Browse Source

Implementation of MUL workaround in "LLVM pass" way.

cl-refactor
Paweł Bylica 10 years ago
parent
commit
463025e15d
  1. 125
      libevmjit/Arith256.cpp
  2. 5
      libevmjit/Arith256.h
  3. 2
      libevmjit/Compiler.cpp
  4. 2
      libevmjit/ExecutionEngine.cpp
  5. 59
      libevmjit/Optimizer.cpp
  6. 2
      libevmjit/Optimizer.h

125
libevmjit/Arith256.cpp

@ -4,6 +4,7 @@
#include <iomanip>
#include "preprocessor/llvm_includes_start.h"
#include <llvm/IR/Module.h>
#include <llvm/IR/IntrinsicInst.h>
#include "preprocessor/llvm_includes_end.h"
@ -32,57 +33,56 @@ void Arith256::debug(llvm::Value* _value, char _c)
createCall(m_debug, {m_builder.CreateZExtOrTrunc(_value, Type::Word), m_builder.getInt8(_c)});
}
llvm::Function* Arith256::getMulFunc()
llvm::Function* Arith256::getMulFunc(llvm::Module& _module)
{
auto& func = m_mul;
if (!func)
{
llvm::Type* argTypes[] = {Type::Word, Type::Word};
func = llvm::Function::Create(llvm::FunctionType::get(Type::Word, argTypes, false), llvm::Function::PrivateLinkage, "mul", getModule());
func->setDoesNotThrow();
func->setDoesNotAccessMemory();
auto x = &func->getArgumentList().front();
x->setName("x");
auto y = x->getNextNode();
y->setName("y");
InsertPointGuard guard{m_builder};
auto bb = llvm::BasicBlock::Create(m_builder.getContext(), {}, func);
m_builder.SetInsertPoint(bb);
auto i64 = Type::Size;
auto i128 = m_builder.getIntNTy(128);
auto i256 = Type::Word;
auto c64 = Constant::get(64);
auto c128 = Constant::get(128);
auto c192 = Constant::get(192);
auto x_lo = m_builder.CreateTrunc(x, i64, "x.lo");
auto y_lo = m_builder.CreateTrunc(y, i64, "y.lo");
auto x_mi = m_builder.CreateTrunc(m_builder.CreateLShr(x, c64), i64);
auto y_mi = m_builder.CreateTrunc(m_builder.CreateLShr(y, c64), i64);
auto x_hi = m_builder.CreateTrunc(m_builder.CreateLShr(x, c128), i128);
auto y_hi = m_builder.CreateTrunc(m_builder.CreateLShr(y, c128), i128);
auto t1 = m_builder.CreateMul(m_builder.CreateZExt(x_lo, i128), m_builder.CreateZExt(y_lo, i128));
auto t2 = m_builder.CreateMul(m_builder.CreateZExt(x_lo, i128), m_builder.CreateZExt(y_mi, i128));
auto t3 = m_builder.CreateMul(m_builder.CreateZExt(x_lo, i128), y_hi);
auto t4 = m_builder.CreateMul(m_builder.CreateZExt(x_mi, i128), m_builder.CreateZExt(y_lo, i128));
auto t5 = m_builder.CreateMul(m_builder.CreateZExt(x_mi, i128), m_builder.CreateZExt(y_mi, i128));
auto t6 = m_builder.CreateMul(m_builder.CreateZExt(x_mi, i128), y_hi);
auto t7 = m_builder.CreateMul(x_hi, m_builder.CreateZExt(y_lo, i128));
auto t8 = m_builder.CreateMul(x_hi, m_builder.CreateZExt(y_mi, i128));
auto p = m_builder.CreateZExt(t1, i256);
p = m_builder.CreateAdd(p, m_builder.CreateShl(m_builder.CreateZExt(t2, i256), c64));
p = m_builder.CreateAdd(p, m_builder.CreateShl(m_builder.CreateZExt(t3, i256), c128));
p = m_builder.CreateAdd(p, m_builder.CreateShl(m_builder.CreateZExt(t4, i256), c64));
p = m_builder.CreateAdd(p, m_builder.CreateShl(m_builder.CreateZExt(t5, i256), c128));
p = m_builder.CreateAdd(p, m_builder.CreateShl(m_builder.CreateZExt(t6, i256), c192));
p = m_builder.CreateAdd(p, m_builder.CreateShl(m_builder.CreateZExt(t7, i256), c128));
p = m_builder.CreateAdd(p, m_builder.CreateShl(m_builder.CreateZExt(t8, i256), c192));
m_builder.CreateRet(p);
}
static const auto funcName = "evm.mul.i256";
if (auto func = _module.getFunction(funcName))
return func;
llvm::Type* argTypes[] = {Type::Word, Type::Word};
auto func = llvm::Function::Create(llvm::FunctionType::get(Type::Word, argTypes, false), llvm::Function::PrivateLinkage, funcName, &_module);
func->setDoesNotThrow();
func->setDoesNotAccessMemory();
auto x = &func->getArgumentList().front();
x->setName("x");
auto y = x->getNextNode();
y->setName("y");
auto bb = llvm::BasicBlock::Create(_module.getContext(), {}, func);
auto builder = llvm::IRBuilder<>{bb};
auto i64 = Type::Size;
auto i128 = builder.getIntNTy(128);
auto i256 = Type::Word;
auto c64 = Constant::get(64);
auto c128 = Constant::get(128);
auto c192 = Constant::get(192);
auto x_lo = builder.CreateTrunc(x, i64, "x.lo");
auto y_lo = builder.CreateTrunc(y, i64, "y.lo");
auto x_mi = builder.CreateTrunc(builder.CreateLShr(x, c64), i64);
auto y_mi = builder.CreateTrunc(builder.CreateLShr(y, c64), i64);
auto x_hi = builder.CreateTrunc(builder.CreateLShr(x, c128), i128);
auto y_hi = builder.CreateTrunc(builder.CreateLShr(y, c128), i128);
auto t1 = builder.CreateMul(builder.CreateZExt(x_lo, i128), builder.CreateZExt(y_lo, i128));
auto t2 = builder.CreateMul(builder.CreateZExt(x_lo, i128), builder.CreateZExt(y_mi, i128));
auto t3 = builder.CreateMul(builder.CreateZExt(x_lo, i128), y_hi);
auto t4 = builder.CreateMul(builder.CreateZExt(x_mi, i128), builder.CreateZExt(y_lo, i128));
auto t5 = builder.CreateMul(builder.CreateZExt(x_mi, i128), builder.CreateZExt(y_mi, i128));
auto t6 = builder.CreateMul(builder.CreateZExt(x_mi, i128), y_hi);
auto t7 = builder.CreateMul(x_hi, builder.CreateZExt(y_lo, i128));
auto t8 = builder.CreateMul(x_hi, builder.CreateZExt(y_mi, i128));
auto p = builder.CreateZExt(t1, i256);
p = builder.CreateAdd(p, builder.CreateShl(builder.CreateZExt(t2, i256), c64));
p = builder.CreateAdd(p, builder.CreateShl(builder.CreateZExt(t3, i256), c128));
p = builder.CreateAdd(p, builder.CreateShl(builder.CreateZExt(t4, i256), c64));
p = builder.CreateAdd(p, builder.CreateShl(builder.CreateZExt(t5, i256), c128));
p = builder.CreateAdd(p, builder.CreateShl(builder.CreateZExt(t6, i256), c192));
p = builder.CreateAdd(p, builder.CreateShl(builder.CreateZExt(t7, i256), c128));
p = builder.CreateAdd(p, builder.CreateShl(builder.CreateZExt(t8, i256), c192));
builder.CreateRet(p);
return func;
}
@ -112,10 +112,11 @@ llvm::Function* Arith256::getMul512Func()
auto x_hi = m_builder.CreateZExt(m_builder.CreateTrunc(m_builder.CreateLShr(x, Constant::get(128)), i128, "x.hi"), i256);
auto y_hi = m_builder.CreateZExt(m_builder.CreateTrunc(m_builder.CreateLShr(y, Constant::get(128)), i128, "y.hi"), i256);
auto t1 = createCall(getMulFunc(), {x_lo, y_lo});
auto t2 = createCall(getMulFunc(), {x_lo, y_hi});
auto t3 = createCall(getMulFunc(), {x_hi, y_lo});
auto t4 = createCall(getMulFunc(), {x_hi, y_hi});
auto mul256Func = getMulFunc(*getModule());
auto t1 = createCall(mul256Func, {x_lo, y_lo});
auto t2 = createCall(mul256Func, {x_lo, y_hi});
auto t3 = createCall(mul256Func, {x_hi, y_lo});
auto t4 = createCall(mul256Func, {x_hi, y_hi});
auto p = m_builder.CreateZExt(t1, i512);
p = m_builder.CreateAdd(p, m_builder.CreateShl(m_builder.CreateZExt(t2, i512), m_builder.getIntN(512, 128)));
@ -260,14 +261,15 @@ llvm::Function* Arith256::getExpFunc()
m_builder.CreateCondBr(eOdd, updateBB, continueBB);
m_builder.SetInsertPoint(updateBB);
auto r0 = createCall(getMulFunc(), {r, b});
auto mul256Func = getMulFunc(*getModule());
auto r0 = createCall(mul256Func, {r, b});
m_builder.CreateBr(continueBB);
m_builder.SetInsertPoint(continueBB);
auto r1 = m_builder.CreatePHI(Type::Word, 2, "r1");
r1->addIncoming(r, bodyBB);
r1->addIncoming(r0, updateBB);
auto b1 = createCall(getMulFunc(), {b, b});
auto b1 = createCall(mul256Func, {b, b});
auto e1 = m_builder.CreateLShr(e, Constant::get(1), "e1");
m_builder.CreateBr(headerBB);
@ -347,17 +349,6 @@ llvm::Function* Arith256::getMulModFunc()
return m_mulmod;
}
llvm::Value* Arith256::mul(llvm::Value* _arg1, llvm::Value* _arg2)
{
if (auto c1 = llvm::dyn_cast<llvm::ConstantInt>(_arg1))
{
if (auto c2 = llvm::dyn_cast<llvm::ConstantInt>(_arg2))
return Constant::get(c1->getValue() * c2->getValue());
}
return createCall(getMulFunc(), {_arg1, _arg2});
}
std::pair<llvm::Value*, llvm::Value*> Arith256::div(llvm::Value* _arg1, llvm::Value* _arg2)
{
if (auto c1 = llvm::dyn_cast<llvm::ConstantInt>(_arg1))

5
libevmjit/Arith256.h

@ -14,7 +14,6 @@ class Arith256 : public CompilerHelper
public:
Arith256(llvm::IRBuilder<>& _builder);
llvm::Value* mul(llvm::Value* _arg1, llvm::Value* _arg2);
std::pair<llvm::Value*, llvm::Value*> div(llvm::Value* _arg1, llvm::Value* _arg2);
std::pair<llvm::Value*, llvm::Value*> sdiv(llvm::Value* _arg1, llvm::Value* _arg2);
llvm::Value* exp(llvm::Value* _arg1, llvm::Value* _arg2);
@ -23,15 +22,15 @@ public:
void debug(llvm::Value* _value, char _c);
static llvm::Function* getMulFunc(llvm::Module& _module);
private:
llvm::Function* getMulFunc();
llvm::Function* getMul512Func();
llvm::Function* getDivFunc(llvm::Type* _type);
llvm::Function* getExpFunc();
llvm::Function* getAddModFunc();
llvm::Function* getMulModFunc();
llvm::Function* m_mul = nullptr;
llvm::Function* m_mul512 = nullptr;
llvm::Function* m_div = nullptr;
llvm::Function* m_div512 = nullptr;

2
libevmjit/Compiler.cpp

@ -270,7 +270,7 @@ void Compiler::compileBasicBlock(BasicBlock& _basicBlock, RuntimeManager& _runti
{
auto lhs = stack.pop();
auto rhs = stack.pop();
auto res = _arith.mul(lhs, rhs);
auto res = m_builder.CreateMul(lhs, rhs);
stack.push(res);
break;
}

2
libevmjit/ExecutionEngine.cpp

@ -162,6 +162,8 @@ ReturnCode ExecutionEngine::run(RuntimeData* _data, Env* _env)
listener->stateChanged(ExecState::Optimization);
optimize(*module);
}
prepare(*module);
}
if (g_dump)
module->dump();

59
libevmjit/Optimizer.cpp

@ -1,11 +1,16 @@
#include "Optimizer.h"
#include "preprocessor/llvm_includes_start.h"
#include <llvm/IR/BasicBlock.h>
#include <llvm/IR/Function.h>
#include <llvm/IR/LegacyPassManager.h>
#include <llvm/Transforms/Scalar.h>
#include <llvm/Transforms/IPO.h>
#include "preprocessor/llvm_includes_end.h"
#include "Arith256.h"
#include "Type.h"
namespace dev
{
namespace eth
@ -24,6 +29,60 @@ bool optimize(llvm::Module& _module)
return pm.run(_module);
}
namespace
{
class LowerEVMPass : public llvm::BasicBlockPass
{
static char ID;
bool m_mulFuncNeeded = false;
public:
LowerEVMPass():
llvm::BasicBlockPass(ID)
{}
virtual bool runOnBasicBlock(llvm::BasicBlock& _bb) override;
virtual bool doFinalization(llvm::Module& _module) override;
};
char LowerEVMPass::ID = 0;
bool LowerEVMPass::runOnBasicBlock(llvm::BasicBlock& _bb)
{
auto modified = false;
auto module = _bb.getParent()->getParent();
for (auto&& inst : _bb)
{
if (inst.getOpcode() == llvm::Instruction::Mul)
{
if (inst.getType() == Type::Word)
{
auto call = llvm::CallInst::Create(Arith256::getMulFunc(*module), {inst.getOperand(0), inst.getOperand(1)}, "", &inst);
inst.replaceAllUsesWith(call);
modified = true;
}
}
}
return modified;
}
bool LowerEVMPass::doFinalization(llvm::Module&)
{
return false;
}
}
bool prepare(llvm::Module& _module)
{
auto pm = llvm::legacy::PassManager{};
pm.add(new LowerEVMPass{});
return pm.run(_module);
}
}
}
}

2
libevmjit/Optimizer.h

@ -14,6 +14,8 @@ namespace jit
bool optimize(llvm::Module& _module);
bool prepare(llvm::Module& _module);
}
}
}

Loading…
Cancel
Save