable)
{ { Instruction::ISZERO, Instruction::ISZERO }, [](AssemblyItemsConstRef) -> AssemblyItems { return {}; } },
};
for (auto const& i: c_associative)
{
rules.push_back({ { Push, Push, i.first }, [&](AssemblyItemsConstRef m) -> AssemblyItems { return { i.second(m[1].data(), m[0].data()) }; } });
rules.push_back({ { Push, i.first, Push, i.first }, [&](AssemblyItemsConstRef m) -> AssemblyItems { return { i.second(m[2].data(), m[0].data()), i.first }; } });
}
for (auto const& i: c_identities)
rules.push_back({{Push, i.first}, [&](AssemblyItemsConstRef m) -> AssemblyItems
{ return m[0].data() == i.second ? AssemblyItems() : m.toVector(); }});
// jump to next instruction
rules.push_back({ { PushTag, Instruction::JUMP, Tag }, [](AssemblyItemsConstRef m) -> AssemblyItems { if (m[0].m_data == m[2].m_data) return {m[2]}; else return m.toVector(); }});
copt << *this;
unsigned total = 0;
for (unsigned count = 1; count > 0; total += count)
{
copt << *this;
count = 0;
copt << "Performing common subexpression elimination...";
@ -353,11 +214,22 @@ Assembly& Assembly::optimise(bool _enable)
CommonSubexpressionEliminator eliminator;
auto orig = iter;
iter = eliminator.feedItems(iter, m_items.end());
AssemblyItems optItems = eliminator.getOptimizedItems();
copt << "Old size: " << (iter - orig) << ", new size: " << optItems.size();
if (optItems.size() < size_t(iter - orig))
AssemblyItems optItems;
bool shouldReplace = false;
try
{
optItems = eliminator.getOptimizedItems();
shouldReplace = (optItems.size() < size_t(iter - orig));
}
catch (StackTooDeepException const&)
{
// This might happen if the opcode reconstruction is not as efficient
// as the hand-crafted code.
}
if (shouldReplace)
{
// replace items
copt << "Old size: " << (iter - orig) << ", new size: " << optItems.size();
count++;
for (auto moveIter = optItems.begin(); moveIter != optItems.end(); ++orig, ++moveIter)
*orig = move(*moveIter);

55
libevmcore/Assembly.h

@ -27,6 +27,7 @@
#include <libdevcore/Assertions.h>
#include <libevmcore/SourceLocation.h>
#include <libevmcore/Instruction.h>
#include <libevmcore/AssemblyItem.h>
#include "Exceptions.h"
namespace dev
@ -34,60 +35,6 @@ namespace dev
namespace eth
{
enum AssemblyItemType { UndefinedItem, Operation, Push, PushString, PushTag, PushSub, PushSubSize, PushProgramSize, Tag, PushData };
class Assembly;
class AssemblyItem
{
friend class Assembly;
public:
enum class JumpType { Ordinary, IntoFunction, OutOfFunction };
AssemblyItem(u256 _push): m_type(Push), m_data(_push) {}
AssemblyItem(Instruction _i): m_type(Operation), m_data((byte)_i) {}
AssemblyItem(AssemblyItemType _type, u256 _data = 0): m_type(_type), m_data(_data) {}
AssemblyItem tag() const { assertThrow(m_type == PushTag || m_type == Tag, Exception, ""); return AssemblyItem(Tag, m_data); }
AssemblyItem pushTag() const { assertThrow(m_type == PushTag || m_type == Tag, Exception, ""); return AssemblyItem(PushTag, m_data); }
AssemblyItemType type() const { return m_type; }
u256 const& data() const { return m_data; }
/// @returns the instruction of this item (only valid if type() == Operation)
Instruction instruction() const { return Instruction(byte(m_data)); }
/// @returns true iff the type and data of the items are equal.
bool operator==(AssemblyItem const& _other) const { return m_type == _other.m_type && m_data == _other.m_data; }
bool operator!=(AssemblyItem const& _other) const { return !operator==(_other); }
/// @returns an upper bound for the number of bytes required by this item, assuming that
/// the value of a jump tag takes @a _addressLength bytes.
unsigned bytesRequired(unsigned _addressLength) const;
int deposit() const;
bool match(AssemblyItem const& _i) const { return _i.m_type == UndefinedItem || (m_type == _i.m_type && (m_type != Operation || m_data == _i.m_data)); }
void setLocation(SourceLocation const& _location) { m_location = _location; }
SourceLocation const& getLocation() const { return m_location; }
void setJumpType(JumpType _jumpType) { m_jumpType = _jumpType; }
JumpType getJumpType() const { return m_jumpType; }
std::string getJumpTypeAsString() const;
private:
AssemblyItemType m_type;
u256 m_data;
SourceLocation m_location;
JumpType m_jumpType = JumpType::Ordinary;
};
using AssemblyItems = std::vector<AssemblyItem>;
using AssemblyItemsConstRef = vector_ref<AssemblyItem const>;
std::ostream& operator<<(std::ostream& _out, AssemblyItem const& _item);
std::ostream& operator<<(std::ostream& _out, AssemblyItemsConstRef _i);
inline std::ostream& operator<<(std::ostream& _out, AssemblyItems const& _i) { return operator<<(_out, AssemblyItemsConstRef(&_i)); }
class Assembly
{
public:

135
libevmcore/AssemblyItem.cpp

@ -0,0 +1,135 @@
/*
This file is part of cpp-ethereum.
cpp-ethereum is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
cpp-ethereum is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with cpp-ethereum. If not, see <http://www.gnu.org/licenses/>.
*/
/** @file Assembly.cpp
* @author Gav Wood <i@gavwood.com>
* @date 2014
*/
#include "AssemblyItem.h"
#include <fstream>
using namespace std;
using namespace dev;
using namespace dev::eth;
unsigned AssemblyItem::bytesRequired(unsigned _addressLength) const
{
switch (m_type)
{
case Operation:
case Tag: // 1 byte for the JUMPDEST
return 1;
case PushString:
return 33;
case Push:
return 1 + max<unsigned>(1, dev::bytesRequired(m_data));
case PushSubSize:
case PushProgramSize:
return 4; // worst case: a 16MB program
case PushTag:
case PushData:
case PushSub:
return 1 + _addressLength;
default:
break;
}
BOOST_THROW_EXCEPTION(InvalidOpcode());
}
int AssemblyItem::deposit() const
{
switch (m_type)
{
case Operation:
return instructionInfo(instruction()).ret - instructionInfo(instruction()).args;
case Push:
case PushString:
case PushTag:
case PushData:
case PushSub:
case PushSubSize:
case PushProgramSize:
return 1;
case Tag:
return 0;
default:;
}
return 0;
}
string AssemblyItem::getJumpTypeAsString() const
{
switch (m_jumpType)
{
case JumpType::IntoFunction:
return "[in]";
case JumpType::OutOfFunction:
return "[out]";
case JumpType::Ordinary:
default:
return "";
}
}
ostream& dev::eth::operator<<(ostream& _out, AssemblyItem const& _item)
{
switch (_item.type())
{
case Operation:
_out << " " << instructionInfo(_item.instruction()).name;
if (_item.instruction() == eth::Instruction::JUMP || _item.instruction() == eth::Instruction::JUMPI)
_out << "\t" << _item.getJumpTypeAsString();
break;
case Push:
_out << " PUSH " << hex << _item.data();
break;
case PushString:
_out << " PushString" << hex << (unsigned)_item.data();
break;
case PushTag:
_out << " PushTag " << _item.data();
break;
case Tag:
_out << " Tag " << _item.data();
break;
case PushData:
_out << " PushData " << hex << (unsigned)_item.data();
break;
case PushSub:
_out << " PushSub " << hex << h256(_item.data()).abridged();
break;
case PushSubSize:
_out << " PushSubSize " << hex << h256(_item.data()).abridged();
break;
case PushProgramSize:
_out << " PushProgramSize";
break;
case UndefinedItem:
_out << " ???";
break;
default:
BOOST_THROW_EXCEPTION(InvalidOpcode());
}
return _out;
}
ostream& dev::eth::operator<<(ostream& _out, AssemblyItemsConstRef _i)
{
for (AssemblyItem const& i: _i)
_out << i;
return _out;
}

92
libevmcore/AssemblyItem.h

@ -0,0 +1,92 @@
/*
This file is part of cpp-ethereum.
cpp-ethereum is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
cpp-ethereum is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with cpp-ethereum. If not, see <http://www.gnu.org/licenses/>.
*/
/** @file Assembly.h
* @author Gav Wood <i@gavwood.com>
* @date 2014
*/
#pragma once
#include <iostream>
#include <sstream>
#include <libdevcore/Common.h>
#include <libdevcore/Assertions.h>
#include <libevmcore/SourceLocation.h>
#include <libevmcore/Instruction.h>
#include "Exceptions.h"
namespace dev
{
namespace eth
{
enum AssemblyItemType { UndefinedItem, Operation, Push, PushString, PushTag, PushSub, PushSubSize, PushProgramSize, Tag, PushData };
class Assembly;
class AssemblyItem
{
friend class Assembly;
public:
enum class JumpType { Ordinary, IntoFunction, OutOfFunction };
AssemblyItem(u256 _push): m_type(Push), m_data(_push) {}
AssemblyItem(Instruction _i): m_type(Operation), m_data((byte)_i) {}
AssemblyItem(AssemblyItemType _type, u256 _data = 0): m_type(_type), m_data(_data) {}
AssemblyItem tag() const { assertThrow(m_type == PushTag || m_type == Tag, Exception, ""); return AssemblyItem(Tag, m_data); }
AssemblyItem pushTag() const { assertThrow(m_type == PushTag || m_type == Tag, Exception, ""); return AssemblyItem(PushTag, m_data); }
AssemblyItemType type() const { return m_type; }
u256 const& data() const { return m_data; }
/// @returns the instruction of this item (only valid if type() == Operation)
Instruction instruction() const { return Instruction(byte(m_data)); }
/// @returns true iff the type and data of the items are equal.
bool operator==(AssemblyItem const& _other) const { return m_type == _other.m_type && m_data == _other.m_data; }
bool operator!=(AssemblyItem const& _other) const { return !operator==(_other); }
/// @returns an upper bound for the number of bytes required by this item, assuming that
/// the value of a jump tag takes @a _addressLength bytes.
unsigned bytesRequired(unsigned _addressLength) const;
int deposit() const;
bool match(AssemblyItem const& _i) const { return _i.m_type == UndefinedItem || (m_type == _i.m_type && (m_type != Operation || m_data == _i.m_data)); }
void setLocation(SourceLocation const& _location) { m_location = _location; }
SourceLocation const& getLocation() const { return m_location; }
void setJumpType(JumpType _jumpType) { m_jumpType = _jumpType; }
JumpType getJumpType() const { return m_jumpType; }
std::string getJumpTypeAsString() const;
private:
AssemblyItemType m_type;
u256 m_data;
SourceLocation m_location;
JumpType m_jumpType = JumpType::Ordinary;
};
using AssemblyItems = std::vector<AssemblyItem>;
using AssemblyItemsConstRef = vector_ref<AssemblyItem const>;
std::ostream& operator<<(std::ostream& _out, AssemblyItem const& _item);
std::ostream& operator<<(std::ostream& _out, AssemblyItemsConstRef _i);
inline std::ostream& operator<<(std::ostream& _out, AssemblyItems const& _i) { return operator<<(_out, AssemblyItemsConstRef(&_i)); }
}
}

213
libevmcore/CommonSubexpressionEliminator.cpp

@ -32,39 +32,34 @@ using namespace dev::eth;
vector<AssemblyItem> CommonSubexpressionEliminator::getOptimizedItems()
{
map<int, EquivalenceClassId> initialStackContents;
map<int, EquivalenceClassId> targetStackContents;
map<int, ExpressionClasses::Id> initialStackContents;
map<int, ExpressionClasses::Id> targetStackContents;
int minHeight = m_stackHeight + 1;
if (!m_stackElements.empty())
minHeight = min(minHeight, m_stackElements.begin()->first.first);
for (int height = minHeight; height <= max(0, m_stackHeight); ++height)
{
// make sure it is created
EquivalenceClassId c = getStackElement(height);
if (height <= 0)
initialStackContents[height] = getClass(AssemblyItem(dupInstruction(1 - height)));
if (height <= m_stackHeight)
targetStackContents[height] = c;
}
minHeight = min(minHeight, m_stackElements.begin()->first);
for (int height = minHeight; height <= 0; ++height)
initialStackContents[height] = initialStackElement(height);
for (int height = minHeight; height <= m_stackHeight; ++height)
targetStackContents[height] = stackElement(height);
// Debug info:
//stream(cout, currentStackContents, targetStackContents);
//stream(cout, initialStackContents, targetStackContents);
return CSECodeGenerator().generateCode(initialStackContents, targetStackContents, m_equivalenceClasses);
return CSECodeGenerator(m_expressionClasses).generateCode(initialStackContents, targetStackContents);
}
ostream& CommonSubexpressionEliminator::stream(
ostream& _out,
map<int, EquivalenceClassId> _currentStack,
map<int, EquivalenceClassId> _targetStack
map<int, ExpressionClasses::Id> _currentStack,
map<int, ExpressionClasses::Id> _targetStack
) const
{
auto streamEquivalenceClass = [this](ostream& _out, EquivalenceClassId _id)
auto streamExpressionClass = [this](ostream& _out, ExpressionClasses::Id _id)
{
auto const& eqClass = m_equivalenceClasses.at(_id);
_out << " " << _id << ": " << *eqClass.first;
auto const& expr = m_expressionClasses.representative(_id);
_out << " " << _id << ": " << *expr.item;
_out << "(";
for (EquivalenceClassId arg: eqClass.second)
for (ExpressionClasses::Id arg: expr.arguments)
_out << dec << arg << ",";
_out << ")" << endl;
};
@ -74,24 +69,24 @@ ostream& CommonSubexpressionEliminator::stream(
_out << "Stack elements: " << endl;
for (auto const& it: m_stackElements)
{
_out << " " << dec << it.first.first << "(" << it.first.second << ") = ";
streamEquivalenceClass(_out, it.second);
_out << " " << dec << it.first << " = ";
streamExpressionClass(_out, it.second);
}
_out << "Equivalence classes: " << endl;
for (EquivalenceClassId eqClass = 0; eqClass < m_equivalenceClasses.size(); ++eqClass)
streamEquivalenceClass(_out, eqClass);
for (ExpressionClasses::Id eqClass = 0; eqClass < m_expressionClasses.size(); ++eqClass)
streamExpressionClass(_out, eqClass);
_out << "Current stack: " << endl;
for (auto const& it: _currentStack)
{
_out << " " << dec << it.first << ": ";
streamEquivalenceClass(_out, it.second);
streamExpressionClass(_out, it.second);
}
_out << "Target stack: " << endl;
for (auto const& it: _targetStack)
{
_out << " " << dec << it.first << ": ";
streamEquivalenceClass(_out, it.second);
streamExpressionClass(_out, it.second);
}
return _out;
@ -103,7 +98,7 @@ void CommonSubexpressionEliminator::feedItem(AssemblyItem const& _item)
{
if (_item.deposit() != 1)
BOOST_THROW_EXCEPTION(InvalidDeposit());
setStackElement(++m_stackHeight, getClass(_item, {}));
setStackElement(++m_stackHeight, m_expressionClasses.find(_item, {}));
}
else
{
@ -112,7 +107,7 @@ void CommonSubexpressionEliminator::feedItem(AssemblyItem const& _item)
if (SemanticInformation::isDupInstruction(_item))
setStackElement(
m_stackHeight + 1,
getStackElement(m_stackHeight - int(instruction) + int(Instruction::DUP1))
stackElement(m_stackHeight - int(instruction) + int(Instruction::DUP1))
);
else if (SemanticInformation::isSwapInstruction(_item))
swapStackElements(
@ -121,139 +116,45 @@ void CommonSubexpressionEliminator::feedItem(AssemblyItem const& _item)
);
else if (instruction != Instruction::POP)
{
vector<EquivalenceClassId> arguments(info.args);
vector<ExpressionClasses::Id> arguments(info.args);
for (int i = 0; i < info.args; ++i)
arguments[i] = getStackElement(m_stackHeight - i);
setStackElement(m_stackHeight + _item.deposit(), getClass(_item, arguments));
arguments[i] = stackElement(m_stackHeight - i);
setStackElement(m_stackHeight + _item.deposit(), m_expressionClasses.find(_item, arguments));
}
m_stackHeight += _item.deposit();
}
}
void CommonSubexpressionEliminator::setStackElement(int _stackHeight, EquivalenceClassId _class)
void CommonSubexpressionEliminator::setStackElement(int _stackHeight, ExpressionClasses::Id _class)
{
unsigned nextSequence = getNextStackElementSequence(_stackHeight);
m_stackElements[make_pair(_stackHeight, nextSequence)] = _class;
m_stackElements[_stackHeight] = _class;
}
void CommonSubexpressionEliminator::swapStackElements(int _stackHeightA, int _stackHeightB)
{
if (_stackHeightA == _stackHeightB)
BOOST_THROW_EXCEPTION(OptimizerException() << errinfo_comment("Swap on same stack elements."));
EquivalenceClassId classA = getStackElement(_stackHeightA);
EquivalenceClassId classB = getStackElement(_stackHeightB);
// ensure they are created
stackElement(_stackHeightA);
stackElement(_stackHeightB);
unsigned nextSequenceA = getNextStackElementSequence(_stackHeightA);
unsigned nextSequenceB = getNextStackElementSequence(_stackHeightB);
m_stackElements[make_pair(_stackHeightA, nextSequenceA)] = classB;
m_stackElements[make_pair(_stackHeightB, nextSequenceB)] = classA;
swap(m_stackElements[_stackHeightA], m_stackElements[_stackHeightB]);
}
EquivalenceClassId CommonSubexpressionEliminator::getStackElement(int _stackHeight)
ExpressionClasses::Id CommonSubexpressionEliminator::stackElement(int _stackHeight)
{
// retrieve class by last sequence number
unsigned nextSequence = getNextStackElementSequence(_stackHeight);
if (nextSequence > 0)
return m_stackElements[make_pair(_stackHeight, nextSequence - 1)];
if (m_stackElements.count(_stackHeight))
return m_stackElements.at(_stackHeight);
// Stack element not found (not assigned yet), create new equivalence class.
if (_stackHeight > 0)
BOOST_THROW_EXCEPTION(OptimizerException() << errinfo_comment("Stack element accessed before assignment."));
if (_stackHeight <= -16)
BOOST_THROW_EXCEPTION(OptimizerException() << errinfo_comment("Stack too deep."));
// This is a special assembly item that refers to elements pre-existing on the initial stack.
m_spareAssemblyItem.push_back(make_shared<AssemblyItem>(dupInstruction(1 - _stackHeight)));
m_equivalenceClasses.push_back(make_pair(m_spareAssemblyItem.back().get(), EquivalenceClassIds()));
return m_stackElements[make_pair(_stackHeight, nextSequence)] = EquivalenceClassId(m_equivalenceClasses.size() - 1);
}
EquivalenceClassId CommonSubexpressionEliminator::getClass(
const AssemblyItem& _item,
EquivalenceClassIds const& _arguments
)
{
// TODO: do a clever search, i.e.
// - check for the presence of constants in the argument classes and do arithmetic
// - check whether the two items are equal for a SUB instruction
// - check whether 0 or 1 is in one of the classes for a MUL
EquivalenceClassIds args = _arguments;
if (SemanticInformation::isCommutativeOperation(_item))
sort(args.begin(), args.end());
//@todo use a better data structure for search here
for (EquivalenceClassId c = 0; c < m_equivalenceClasses.size(); ++c)
{
AssemblyItem const& classItem = *m_equivalenceClasses.at(c).first;
if (classItem != _item)
continue;
assertThrow(
args.size() == m_equivalenceClasses.at(c).second.size(),
OptimizerException,
"Equal assembly items with different number of arguments."
);
if (equal(args.begin(), args.end(), m_equivalenceClasses.at(c).second.begin()))
return c;
}
// constant folding
if (_item.type() == Operation && args.size() == 2 && all_of(
args.begin(),
args.end(),
[this](EquivalenceClassId eqc) { return m_equivalenceClasses.at(eqc).first->match(Push); }))
{
auto signextend = [](u256 const& _a, u256 const& _b) -> u256
{
if (_a >= 31)
return _b;
unsigned testBit = unsigned(_a) * 8 + 7;
u256 mask = (u256(1) << testBit) - 1;
return boost::multiprecision::bit_test(_b, testBit) ? _b | ~mask : _b & mask;
};
map<Instruction, function<u256(u256 const&, u256 const&)>> const arithmetics =
{
{ Instruction::SUB, [](u256 const& _a, u256 const& _b) -> u256 {return _a - _b; } },
{ Instruction::DIV, [](u256 const& _a, u256 const& _b) -> u256 {return _b == 0 ? 0 : _a / _b; } },
{ Instruction::SDIV, [](u256 const& _a, u256 const& _b) -> u256 { return _b == 0 ? 0 : s2u(u2s(_a) / u2s(_b)); } },
{ Instruction::MOD, [](u256 const& _a, u256 const& _b) -> u256 { return _b == 0 ? 0 : _a % _b; } },
{ Instruction::SMOD, [](u256 const& _a, u256 const& _b) -> u256 { return _b == 0 ? 0 : s2u(u2s(_a) % u2s(_b)); } },
{ Instruction::EXP, [](u256 const& _a, u256 const& _b) -> u256 { return (u256)boost::multiprecision::powm(bigint(_a), bigint(_b), bigint(1) << 256); } },
{ Instruction::SIGNEXTEND, signextend },
{ Instruction::LT, [](u256 const& _a, u256 const& _b) -> u256 { return _a < _b ? 1 : 0; } },
{ Instruction::GT, [](u256 const& _a, u256 const& _b) -> u256 { return _a > _b ? 1 : 0; } },
{ Instruction::SLT, [](u256 const& _a, u256 const& _b) -> u256 { return u2s(_a) < u2s(_b) ? 1 : 0; } },
{ Instruction::SGT, [](u256 const& _a, u256 const& _b) -> u256 { return u2s(_a) > u2s(_b) ? 1 : 0; } },
{ Instruction::EQ, [](u256 const& _a, u256 const& _b) -> u256 { return _a == _b ? 1 : 0; } },
{ Instruction::ADD, [](u256 const& _a, u256 const& _b) -> u256 { return _a + _b; } },
{ Instruction::MUL, [](u256 const& _a, u256 const& _b) -> u256 { return _a * _b; } },
{ Instruction::AND, [](u256 const& _a, u256 const& _b) -> u256 { return _a & _b; } },
{ Instruction::OR, [](u256 const& _a, u256 const& _b) -> u256 { return _a | _b; } },
{ Instruction::XOR, [](u256 const& _a, u256 const& _b) -> u256 { return _a ^ _b; } },
};
if (arithmetics.count(_item.instruction()))
{
u256 result = arithmetics.at(_item.instruction())(
m_equivalenceClasses.at(args[0]).first->data(),
m_equivalenceClasses.at(args[1]).first->data()
);
m_spareAssemblyItem.push_back(make_shared<AssemblyItem>(result));
return getClass(*m_spareAssemblyItem.back());
}
}
m_equivalenceClasses.push_back(make_pair(&_item, args));
return m_equivalenceClasses.size() - 1;
return m_stackElements[_stackHeight] = initialStackElement(_stackHeight);
}
unsigned CommonSubexpressionEliminator::getNextStackElementSequence(int _stackHeight)
ExpressionClasses::Id CommonSubexpressionEliminator::initialStackElement(int _stackHeight)
{
auto it = m_stackElements.upper_bound(make_pair(_stackHeight, unsigned(-1)));
if (it == m_stackElements.begin())
return 0;
--it;
if (it->first.first == _stackHeight)
return it->first.second + 1;
else
return 0;
assertThrow(_stackHeight <= 0, OptimizerException, "Initial stack element of positive height requested.");
assertThrow(_stackHeight > -16, StackTooDeepException, "");
// This is a special assembly item that refers to elements pre-existing on the initial stack.
return m_expressionClasses.find(AssemblyItem(dupInstruction(1 - _stackHeight)));
}
bool SemanticInformation::breaksBasicBlock(AssemblyItem const& _item)
@ -318,15 +219,11 @@ bool SemanticInformation::isSwapInstruction(AssemblyItem const& _item)
}
AssemblyItems CSECodeGenerator::generateCode(
map<int, EquivalenceClassId> const& _initialStack,
map<int, EquivalenceClassId> const& _targetStackContents,
vector<pair<AssemblyItem const*, EquivalenceClassIds>> const& _equivalenceClasses
map<int, ExpressionClasses::Id> const& _initialStack,
map<int, ExpressionClasses::Id> const& _targetStackContents
)
{
// reset
*this = move(CSECodeGenerator());
m_stack = _initialStack;
m_equivalenceClasses = _equivalenceClasses;
for (auto const& item: m_stack)
if (!m_classPositions.count(item.second))
m_classPositions[item.second] = item.first;
@ -377,18 +274,18 @@ AssemblyItems CSECodeGenerator::generateCode(
return m_generatedItems;
}
void CSECodeGenerator::addDependencies(EquivalenceClassId _c)
void CSECodeGenerator::addDependencies(ExpressionClasses::Id _c)
{
if (m_neededBy.count(_c))
return;
for (EquivalenceClassId argument: m_equivalenceClasses.at(_c).second)
for (ExpressionClasses::Id argument: m_expressionClasses.representative(_c).arguments)
{
addDependencies(argument);
m_neededBy.insert(make_pair(argument, _c));
}
}
int CSECodeGenerator::generateClassElement(EquivalenceClassId _c)
int CSECodeGenerator::generateClassElement(ExpressionClasses::Id _c)
{
if (m_classPositions.count(_c))
{
@ -399,8 +296,8 @@ int CSECodeGenerator::generateClassElement(EquivalenceClassId _c)
);
return m_classPositions[_c];
}
EquivalenceClassIds const& arguments = m_equivalenceClasses.at(_c).second;
for (EquivalenceClassId arg: boost::adaptors::reverse(arguments))
ExpressionClasses::Ids const& arguments = m_expressionClasses.representative(_c).arguments;
for (ExpressionClasses::Id arg: boost::adaptors::reverse(arguments))
generateClassElement(arg);
// The arguments are somewhere on the stack now, so it remains to move them at the correct place.
@ -458,7 +355,7 @@ int CSECodeGenerator::generateClassElement(EquivalenceClassId _c)
for (size_t i = 0; i < arguments.size(); ++i)
assertThrow(m_stack[m_stackHeight - i] == arguments[i], OptimizerException, "Expected arguments not present." );
AssemblyItem const& item = *m_equivalenceClasses.at(_c).first;
AssemblyItem const& item = *m_expressionClasses.representative(_c).item;
while (SemanticInformation::isCommutativeOperation(item) &&
!m_generatedItems.empty() &&
m_generatedItems.back() == AssemblyItem(Instruction::SWAP1))
@ -469,12 +366,12 @@ int CSECodeGenerator::generateClassElement(EquivalenceClassId _c)
m_classPositions[arg] = c_invalidPosition;
for (size_t i = 0; i < arguments.size(); ++i)
m_stack.erase(m_stackHeight - i);
appendItem(*m_equivalenceClasses.at(_c).first);
appendItem(*m_expressionClasses.representative(_c).item);
m_stack[m_stackHeight] = _c;
return m_classPositions[_c] = m_stackHeight;
}
bool CSECodeGenerator::canBeRemoved(EquivalenceClassId _element, EquivalenceClassId _result)
bool CSECodeGenerator::canBeRemoved(ExpressionClasses::Id _element, ExpressionClasses::Id _result)
{
// Returns false if _element is finally needed or is needed by a class that has not been
// computed yet. Note that m_classPositions also includes classes that were deleted in the meantime.
@ -493,7 +390,7 @@ bool CSECodeGenerator::removeStackTopIfPossible()
if (m_stack.empty())
return false;
assertThrow(m_stack.count(m_stackHeight), OptimizerException, "");
EquivalenceClassId top = m_stack[m_stackHeight];
ExpressionClasses::Id top = m_stack[m_stackHeight];
if (!canBeRemoved(top))
return false;
m_generatedItems.push_back(AssemblyItem(Instruction::POP));
@ -505,7 +402,8 @@ bool CSECodeGenerator::removeStackTopIfPossible()
void CSECodeGenerator::appendDup(int _fromPosition)
{
int nr = 1 + m_stackHeight - _fromPosition;
assertThrow(1 <= nr && nr <= 16, OptimizerException, "Stack too deep.");
assertThrow(nr <= 16, StackTooDeepException, "Stack too deep.");
assertThrow(1 <= nr, OptimizerException, "Invalid stack access.");
m_generatedItems.push_back(AssemblyItem(dupInstruction(nr)));
m_stackHeight++;
m_stack[m_stackHeight] = m_stack[_fromPosition];
@ -516,7 +414,8 @@ void CSECodeGenerator::appendSwapOrRemove(int _fromPosition)
if (_fromPosition == m_stackHeight)
return;
int nr = m_stackHeight - _fromPosition;
assertThrow(1 <= nr && nr <= 16, OptimizerException, "Stack too deep.");
assertThrow(nr <= 16, StackTooDeepException, "Stack too deep.");
assertThrow(1 <= nr, OptimizerException, "Invalid stack access.");
m_generatedItems.push_back(AssemblyItem(swapInstruction(nr)));
// The value of a class can be present in multiple locations on the stack. We only update the
// "canonical" one that is tracked by m_classPositions

60
libevmcore/CommonSubexpressionEliminator.h

@ -28,6 +28,7 @@
#include <ostream>
#include <libdevcore/CommonIO.h>
#include <libdevcore/Exceptions.h>
#include <libevmcore/ExpressionClasses.h>
namespace dev
{
@ -37,9 +38,6 @@ namespace eth
class AssemblyItem;
using AssemblyItems = std::vector<AssemblyItem>;
using EquivalenceClassId = unsigned;
using EquivalenceClassIds = std::vector<EquivalenceClassId>;
/**
* Optimizer step that performs common subexpression elimination and stack reorganisation,
* i.e. it tries to infer equality among expressions and compute the values of two expressions
@ -67,37 +65,32 @@ public:
/// Streams debugging information to @a _out.
std::ostream& stream(
std::ostream& _out,
std::map<int, EquivalenceClassId> _currentStack = std::map<int, EquivalenceClassId>(),
std::map<int, EquivalenceClassId> _targetStack = std::map<int, EquivalenceClassId>()
std::map<int, ExpressionClasses::Id> _currentStack = std::map<int, ExpressionClasses::Id>(),
std::map<int, ExpressionClasses::Id> _targetStack = std::map<int, ExpressionClasses::Id>()
) const;
private:
/// Feeds the item into the system for analysis.
void feedItem(AssemblyItem const& _item);
/// Simplifies the given item using
/// Assigns a new equivalence class to the next sequence number of the given stack element.
void setStackElement(int _stackHeight, EquivalenceClassId _class);
void setStackElement(int _stackHeight, ExpressionClasses::Id _class);
/// Swaps the given stack elements in their next sequence number.
void swapStackElements(int _stackHeightA, int _stackHeightB);
/// Retrieves the current equivalence class fo the given stack element (or generates a new
/// one if it does not exist yet).
EquivalenceClassId getStackElement(int _stackHeight);
/// Retrieves the equivalence class resulting from the given item applied to the given classes,
/// might also create a new one.
EquivalenceClassId getClass(AssemblyItem const& _item, EquivalenceClassIds const& _arguments = {});
/// @returns the next sequence number of the given stack element.
unsigned getNextStackElementSequence(int _stackHeight);
ExpressionClasses::Id stackElement(int _stackHeight);
/// @returns the equivalence class id of the special initial stack element at the given height
/// (must not be positive).
ExpressionClasses::Id initialStackElement(int _stackHeight);
/// Current stack height, can be negative.
int m_stackHeight = 0;
/// Mapping (stack height, sequence number) -> equivalence class
std::map<std::pair<int, unsigned>, EquivalenceClassId> m_stackElements;
/// Vector of equivalence class representatives - we only store one item of an equivalence
/// class and the index is used as identifier.
std::vector<std::pair<AssemblyItem const*, EquivalenceClassIds>> m_equivalenceClasses;
/// List of items generated during analysis.
std::vector<std::shared_ptr<AssemblyItem>> m_spareAssemblyItem;
/// Current stack layout, mapping stack height -> equivalence class
std::map<int, ExpressionClasses::Id> m_stackElements;
/// Structure containing the classes of equivalent expressions.
ExpressionClasses m_expressionClasses;
};
/**
@ -121,27 +114,30 @@ struct SemanticInformation
class CSECodeGenerator
{
public:
CSECodeGenerator(ExpressionClasses const& _expressionClasses):
m_expressionClasses(_expressionClasses)
{}
/// @returns the assembly items generated from the given requirements
/// @param _initialStack current contents of the stack (up to stack height of zero)
/// @param _targetStackContents final contents of the stack, by stack height relative to initial
/// @param _equivalenceClasses equivalence classes as expressions of how to compute them
/// @note resuts the state of the object for each call.
/// @note should only be called once on each object.
AssemblyItems generateCode(
std::map<int, EquivalenceClassId> const& _initialStack,
std::map<int, EquivalenceClassId> const& _targetStackContents,
std::vector<std::pair<AssemblyItem const*, EquivalenceClassIds>> const& _equivalenceClasses
std::map<int, ExpressionClasses::Id> const& _initialStack,
std::map<int, ExpressionClasses::Id> const& _targetStackContents
);
private:
/// Recursively discovers all dependencies to @a m_requests.
void addDependencies(EquivalenceClassId _c);
void addDependencies(ExpressionClasses::Id _c);
/// Produce code that generates the given element if it is not yet present.
/// @returns the stack position of the element.
int generateClassElement(EquivalenceClassId _c);
int generateClassElement(ExpressionClasses::Id _c);
/// @returns true if @a _element can be removed - in general or, if given, while computing @a _result.
bool canBeRemoved(EquivalenceClassId _element, EquivalenceClassId _result = EquivalenceClassId(-1));
bool canBeRemoved(ExpressionClasses::Id _element, ExpressionClasses::Id _result = ExpressionClasses::Id(-1));
/// Appends code to remove the topmost stack element if it can be removed.
bool removeStackTopIfPossible();
@ -160,16 +156,16 @@ private:
/// Current height of the stack relative to the start.
int m_stackHeight = 0;
/// If (b, a) is in m_requests then b is needed to compute a.
std::multimap<EquivalenceClassId, EquivalenceClassId> m_neededBy;
std::multimap<ExpressionClasses::Id, ExpressionClasses::Id> m_neededBy;
/// Current content of the stack.
std::map<int, EquivalenceClassId> m_stack;
std::map<int, ExpressionClasses::Id> m_stack;
/// Current positions of equivalence classes, equal to c_invalidPosition if already deleted.
std::map<EquivalenceClassId, int> m_classPositions;
std::map<ExpressionClasses::Id, int> m_classPositions;
/// The actual eqivalence class items and how to compute them.
std::vector<std::pair<AssemblyItem const*, EquivalenceClassIds>> m_equivalenceClasses;
ExpressionClasses const& m_expressionClasses;
/// The set of equivalence classes that should be present on the stack at the end.
std::set<EquivalenceClassId> m_finalClasses;
std::set<ExpressionClasses::Id> m_finalClasses;
};
template <class _AssemblyItemIterator>

1
libevmcore/Exceptions.h

@ -32,6 +32,7 @@ struct AssemblyException: virtual Exception {};
struct InvalidDeposit: virtual AssemblyException {};
struct InvalidOpcode: virtual AssemblyException {};
struct OptimizerException: virtual AssemblyException {};
struct StackTooDeepException: virtual OptimizerException {};
}
}

371
libevmcore/ExpressionClasses.cpp

@ -0,0 +1,371 @@
/*
This file is part of cpp-ethereum.
cpp-ethereum is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
cpp-ethereum is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with cpp-ethereum. If not, see <http://www.gnu.org/licenses/>.
*/
/**
* @file ExpressionClasses.cpp
* @author Christian <c@ethdev.com>
* @date 2015
* Container for equivalence classes of expressions for use in common subexpression elimination.
*/
#include <libevmcore/ExpressionClasses.h>
#include <utility>
#include <tuple>
#include <functional>
#include <boost/range/adaptor/reversed.hpp>
#include <boost/noncopyable.hpp>
#include <libevmcore/Assembly.h>
#include <libevmcore/CommonSubexpressionEliminator.h>
using namespace std;
using namespace dev;
using namespace dev::eth;
bool ExpressionClasses::Expression::operator<(ExpressionClasses::Expression const& _other) const
{
auto type = item->type();
auto otherType = _other.item->type();
return std::tie(type, item->data(), arguments) <
std::tie(otherType, _other.item->data(), _other.arguments);
}
ExpressionClasses::Id ExpressionClasses::find(AssemblyItem const& _item, Ids const& _arguments)
{
Expression exp;
exp.id = Id(-1);
exp.item = &_item;
exp.arguments = _arguments;
if (SemanticInformation::isCommutativeOperation(_item))
sort(exp.arguments.begin(), exp.arguments.end());
//@todo store all class members (not only the representatives) in an efficient data structure to search here
for (Expression const& e: m_representatives)
if (!(e < exp || exp < e))
return e.id;
if (SemanticInformation::isDupInstruction(_item))
{
// Special item that refers to values pre-existing on the stack
m_spareAssemblyItem.push_back(make_shared<AssemblyItem>(_item));
exp.item = m_spareAssemblyItem.back().get();
}
ExpressionClasses::Id id = tryToSimplify(exp);
if (id < m_representatives.size())
return id;
exp.id = m_representatives.size();
m_representatives.push_back(exp);
return exp.id;
}
string ExpressionClasses::fullDAGToString(ExpressionClasses::Id _id) const
{
Expression const& expr = representative(_id);
stringstream str;
str << dec << expr.id << ":" << *expr.item << "(";
for (Id arg: expr.arguments)
str << fullDAGToString(arg) << ",";
str << ")";
return str.str();
}
class Rules: public boost::noncopyable
{
public:
Rules();
void resetMatchGroups() { m_matchGroups.clear(); }
vector<pair<Pattern, function<Pattern()>>> rules() const { return m_rules; }
private:
using Expression = ExpressionClasses::Expression;
map<unsigned, Expression const*> m_matchGroups;
vector<pair<Pattern, function<Pattern()>>> m_rules;
};
Rules::Rules()
{
// Multiple occurences of one of these inside one rule must match the same equivalence class.
// Constants.
Pattern A(Push);
Pattern B(Push);
Pattern C(Push);
// Anything.
Pattern X;
Pattern Y;
Pattern Z;
A.setMatchGroup(1, m_matchGroups);
B.setMatchGroup(2, m_matchGroups);
C.setMatchGroup(3, m_matchGroups);
X.setMatchGroup(4, m_matchGroups);
Y.setMatchGroup(5, m_matchGroups);
Z.setMatchGroup(6, m_matchGroups);
m_rules = vector<pair<Pattern, function<Pattern()>>>{
// arithmetics on constants
{{Instruction::ADD, {A, B}}, [=]{ return A.d() + B.d(); }},
{{Instruction::MUL, {A, B}}, [=]{ return A.d() * B.d(); }},
{{Instruction::SUB, {A, B}}, [=]{ return A.d() - B.d(); }},
{{Instruction::DIV, {A, B}}, [=]{ return B.d() == 0 ? 0 : A.d() / B.d(); }},
{{Instruction::SDIV, {A, B}}, [=]{ return B.d() == 0 ? 0 : s2u(u2s(A.d()) / u2s(B.d())); }},
{{Instruction::MOD, {A, B}}, [=]{ return B.d() == 0 ? 0 : A.d() % B.d(); }},
{{Instruction::SMOD, {A, B}}, [=]{ return B.d() == 0 ? 0 : s2u(u2s(A.d()) % u2s(B.d())); }},
{{Instruction::EXP, {A, B}}, [=]{ return u256(boost::multiprecision::powm(bigint(A.d()), bigint(B.d()), bigint(1) << 256)); }},
{{Instruction::NOT, {A}}, [=]{ return ~A.d(); }},
{{Instruction::LT, {A, B}}, [=]() { return A.d() < B.d() ? u256(1) : 0; }},
{{Instruction::GT, {A, B}}, [=]() -> u256 { return A.d() > B.d() ? 1 : 0; }},
{{Instruction::SLT, {A, B}}, [=]() -> u256 { return u2s(A.d()) < u2s(B.d()) ? 1 : 0; }},
{{Instruction::SGT, {A, B}}, [=]() -> u256 { return u2s(A.d()) > u2s(B.d()) ? 1 : 0; }},
{{Instruction::EQ, {A, B}}, [=]() -> u256 { return A.d() == B.d() ? 1 : 0; }},
{{Instruction::ISZERO, {A}}, [=]() -> u256 { return A.d() == 0 ? 1 : 0; }},
{{Instruction::AND, {A, B}}, [=]{ return A.d() & B.d(); }},
{{Instruction::OR, {A, B}}, [=]{ return A.d() | B.d(); }},
{{Instruction::XOR, {A, B}}, [=]{ return A.d() ^ B.d(); }},
{{Instruction::BYTE, {A, B}}, [=]{ return A.d() >= 32 ? 0 : (B.d() >> unsigned(8 * (31 - A.d()))) & 0xff; }},
{{Instruction::ADDMOD, {A, B, C}}, [=]{ return C.d() == 0 ? 0 : u256((bigint(A.d()) + bigint(B.d())) % C.d()); }},
{{Instruction::MULMOD, {A, B, C}}, [=]{ return C.d() == 0 ? 0 : u256((bigint(A.d()) * bigint(B.d())) % C.d()); }},
{{Instruction::MULMOD, {A, B, C}}, [=]{ return A.d() * B.d(); }},
{{Instruction::SIGNEXTEND, {A, B}}, [=]() -> u256 {
if (A.d() >= 31)
return B.d();
unsigned testBit = unsigned(A.d()) * 8 + 7;
u256 mask = (u256(1) << testBit) - 1;
return u256(boost::multiprecision::bit_test(B.d(), testBit) ? B.d() | ~mask : B.d() & mask);
}},
// invariants involving known constants
{{Instruction::ADD, {X, 0}}, [=]{ return X; }},
{{Instruction::MUL, {X, 1}}, [=]{ return X; }},
{{Instruction::DIV, {X, 1}}, [=]{ return X; }},
{{Instruction::SDIV, {X, 1}}, [=]{ return X; }},
{{Instruction::OR, {X, 0}}, [=]{ return X; }},
{{Instruction::XOR, {X, 0}}, [=]{ return X; }},
{{Instruction::AND, {X, ~u256(0)}}, [=]{ return X; }},
{{Instruction::MUL, {X, 0}}, [=]{ return u256(0); }},
{{Instruction::DIV, {X, 0}}, [=]{ return u256(0); }},
{{Instruction::MOD, {X, 0}}, [=]{ return u256(0); }},
{{Instruction::MOD, {0, X}}, [=]{ return u256(0); }},
{{Instruction::AND, {X, 0}}, [=]{ return u256(0); }},
{{Instruction::OR, {X, ~u256(0)}}, [=]{ return ~u256(0); }},
// operations involving an expression and itself
{{Instruction::AND, {X, X}}, [=]{ return X; }},
{{Instruction::OR, {X, X}}, [=]{ return X; }},
{{Instruction::SUB, {X, X}}, [=]{ return u256(0); }},
{{Instruction::EQ, {X, X}}, [=]{ return u256(1); }},
{{Instruction::LT, {X, X}}, [=]{ return u256(0); }},
{{Instruction::SLT, {X, X}}, [=]{ return u256(0); }},
{{Instruction::GT, {X, X}}, [=]{ return u256(0); }},
{{Instruction::SGT, {X, X}}, [=]{ return u256(0); }},
{{Instruction::MOD, {X, X}}, [=]{ return u256(0); }},
{{Instruction::NOT, {{Instruction::NOT, {X}}}}, [=]{ return X; }},
};
// Associative operations
for (auto const& opFun: vector<pair<Instruction,function<u256(u256 const&,u256 const&)>>>{
{Instruction::ADD, plus<u256>()},
{Instruction::MUL, multiplies<u256>()},
{Instruction::AND, bit_and<u256>()},
{Instruction::OR, bit_or<u256>()},
{Instruction::XOR, bit_xor<u256>()}
})
{
auto op = opFun.first;
auto fun = opFun.second;
// Moving constants to the outside, order matters here!
// we need actions that return expressions (or patterns?) here, and we need also reversed rules
// (X+A)+B -> X+(A+B)
m_rules.push_back({
{op, {{op, {X, A}}, B}},
[=]() -> Pattern { return {op, {X, fun(A.d(), B.d())}}; }
});
// X+(Y+A) -> (X+Y)+A
m_rules.push_back({
{op, {{op, {X, A}}, Y}},
[=]() -> Pattern { return {op, {{op, {X, Y}}, A}}; }
});
// For now, we still need explicit commutativity for the inner pattern
m_rules.push_back({
{op, {{op, {A, X}}, B}},
[=]() -> Pattern { return {op, {X, fun(A.d(), B.d())}}; }
});
m_rules.push_back({
{op, {{op, {A, X}}, Y}},
[=]() -> Pattern { return {op, {{op, {X, Y}}, A}}; }
});
};
//@todo: (x+8)-3 and other things
}
ExpressionClasses::Id ExpressionClasses::tryToSimplify(Expression const& _expr, bool _secondRun)
{
static Rules rules;
if (_expr.item->type() != Operation)
return -1;
for (auto const& rule: rules.rules())
{
rules.resetMatchGroups();
if (rule.first.matches(_expr, *this))
{
// Debug info
//cout << "Simplifying " << *_expr.item << "(";
//for (Id arg: _expr.arguments)
// cout << fullDAGToString(arg) << ", ";
//cout << ")" << endl;
//cout << "with rule " << rule.first.toString() << endl;
//ExpressionTemplate t(rule.second());
//cout << "to" << rule.second().toString() << endl;
return rebuildExpression(ExpressionTemplate(rule.second()));
}
}
if (!_secondRun && _expr.arguments.size() == 2 && SemanticInformation::isCommutativeOperation(*_expr.item))
{
Expression expr = _expr;
swap(expr.arguments[0], expr.arguments[1]);
return tryToSimplify(expr, true);
}
return -1;
}
ExpressionClasses::Id ExpressionClasses::rebuildExpression(ExpressionTemplate const& _template)
{
if (_template.hasId)
return _template.id;
Ids arguments;
for (ExpressionTemplate const& t: _template.arguments)
arguments.push_back(rebuildExpression(t));
m_spareAssemblyItem.push_back(make_shared<AssemblyItem>(_template.item));
return find(*m_spareAssemblyItem.back(), arguments);
}
Pattern::Pattern(Instruction _instruction, std::vector<Pattern> const& _arguments):
m_type(Operation),
m_requireDataMatch(true),
m_data(_instruction),
m_arguments(_arguments)
{
}
void Pattern::setMatchGroup(unsigned _group, map<unsigned, Expression const*>& _matchGroups)
{
m_matchGroup = _group;
m_matchGroups = &_matchGroups;
}
bool Pattern::matches(Expression const& _expr, ExpressionClasses const& _classes) const
{
if (!matchesBaseItem(*_expr.item))
return false;
if (m_matchGroup)
{
if (!m_matchGroups->count(m_matchGroup))
(*m_matchGroups)[m_matchGroup] = &_expr;
else if ((*m_matchGroups)[m_matchGroup]->id != _expr.id)
return false;
}
assertThrow(m_arguments.size() == 0 || _expr.arguments.size() == m_arguments.size(), OptimizerException, "");
for (size_t i = 0; i < m_arguments.size(); ++i)
if (!m_arguments[i].matches(_classes.representative(_expr.arguments[i]), _classes))
return false;
return true;
}
string Pattern::toString() const
{
stringstream s;
switch (m_type)
{
case Operation:
s << instructionInfo(Instruction(unsigned(m_data))).name;
break;
case Push:
s << "PUSH " << hex << m_data;
break;
case UndefinedItem:
s << "ANY";
break;
default:
s << "t=" << dec << m_type << " d=" << hex << m_data;
break;
}
if (!m_requireDataMatch)
s << " ~";
if (m_matchGroup)
s << "[" << dec << m_matchGroup << "]";
s << "(";
for (Pattern const& p: m_arguments)
s << p.toString() << ", ";
s << ")";
return s.str();
}
bool Pattern::matchesBaseItem(AssemblyItem const& _item) const
{
if (m_type == UndefinedItem)
return true;
if (m_type != _item.type())
return false;
if (m_requireDataMatch && m_data != _item.data())
return false;
return true;
}
Pattern::Expression const& Pattern::matchGroupValue() const
{
assertThrow(m_matchGroup > 0, OptimizerException, "");
assertThrow(!!m_matchGroups, OptimizerException, "");
assertThrow((*m_matchGroups)[m_matchGroup], OptimizerException, "");
return *(*m_matchGroups)[m_matchGroup];
}
ExpressionTemplate::ExpressionTemplate(Pattern const& _pattern)
{
if (_pattern.matchGroup())
{
hasId = true;
id = _pattern.id();
}
else
{
hasId = false;
item = _pattern.toAssemblyItem();
}
for (auto const& arg: _pattern.arguments())
arguments.push_back(ExpressionTemplate(arg));
}
string ExpressionTemplate::toString() const
{
stringstream s;
if (hasId)
s << id;
else
s << item;
s << "(";
for (auto const& arg: arguments)
s << arg.toString();
s << ")";
return s.str();
}

150
libevmcore/ExpressionClasses.h

@ -0,0 +1,150 @@
/*
This file is part of cpp-ethereum.
cpp-ethereum is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
cpp-ethereum is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with cpp-ethereum. If not, see <http://www.gnu.org/licenses/>.
*/
/**
* @file ExpressionClasses.h
* @author Christian <c@ethdev.com>
* @date 2015
* Container for equivalence classes of expressions for use in common subexpression elimination.
*/
#pragma once
#include <vector>
#include <map>
#include <memory>
#include <libdevcore/Common.h>
#include <libevmcore/AssemblyItem.h>
namespace dev
{
namespace eth
{
class Pattern;
struct ExpressionTemplate;
/**
* Collection of classes of equivalent expressions that can also determine the class of an expression.
* Identifiers are contiguously assigned to new classes starting from zero.
*/
class ExpressionClasses
{
public:
using Id = unsigned;
using Ids = std::vector<Id>;
struct Expression
{
Id id;
AssemblyItem const* item;
Ids arguments;
bool operator<(Expression const& _other) const;
};
/// Retrieves the id of the expression equivalence class resulting from the given item applied to the
/// given classes, might also create a new one.
Id find(AssemblyItem const& _item, Ids const& _arguments = {});
/// @returns the canonical representative of an expression class.
Expression const& representative(Id _id) const { return m_representatives.at(_id); }
/// @returns the number of classes.
Id size() const { return m_representatives.size(); }
std::string fullDAGToString(Id _id) const;
private:
/// Tries to simplify the given expression.
/// @returns its class if it possible or Id(-1) otherwise.
/// @param _secondRun is set to true for the second run where arguments of commutative expressions are reversed
Id tryToSimplify(Expression const& _expr, bool _secondRun = false);
/// Rebuilds an expression from a (matched) pattern.
Id rebuildExpression(ExpressionTemplate const& _template);
std::vector<std::pair<Pattern, std::function<Pattern()>>> createRules() const;
/// Expression equivalence class representatives - we only store one item of an equivalence.
std::vector<Expression> m_representatives;
std::vector<std::shared_ptr<AssemblyItem>> m_spareAssemblyItem;
};
/**
* Pattern to match against an expression.
* Also stores matched expressions to retrieve them later, for constructing new expressions using
* ExpressionTemplate.
*/
class Pattern
{
public:
using Expression = ExpressionClasses::Expression;
using Id = ExpressionClasses::Id;
// Matches a specific constant value.
Pattern(unsigned _value): Pattern(u256(_value)) {}
// Matches a specific constant value.
Pattern(u256 const& _value): m_type(Push), m_requireDataMatch(true), m_data(_value) {}
// Matches a specific assembly item type or anything if not given.
Pattern(AssemblyItemType _type = UndefinedItem): m_type(_type) {}
// Matches a given instruction with given arguments
Pattern(Instruction _instruction, std::vector<Pattern> const& _arguments = {});
/// Sets this pattern to be part of the match group with the identifier @a _group.
/// Inside one rule, all patterns in the same match group have to match expressions from the
/// same expression equivalence class.
void setMatchGroup(unsigned _group, std::map<unsigned, Expression const*>& _matchGroups);
unsigned matchGroup() const { return m_matchGroup; }
bool matches(Expression const& _expr, ExpressionClasses const& _classes) const;
AssemblyItem toAssemblyItem() const { return AssemblyItem(m_type, m_data); }
std::vector<Pattern> arguments() const { return m_arguments; }
/// @returns the id of the matched expression if this pattern is part of a match group.
Id id() const { return matchGroupValue().id; }
/// @returns the data of the matched expression if this pattern is part of a match group.
u256 d() const { return matchGroupValue().item->data(); }
std::string toString() const;
private:
bool matchesBaseItem(AssemblyItem const& _item) const;
Expression const& matchGroupValue() const;
AssemblyItemType m_type;
bool m_requireDataMatch = false;
u256 m_data = 0;
std::vector<Pattern> m_arguments;
unsigned m_matchGroup = 0;
std::map<unsigned, Expression const*>* m_matchGroups = nullptr;
};
/**
* Template for a new expression that can be built from matched patterns.
*/
struct ExpressionTemplate
{
using Expression = ExpressionClasses::Expression;
using Id = ExpressionClasses::Id;
explicit ExpressionTemplate(Pattern const& _pattern);
std::string toString() const;
bool hasId = false;
/// Id of the matched expression, if available.
Id id = Id(-1);
// Otherwise, assembly item.
AssemblyItem item = UndefinedItem;
std::vector<ExpressionTemplate> arguments;
};
}
}

115
test/SolidityOptimizer.cpp

@ -74,6 +74,14 @@ public:
"\nOptimized: " + toHex(optimizedOutput));
}
void checkCSE(AssemblyItems const& _input, AssemblyItems const& _expectation)
{
eth::CommonSubexpressionEliminator cse;
BOOST_REQUIRE(cse.feedItems(_input.begin(), _input.end()) == _input.end());
AssemblyItems output = cse.getOptimizedItems();
BOOST_CHECK_EQUAL_COLLECTIONS(_expectation.begin(), _expectation.end(), output.begin(), output.end());
}
protected:
Address m_optimizedContract;
Address m_nonOptimizedContract;
@ -199,61 +207,100 @@ BOOST_AUTO_TEST_CASE(cse_intermediate_swap)
BOOST_AUTO_TEST_CASE(cse_negative_stack_access)
{
eth::CommonSubexpressionEliminator cse;
AssemblyItems input{AssemblyItem(Instruction::DUP2), AssemblyItem(u256(0))};
BOOST_REQUIRE(cse.feedItems(input.begin(), input.end()) == input.end());
AssemblyItems output = cse.getOptimizedItems();
BOOST_CHECK_EQUAL_COLLECTIONS(input.begin(), input.end(), output.begin(), output.end());
AssemblyItems input{Instruction::DUP2, u256(0)};
checkCSE(input, input);
}
BOOST_AUTO_TEST_CASE(cse_negative_stack_end)
{
eth::CommonSubexpressionEliminator cse;
AssemblyItems input{Instruction::ADD};
checkCSE(input, input);
}
BOOST_AUTO_TEST_CASE(cse_intermediate_negative_stack)
{
AssemblyItems input{Instruction::ADD, u256(1), Instruction::DUP1};
checkCSE(input, input);
}
BOOST_AUTO_TEST_CASE(cse_pop)
{
checkCSE({Instruction::POP}, {Instruction::POP});
}
BOOST_AUTO_TEST_CASE(cse_unneeded_items)
{
AssemblyItems input{
AssemblyItem(Instruction::ADD)
Instruction::ADD,
Instruction::SWAP1,
Instruction::POP,
u256(7),
u256(8),
};
BOOST_REQUIRE(cse.feedItems(input.begin(), input.end()) == input.end());
AssemblyItems output = cse.getOptimizedItems();
BOOST_CHECK_EQUAL_COLLECTIONS(input.begin(), input.end(), output.begin(), output.end());
checkCSE(input, input);
}
BOOST_AUTO_TEST_CASE(cse_intermediate_negative_stack)
BOOST_AUTO_TEST_CASE(cse_constant_addition)
{
AssemblyItems input{u256(7), u256(8), Instruction::ADD};
checkCSE(input, {u256(7 + 8)});
}
BOOST_AUTO_TEST_CASE(cse_invariants)
{
eth::CommonSubexpressionEliminator cse;
AssemblyItems input{
AssemblyItem(Instruction::ADD),
AssemblyItem(u256(1)),
AssemblyItem(Instruction::DUP2)
Instruction::DUP1,
Instruction::DUP1,
u256(0),
Instruction::OR,
Instruction::OR
};
BOOST_REQUIRE(cse.feedItems(input.begin(), input.end()) == input.end());
AssemblyItems output = cse.getOptimizedItems();
BOOST_CHECK_EQUAL_COLLECTIONS(input.begin(), input.end(), output.begin(), output.end());
checkCSE(input, {Instruction::DUP1});
}
BOOST_AUTO_TEST_CASE(cse_pop)
BOOST_AUTO_TEST_CASE(cse_subself)
{
checkCSE({Instruction::DUP1, Instruction::SUB}, {Instruction::POP, u256(0)});
}
BOOST_AUTO_TEST_CASE(cse_subother)
{
checkCSE({Instruction::SUB}, {Instruction::SUB});
}
BOOST_AUTO_TEST_CASE(cse_double_negation)
{
checkCSE({Instruction::DUP5, Instruction::NOT, Instruction::NOT}, {Instruction::DUP5});
}
BOOST_AUTO_TEST_CASE(cse_associativity)
{
eth::CommonSubexpressionEliminator cse;
AssemblyItems input{
AssemblyItem(Instruction::POP)
Instruction::DUP1,
Instruction::DUP1,
u256(0),
Instruction::OR,
Instruction::OR
};
BOOST_REQUIRE(cse.feedItems(input.begin(), input.end()) == input.end());
AssemblyItems output = cse.getOptimizedItems();
BOOST_CHECK_EQUAL_COLLECTIONS(input.begin(), input.end(), output.begin(), output.end());
checkCSE(input, {Instruction::DUP1});
}
BOOST_AUTO_TEST_CASE(cse_unneeded_items)
BOOST_AUTO_TEST_CASE(cse_associativity2)
{
eth::CommonSubexpressionEliminator cse;
AssemblyItems input{
AssemblyItem(Instruction::ADD),
AssemblyItem(Instruction::SWAP1),
AssemblyItem(Instruction::POP),
AssemblyItem(u256(7)),
AssemblyItem(u256(8)),
u256(0),
Instruction::DUP2,
u256(2),
u256(1),
Instruction::DUP6,
Instruction::ADD,
u256(2),
Instruction::ADD,
Instruction::ADD,
Instruction::ADD,
Instruction::ADD
};
BOOST_REQUIRE(cse.feedItems(input.begin(), input.end()) == input.end());
AssemblyItems output = cse.getOptimizedItems();
BOOST_CHECK_EQUAL_COLLECTIONS(input.begin(), input.end(), output.begin(), output.end());
checkCSE(input, {Instruction::DUP2, Instruction::DUP2, Instruction::ADD, u256(5), Instruction::ADD});
}
BOOST_AUTO_TEST_SUITE_END()

Loading…
Cancel
Save