Instructions Bytecode::parser(const Tokens& tokens) const {
Tokens tokens_group;
// Abstract syntax tree.
Instructions ast;
int funcs = 0, params = 0, specs = 0;
for (int i = 0; i < tokens.size(); i++) {
if (tokens[i].type == DELIMITER) {
checkFunctions(tokens_group[0], funcs);
checkById(tokens_group[0].function_id, params, specs);
Instruction inst =
makeInstruction(params, specs, tokens_group);
ast.push_back(inst);
funcs = 0, params = 0, specs = 0;
tokens_group.clear();
} else {
if (tokens[i].type == FUNCTION) {
funcs++;
} else if (tokens[i].type == PARAMETER) {
params++;
} else {
specs++;
}
tokens_group.push_back(tokens[i]);
}
}
return ast;
}
void constructBlocks(Blocks &blocks, Instructions &instructions) {
/* Create the first block containing the very first instruction in this script.
* Then follow the complete code flow from this instruction onwards. */
assert(blocks.empty());
if (instructions.empty())
return;
blocks.push_back(Block(instructions.front().address));
constructBlocks(blocks, blocks.back(), instructions.front());
}
void ConstantInsertExtractElementIndex::fixOutOfRangeConstantIndices(
BasicBlock &BB, const Instructions &Instrs) const {
for (Instructions::const_iterator IB = Instrs.begin(), IE = Instrs.end();
IB != IE; ++IB) {
Instruction *I = *IB;
const APInt &Idx =
cast<ConstantInt>(getInsertExtractElementIdx(I))->getValue();
APInt NumElements = APInt(Idx.getBitWidth(), vectorNumElements(I));
APInt NewIdx = Idx.urem(NumElements);
setInsertExtractElementIdx(I, ConstantInt::get(M->getContext(), NewIdx));
}
}
void ConstantInsertExtractElementIndex::findNonConstantInsertExtractElements(
const BasicBlock &BB, Instructions &OutOfRangeConstantIndices,
Instructions &NonConstantVectorIndices) const {
for (BasicBlock::const_iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE;
++BBI) {
const Instruction *I = &*BBI;
if (Value *Idx = getInsertExtractElementIdx(I)) {
if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
if (!CI->getValue().ult(vectorNumElements(I)))
OutOfRangeConstantIndices.push_back(const_cast<Instruction *>(I));
} else
NonConstantVectorIndices.push_back(const_cast<Instruction *>(I));
}
}
}
bool ConstantInsertExtractElementIndex::runOnBasicBlock(BasicBlock &BB) {
bool Changed = false;
if (!DL)
DL = &getAnalysis<DataLayoutPass>().getDataLayout();
Instructions OutOfRangeConstantIndices;
Instructions NonConstantVectorIndices;
findNonConstantInsertExtractElements(BB, OutOfRangeConstantIndices,
NonConstantVectorIndices);
if (!OutOfRangeConstantIndices.empty()) {
Changed = true;
fixOutOfRangeConstantIndices(BB, OutOfRangeConstantIndices);
}
if (!NonConstantVectorIndices.empty()) {
Changed = true;
fixNonConstantVectorIndices(BB, NonConstantVectorIndices);
}
return Changed;
}
void Bytecode::compile(const Instructions& instructions) {
for (int i = 0; i < instructions.size(); i++) {
int func_id = instructions[i].function.function_id;
int p1 = instructions[i].p1.parameter;
int p2 = instructions[i].p2.parameter;
bool spec = instructions[i].spec.spec;
PackedInstruction packed_inst(func_id, p1, p2, spec);
bytecode_.push_back(packed_inst);
}
}
ClobFile::ClobFile(std::string _path, Instructions _instructions, strings_t _data) {
path = _path;
instructions = _instructions;
header = {0, _instructions.size(), _data.size()};
for (auto _d : _data) {
DataEntry_t entry;
entry.string = _d;
entry.length = entry.string.size();
data.push_back(entry);
}
}
void ConstantInsertExtractElementIndex::fixNonConstantVectorIndices(
BasicBlock &BB, const Instructions &Instrs) const {
for (Instructions::const_iterator IB = Instrs.begin(), IE = Instrs.end();
IB != IE; ++IB) {
Instruction *I = *IB;
Value *Vec = I->getOperand(0);
Value *Idx = getInsertExtractElementIdx(I);
VectorType *VecTy = cast<VectorType>(Vec->getType());
Type *ElemTy = VecTy->getElementType();
unsigned ElemAlign = DL->getPrefTypeAlignment(ElemTy);
unsigned VecAlign = std::max(ElemAlign, DL->getPrefTypeAlignment(VecTy));
IRBuilder<> IRB(I);
AllocaInst *Alloca = IRB.CreateAlloca(
ElemTy, ConstantInt::get(Type::getInt32Ty(M->getContext()),
vectorNumElements(I)));
Alloca->setAlignment(VecAlign);
Value *AllocaAsVec = IRB.CreateBitCast(Alloca, VecTy->getPointerTo());
IRB.CreateAlignedStore(Vec, AllocaAsVec, Alloca->getAlignment());
Value *GEP = IRB.CreateGEP(Alloca, Idx);
Value *Res;
switch (I->getOpcode()) {
default:
llvm_unreachable("expected InsertElement or ExtractElement");
case Instruction::InsertElement:
IRB.CreateAlignedStore(I->getOperand(1), GEP, ElemAlign);
Res = IRB.CreateAlignedLoad(AllocaAsVec, Alloca->getAlignment());
break;
case Instruction::ExtractElement:
Res = IRB.CreateAlignedLoad(GEP, ElemAlign);
break;
}
I->replaceAllUsesWith(Res);
I->eraseFromParent();
}
}
void AArch64AddressTypePromotion::mergeSExts(ValueToInsts &ValToSExtendedUses,
SetOfInstructions &ToRemove) {
DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
for (auto &Entry : ValToSExtendedUses) {
Instructions &Insts = Entry.second;
Instructions CurPts;
for (Instruction *Inst : Insts) {
if (ToRemove.count(Inst))
continue;
bool inserted = false;
for (auto &Pt : CurPts) {
if (DT.dominates(Inst, Pt)) {
DEBUG(dbgs() << "Replace all uses of:\n" << *Pt << "\nwith:\n"
<< *Inst << '\n');
Pt->replaceAllUsesWith(Inst);
ToRemove.insert(Pt);
Pt = Inst;
inserted = true;
break;
}
if (!DT.dominates(Pt, Inst))
// Give up if we need to merge in a common dominator as the
// expermients show it is not profitable.
continue;
DEBUG(dbgs() << "Replace all uses of:\n" << *Inst << "\nwith:\n"
<< *Pt << '\n');
Inst->replaceAllUsesWith(Pt);
ToRemove.insert(Inst);
inserted = true;
break;
}
if (!inserted)
CurPts.push_back(Inst);
}
}
}
/**
@brief opens instructions window after pushing button
*/
void MainWindow::help() {
//open new window
Instructions* instruct = new Instructions(this);
instruct->show();
}
void AArch64AddressTypePromotion::analyzeSExtension(Instructions &SExtInsts) {
DEBUG(dbgs() << "*** Analyze Sign Extensions ***\n");
DenseMap<Value *, Instruction *> SeenChains;
for (auto &BB : *Func) {
for (auto &II : BB) {
Instruction *SExt = &II;
// Collect all sext operation per type.
if (!isa<SExtInst>(SExt) || !shouldConsiderSExt(SExt))
continue;
DEBUG(dbgs() << "Found:\n" << (*SExt) << '\n');
// Cases where we actually perform the optimization:
// 1. SExt is used in a getelementptr with more than 2 operand =>
// likely we can merge some computation if they are done on 64 bits.
// 2. The beginning of the SExt chain is SExt several time. =>
// code sharing is possible.
bool insert = false;
// #1.
for (const User *U : SExt->users()) {
const Instruction *Inst = dyn_cast<GetElementPtrInst>(U);
if (Inst && Inst->getNumOperands() > 2) {
DEBUG(dbgs() << "Interesting use in GetElementPtrInst\n" << *Inst
<< '\n');
insert = true;
break;
}
}
// #2.
// Check the head of the chain.
Instruction *Inst = SExt;
Value *Last;
do {
int OpdIdx = 0;
const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
if (BinOp && isa<ConstantInt>(BinOp->getOperand(0)))
OpdIdx = 1;
Last = Inst->getOperand(OpdIdx);
Inst = dyn_cast<Instruction>(Last);
} while (Inst && canGetThrough(Inst) && shouldGetThrough(Inst));
DEBUG(dbgs() << "Head of the chain:\n" << *Last << '\n');
DenseMap<Value *, Instruction *>::iterator AlreadySeen =
SeenChains.find(Last);
if (insert || AlreadySeen != SeenChains.end()) {
DEBUG(dbgs() << "Insert\n");
SExtInsts.push_back(SExt);
if (AlreadySeen != SeenChains.end() && AlreadySeen->second != nullptr) {
DEBUG(dbgs() << "Insert chain member\n");
SExtInsts.push_back(AlreadySeen->second);
SeenChains[Last] = nullptr;
}
} else {
DEBUG(dbgs() << "Record its chain membership\n");
SeenChains[Last] = SExt;
}
}
}
}
// Input:
// - SExtInsts contains all the sext instructions that are used directly in
// GetElementPtrInst, i.e., access to memory.
// Algorithm:
// - For each sext operation in SExtInsts:
// Let var be the operand of sext.
// while it is profitable (see shouldGetThrough), legal, and safe
// (see canGetThrough) to move sext through var's definition:
// * promote the type of var's definition.
// * fold var into sext uses.
// * move sext above var's definition.
// * update sext operand to use the operand of var that should be sign
// extended (by construction there is only one).
//
// E.g.,
// a = ... i32 c, 3
// b = sext i32 a to i64 <- is it legal/safe/profitable to get through 'a'
// ...
// = b
// => Yes, update the code
// b = sext i32 c to i64
// a = ... i64 b, 3
// ...
// = a
// Iterate on 'c'.
bool
AArch64AddressTypePromotion::propagateSignExtension(Instructions &SExtInsts) {
DEBUG(dbgs() << "*** Propagate Sign Extension ***\n");
bool LocalChange = false;
SetOfInstructions ToRemove;
ValueToInsts ValToSExtendedUses;
while (!SExtInsts.empty()) {
// Get through simple chain.
Instruction *SExt = SExtInsts.pop_back_val();
DEBUG(dbgs() << "Consider:\n" << *SExt << '\n');
// If this SExt has already been merged continue.
if (SExt->use_empty() && ToRemove.count(SExt)) {
DEBUG(dbgs() << "No uses => marked as delete\n");
continue;
}
// Now try to get through the chain of definitions.
while (auto *Inst = dyn_cast<Instruction>(SExt->getOperand(0))) {
DEBUG(dbgs() << "Try to get through:\n" << *Inst << '\n');
if (!canGetThrough(Inst) || !shouldGetThrough(Inst)) {
// We cannot get through something that is not an Instruction
// or not safe to SExt.
DEBUG(dbgs() << "Cannot get through\n");
break;
}
LocalChange = true;
// If this is a sign extend, it becomes useless.
if (isa<SExtInst>(Inst) || isa<TruncInst>(Inst)) {
DEBUG(dbgs() << "SExt or trunc, mark it as to remove\n");
// We cannot use replaceAllUsesWith here because we may trigger some
// assertion on the type as all involved sext operation may have not
// been moved yet.
while (!Inst->use_empty()) {
Use &U = *Inst->use_begin();
Instruction *User = dyn_cast<Instruction>(U.getUser());
assert(User && "User of sext is not an Instruction!");
User->setOperand(U.getOperandNo(), SExt);
}
ToRemove.insert(Inst);
SExt->setOperand(0, Inst->getOperand(0));
SExt->moveBefore(Inst);
continue;
}
// Get through the Instruction:
// 1. Update its type.
// 2. Replace the uses of SExt by Inst.
// 3. Sign extend each operand that needs to be sign extended.
// Step #1.
Inst->mutateType(SExt->getType());
// Step #2.
SExt->replaceAllUsesWith(Inst);
// Step #3.
Instruction *SExtForOpnd = SExt;
DEBUG(dbgs() << "Propagate SExt to operands\n");
for (int OpIdx = 0, EndOpIdx = Inst->getNumOperands(); OpIdx != EndOpIdx;
++OpIdx) {
DEBUG(dbgs() << "Operand:\n" << *(Inst->getOperand(OpIdx)) << '\n');
if (Inst->getOperand(OpIdx)->getType() == SExt->getType() ||
!shouldSExtOperand(Inst, OpIdx)) {
DEBUG(dbgs() << "No need to propagate\n");
continue;
}
// Check if we can statically sign extend the operand.
Value *Opnd = Inst->getOperand(OpIdx);
if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
DEBUG(dbgs() << "Statically sign extend\n");
Inst->setOperand(OpIdx, ConstantInt::getSigned(SExt->getType(),
Cst->getSExtValue()));
continue;
}
// UndefValue are typed, so we have to statically sign extend them.
if (isa<UndefValue>(Opnd)) {
DEBUG(dbgs() << "Statically sign extend\n");
Inst->setOperand(OpIdx, UndefValue::get(SExt->getType()));
continue;
}
// Otherwise we have to explicity sign extend it.
assert(SExtForOpnd &&
"Only one operand should have been sign extended");
SExtForOpnd->setOperand(0, Opnd);
DEBUG(dbgs() << "Move before:\n" << *Inst << "\nSign extend\n");
// Move the sign extension before the insertion point.
SExtForOpnd->moveBefore(Inst);
Inst->setOperand(OpIdx, SExtForOpnd);
// If more sext are required, new instructions will have to be created.
SExtForOpnd = nullptr;
}
if (SExtForOpnd == SExt) {
DEBUG(dbgs() << "Sign extension is useless now\n");
ToRemove.insert(SExt);
break;
}
}
// If the use is already of the right type, connect its uses to its argument
// and delete it.
// This can happen for an Instruction all uses of which are sign extended.
if (!ToRemove.count(SExt) &&
SExt->getType() == SExt->getOperand(0)->getType()) {
DEBUG(dbgs() << "Sign extension is useless, attach its use to "
"its argument\n");
SExt->replaceAllUsesWith(SExt->getOperand(0));
ToRemove.insert(SExt);
} else
ValToSExtendedUses[SExt->getOperand(0)].push_back(SExt);
}
if (EnableMerge)
mergeSExts(ValToSExtendedUses, ToRemove);
// Remove all instructions marked as ToRemove.
for (Instruction *I: ToRemove)
I->eraseFromParent();
return LocalChange;
}
// Pre-Processamento:
// Remove os comentarios e linhas em branco
// E coloca o codigo numa estrutura do tipo CodeLines
// Tambem verifica os labels e equs e ifs
void readAndPreProcess (const char* fileName) {
ifstream infile(fileName);
string line;
int textMemAddr = textStartAddress;
int dataMemAddr = 0;
int BSSMemAddr = 0;
stringstream tempSS;
CodeSection codeSection = NONE;
// Le linha a linha
for (int lineCount = 1; getline(infile, line); ++lineCount) {
// Troca virgulas por espaco
strReplace(line, ",", " ");
// Ignora linhas em branco
if (line.empty()) continue;
// Pega palavra a palavra de acordo com os espacos
istringstream iss(line);
string tempStr;
while (iss >> tempStr) {
if ("SECTION" == tempStr) {
string tempStr2;
iss >> tempStr2;
if ("TEXT" == tempStr2)
codeSection = TEXT;
else if ("DATA" == tempStr2)
codeSection = DATA;
codeLines[lineCount].push_back(tempStr);
codeLines[lineCount].push_back(tempStr2);
continue;
}
// Ignora comentarios
if (";" == tempStr.substr(0,1)) break;
// Desconsidera o caso (maiusculas/minusculas)
transform(tempStr.begin(), tempStr.end(), tempStr.begin(), ::toupper);
// Ve se eh um label / define
if (":" == tempStr.substr(tempStr.length() - 1, 1)) {
// Ve se ainda restam tokens na linha
if (iss.rdbuf()->in_avail() != 0) {
// Remove o ':'
tempStr = tempStr.substr(0, tempStr.length() - 1);
string tempStr2;
iss >> tempStr2;
// Ve se o proximo token eh EQU
if ("EQU" == tempStr2) {
string tempStr3;
iss >> tempStr3;
// Se define já existe
if (defines.find(tempStr3) != defines.end()){
tempSS << lineCount;
errors.push("ERRO NA LINHA " + tempSS.str() + ": EQU ja declarado.");
tempSS.str("");
} else {
// Coloca o valor do EQU na tabela de defines
defines[tempStr] = tempStr3;
}
// Se nao eh so um label
// Com algo a mais na mesma linha
} else {
if ( (labels.find(tempStr) != labels.end()) ||
(dataLabels.find(tempStr) != dataLabels.end()) ){
tempSS << lineCount;
errors.push("ERRO NA LINHA " + tempSS.str() + ": Label ja declarado.");
tempSS.str("");
} else {
// Adiciona na tabela de labels
if(codeSection == TEXT){
labels[tempStr] = textMemAddr;
} else if (codeSection == DATA) {
dataLabels[tempStr] = dataMemAddr;
dataMemAddr += 4;
}
}
// Adiciona endereco de memoria
if (instructions.find(tempStr2) != instructions.end())
textMemAddr += get<3>(instructions[tempStr2]);
// Adiciona os tokens ao vetor
codeLines[lineCount].push_back(tempStr+":");
codeLines[lineCount].push_back(tempStr2);
}
// Se nao eh um label "sozinho"
// Adiciona no vetor
} else {
// Syntax of instruction
//
// mnemonic {op} '|' ( prefix )* opcode args
//
// mnemonic = Intel instruction lexium
//
// op = sr %es, ..., %gs
// xb %al, .., %bh
// xd %ax, .., %dil
// xw %eax, ..., %edi
// xq %rax, ..., %rdi
// rb %rNb
// rd %rNd
// rw %rNw
// rq %rN
// st %stN
// mm %mmN
// xmm %xmmN
// ymm %ymmN
// zmm %zmmN
// cr %crN
// dr %drN
// k %kN
// addr.disp.rq, 10(%r10)
// addr.disp.xq, 20(%rdx)
// addr.disp.rq.rq, 30(%r10,%r11,2)
// addr.disp.rq.xq, 40(%r10,%rdx,2)
// addr.disp.xq.rq, 50(%rdx,%r11,2)
// addr.disp.xq.xq, 60(%rdx,%rax,2)
// addr.disp.rd, 70(%r10d)
// addr.disp.xd, 80(%dx)
// addr.disp.rd.rd, 90(%r10d,%r11d,2)
// addr.disp.rd.xd, 10(%r10d,%dx,2)
// addr.disp.xd.rd, 11(%dx,%r11d,2)
// addr.disp.xd.xd, 12(%dx,%ax,2)
// addr.disp.rw, 13(%r10w)
// addr.disp.xw, 14(%edx)
// addr.disp.rw.rw, 15(%r10w,%r11w,1)
// addr.disp.rw.xw, 16(%r10w,%edx,2)
// addr.disp.xw.rw, 17(%edx,%r11w,4)
// addr.disp.xw.xw, 18(%edx,%eax,8)
// addr.rq, (%r10)
// addr.xq, (%rdx)
// addr.rq.rq, (%r10,%r11,2)
// addr.rq.xq, (%r10,%rdx,2)
// addr.xq.rq, (%rdx,%r11,2)
// addr.xq.xq, (%rdx,%rax,2)
// addr.rd, (%r10d)
// addr.xd, (%dx)
// addr.rd.rd, (%r10d,%r11d,2)
// addr.rd.xd, (%r10d,%dx,2)
// addr.xd.rd, (%dx,%r11d,2)
// addr.xd.xd, (%dx,%ax,2)
// addr.rw, (%r10w)
// addr.xw, (%edx)
// addr.rw.rw, (%r10w,%r11w,1)
// addr.rw.xw, (%r10w,%edx,2)
// addr.xw.rw, (%edx,%r11w,4)
// addr.xw.xw, (%edx,%eax,8)
//
// imm8 Immediate value
// imm32 Immediate value
// disp Memory addressing displayment as the exponent of power of 2
//
// prefix = rex:BRWX
// vex2:R:vvvv:L:pp
// vex3:RXB:mmmm:W:vvvv:L:pp
// evex:RXB:R':mm:W:vvvv:pp:z:L':L:b:V':aaa
// es
// cs
// ss
// ds
// os
// as
// rep
// ne
// lock
// bhnt
// bnht
//
// opcode = one or more bytes
// args = reg:<len>:<n> | imm:<len>:<n> | scale:<len>:<n>
//
bool loadDef( FILE * fh, Instructions & insts )
{
const int BUF_SIZE = 256;
char buff[BUF_SIZE];
enum Pstate
{
start,
nmemonicSeen,
operandSeen,
sepSeen,
opcodeSeen,
argsSeen,
};
static const char * opTypes[] =
{
"addr.disp32.rd",
"addr.disp32.rd.rd",
"addr.disp32.rd.rd.scale",
"addr.disp32.rd.xd",
"addr.disp32.rd.xd.scale",
"addr.disp32.rq",
"addr.disp32.rq.rq",
"addr.disp32.rq.rq.scale",
"addr.disp32.rq.xq",
"addr.disp32.rq.xq.scale",
"addr.disp32.xd",
"addr.disp32.xd.rd",
"addr.disp32.xd.rd.scale",
"addr.disp32.xd.xd",
"addr.disp32.xd.xd.scale",
"addr.disp32.xq",
"addr.disp32.xq.rq",
"addr.disp32.xq.rq.scale",
"addr.disp32.xq.xq",
"addr.disp32.xq.xq.scale",
"addr.disp8.rd",
"addr.disp8.rd.rd",
"addr.disp8.rd.rd.scale",
"addr.disp8.rd.xd",
"addr.disp8.rd.xd.scale",
"addr.disp8.rq",
"addr.disp8.rq.rq",
"addr.disp8.rq.rq.scale",
"addr.disp8.rq.xq",
"addr.disp8.rq.xq.scale",
"addr.disp8.xd",
"addr.disp8.xd.rd",
"addr.disp8.xd.rd.scale",
"addr.disp8.xd.xd",
"addr.disp8.xd.xd.scale",
"addr.disp8.xq",
"addr.disp8.xq.rq",
"addr.disp8.xq.rq.scale",
"addr.disp8.xq.xq",
"addr.disp8.xq.xq.scale",
"addr.rd",
"addr.rd.rd",
"addr.rd.rd.scale",
"addr.rd.xd",
"addr.rd.xd.scale",
"addr.rq",
"addr.rq.rq",
"addr.rq.rq.scale",
"addr.rq.xq",
"addr.rq.xq.scale",
"addr.xd",
"addr.xd.rd",
"addr.xd.rd.scale",
"addr.xd.xd",
"addr.xd.xd.scale",
"addr.xq",
"addr.xq.rq",
"addr.xq.rq.scale",
"addr.xq.xq",
"addr.xq.xq.scale",
"cr",
"dr",
"imm32",
"imm8",
"k",
"mm",
"rb",
"rd",
"rq",
"rw",
"sr",
"st",
"xb",
"xb64",
"xd",
"xmm",
"xq",
"xw",
"ymm",
"zmm",
};
while( fgets( buff, BUF_SIZE, fh ) )
{
vector<string> tokens = split(buff);
if( tokens.size() > 0 )
{
if( tokens.size() > 1 )
{
try
{
Pstate state = start;
Nmemonic nm;
vector<Otype> operands;
vector<unique_ptr<Prefix>> prefixes;
vector<uint8_t> opcode;
set<MC_Comp,Instruction::Comp>args;
int offset = 0;
for( auto & token:tokens)
{
switch(state)
{
case start:
nm = Instruction::nmemonic(token);
state = nmemonicSeen;
break;
case nmemonicSeen:
{
static auto end = opTypes + sizeof(opTypes)/sizeof(char *);
auto p = equal_range( opTypes, end, token.c_str(),
[]( const char * a, const char * b){ return strcmp(a,b) < 0; } );
if(p.first < end && token == opTypes[p.first-opTypes])
operands.push_back(static_cast<Otype>(p.first-opTypes));
else if( token == "|" )
state = sepSeen;
else
throw "Invalid instruction: Expected operand or separator";
break;
}
case sepSeen:
{
unique_ptr<Prefix> prefix;
if( createPrefix( prefix, token, offset ))
prefixes.push_back(move(prefix));
else
{
offset += 8;
opcode.push_back(strtol( token.c_str(), nullptr, 16 ));
state = opcodeSeen;
}
break;
}
case opcodeSeen:
if(isArgOrConst(token))
{
args.insert(MC_Comp(token, offset ));
state = argsSeen;
}
else if(isOpcodeByte(token))
{
opcode.push_back(strtol( token.c_str(), nullptr, 16 ));
offset += 8;
}
else
throw "Invalid instruction: Expected opcode or separator ";
break;
case argsSeen:
if(isArgOrConst(token))
{
args.insert(MC_Comp(token, offset ));
}
else
throw "Invalid instruction";
break;
}
}
if( opcode.size() == 0 )
fprintf( stderr, "%s missing opcode. Ignored\n", Instruction::lexium(nm) );
else
insts.insert(
Instruction(nm,move(operands),move(prefixes),move(opcode), move(args)));
}
catch( ... )
{
fprintf( stderr, "Failed to parse instruction on %s\n", buff );
return false;
}
}
}
}
return true;
}
/**
* Compiles a brainfreeze program. It converts the program's
* textual representation into a program containing only the
* brainfreeze vm instructions.
*
* Additionally, it performs several compile time optimizations in order to
* speed up execution time. These optimizations include pre-calculating
* jump offsets and fusing sequential identical instructions.
*
* \return Returns true if the program was succesfully compiled, otherwise
* it will return false to indicate the presence of errors
*/
bool BFProgram::compile()
{
Instructions temp;
Instruction last = Instruction( OP_EOF, 0 );
std::stack<int> jumps; // record positions for [
//
// Scan the code string. Replace each BF instruction
// with its symbolic equivilant.
//
int icount = 0;
for( std::string::iterator itr = m_codestr.begin();
itr != m_codestr.end();
++itr )
{
if( BF::isInstruction( *itr ) )
{
// Convert the character into an instruction
Instruction instr = BF::convert( *itr );
//
// Is this a jump instruction?
//
if ( instr.opcode() == OP_JMP_FWD )
{
// This is a forward jump. Record its position
jumps.push(icount);
}
else if ( instr.opcode() == OP_JMP_BAC )
{
// This is a backward jump. Pop the corresponding
// forward jump marker off the stack, and update both
// of the instructions with the location of their
// corresponding targets.
int backpos = jumps.top(); jumps.pop();
int dist = icount - backpos;
assert( dist > 0 );
// Update the [ character
temp[backpos].setParam( dist );
// Update the ] (current) character
instr.setParam( dist );
}
//
// Was the last character a repeat of +, -, >, <
//
if ( ( instr.opcode() == last.opcode() ) &&
( instr.opcode() == OP_PTR_INC ||
instr.opcode() == OP_PTR_DEC ||
instr.opcode() == OP_MEM_INC ||
instr.opcode() == OP_MEM_DEC ) )
{
// Get the last character, and update its occurrence
temp[icount-1].setParam( temp[icount-1].param() + 1 );
}
else
{
// Add the instruction to the instruction stream
temp.push_back( instr );
// Remember last instruction
last = instr;
// Increment instruction counter
icount += 1;
}
// Remember the last instruction
last = instr;
}
}
// Verify the jump stack is empty. If not, then there is a
// mismatched jump somewhere!
if (! jumps.empty() )
{
raiseError( "Mismatched jump detected when compiling" );
return false;
}
// Insert end of program instruction
temp.push_back( Instruction( OP_EOF, 0 ) );
// Save it to m_instructions
m_instructions.swap( temp );
m_ip = m_instructions.begin();
return true;
}
void LowerEmAsyncify::transformAsyncFunction(Function &F, Instructions const& AsyncCalls) {
assert(!AsyncCalls.empty());
// Pass 0
// collect all the return instructions from the original function
// will use later
std::vector<ReturnInst*> OrigReturns;
for (inst_iterator I = inst_begin(&F), E = inst_end(&F); I != E; ++I) {
if (ReturnInst *RI = dyn_cast<ReturnInst>(&*I)) {
OrigReturns.push_back(RI);
}
}
// Pass 1
// Scan each async call and make the basic structure:
// All these will be cloned into the callback functions
// - allocate the async context before calling an async function
// - check async right after calling an async function, save context & return if async, continue if not
// - retrieve the async return value and free the async context if the called function turns out to be sync
std::vector<AsyncCallEntry> AsyncCallEntries;
AsyncCallEntries.reserve(AsyncCalls.size());
for (Instructions::const_iterator I = AsyncCalls.begin(), E = AsyncCalls.end(); I != E; ++I) {
// prepare blocks
Instruction *CurAsyncCall = *I;
// The block containing the async call
BasicBlock *CurBlock = CurAsyncCall->getParent();
// The block should run after the async call
BasicBlock *AfterCallBlock = SplitBlock(CurBlock, CurAsyncCall->getNextNode());
// The block where we store the context and return
BasicBlock *SaveAsyncCtxBlock = BasicBlock::Create(TheModule->getContext(), "SaveAsyncCtx", &F, AfterCallBlock);
// return a dummy value at the end, to make the block valid
new UnreachableInst(TheModule->getContext(), SaveAsyncCtxBlock);
// allocate the context before making the call
// we don't know the size yet, will fix it later
// we cannot insert the instruction later because,
// we need to make sure that all the instructions and blocks are fixed before we can generate DT and find context variables
// In CallHandler.h `sp` will be put as the second parameter
// such that we can take a note of the original sp
CallInst *AllocAsyncCtxInst = CallInst::Create(AllocAsyncCtxFunction, Constant::getNullValue(I32), "AsyncCtx", CurAsyncCall);
// Right after the call
// check async and return if so
// TODO: we can define truly async functions and partial async functions
{
// remove old terminator, which came from SplitBlock
CurBlock->getTerminator()->eraseFromParent();
// go to SaveAsyncCtxBlock if the previous call is async
// otherwise just continue to AfterCallBlock
CallInst *CheckAsync = CallInst::Create(CheckAsyncFunction, "IsAsync", CurBlock);
BranchInst::Create(SaveAsyncCtxBlock, AfterCallBlock, CheckAsync, CurBlock);
}
// take a note of this async call
AsyncCallEntry CurAsyncCallEntry;
CurAsyncCallEntry.AsyncCallInst = CurAsyncCall;
CurAsyncCallEntry.AfterCallBlock = AfterCallBlock;
CurAsyncCallEntry.AllocAsyncCtxInst = AllocAsyncCtxInst;
CurAsyncCallEntry.SaveAsyncCtxBlock = SaveAsyncCtxBlock;
// create an empty function for the callback, which will be constructed later
CurAsyncCallEntry.CallbackFunc = Function::Create(CallbackFunctionType, F.getLinkage(), F.getName() + "__async_cb", TheModule);
AsyncCallEntries.push_back(CurAsyncCallEntry);
}
// Pass 2
// analyze the context variables and construct SaveAsyncCtxBlock for each async call
// also calculate the size of the context and allocate the async context accordingly
for (std::vector<AsyncCallEntry>::iterator EI = AsyncCallEntries.begin(), EE = AsyncCallEntries.end(); EI != EE; ++EI) {
AsyncCallEntry & CurEntry = *EI;
// Collect everything to be saved
FindContextVariables(CurEntry);
// Pack the variables as a struct
{
// TODO: sort them from large memeber to small ones, in order to make the struct compact even when aligned
SmallVector<Type*, 8> Types;
Types.push_back(CallbackFunctionType->getPointerTo());
for (Values::iterator VI = CurEntry.ContextVariables.begin(), VE = CurEntry.ContextVariables.end(); VI != VE; ++VI) {
Types.push_back((*VI)->getType());
}
CurEntry.ContextStructType = StructType::get(TheModule->getContext(), Types);
}
// fix the size of allocation
CurEntry.AllocAsyncCtxInst->setOperand(0,
ConstantInt::get(I32, DL->getTypeStoreSize(CurEntry.ContextStructType)));
// construct SaveAsyncCtxBlock
{
// fill in SaveAsyncCtxBlock
// temporarily remove the terminator for convenience
CurEntry.SaveAsyncCtxBlock->getTerminator()->eraseFromParent();
assert(CurEntry.SaveAsyncCtxBlock->empty());
Type *AsyncCtxAddrTy = CurEntry.ContextStructType->getPointerTo();
BitCastInst *AsyncCtxAddr = new BitCastInst(CurEntry.AllocAsyncCtxInst, AsyncCtxAddrTy, "AsyncCtxAddr", CurEntry.SaveAsyncCtxBlock);
SmallVector<Value*, 2> Indices;
// store the callback
{
Indices.push_back(ConstantInt::get(I32, 0));
Indices.push_back(ConstantInt::get(I32, 0));
GetElementPtrInst *AsyncVarAddr = GetElementPtrInst::Create(AsyncCtxAddrTy, AsyncCtxAddr, Indices, "", CurEntry.SaveAsyncCtxBlock);
new StoreInst(CurEntry.CallbackFunc, AsyncVarAddr, CurEntry.SaveAsyncCtxBlock);
}
// store the context variables
for (size_t i = 0; i < CurEntry.ContextVariables.size(); ++i) {
Indices.clear();
Indices.push_back(ConstantInt::get(I32, 0));
Indices.push_back(ConstantInt::get(I32, i + 1)); // the 0th element is the callback function
GetElementPtrInst *AsyncVarAddr = GetElementPtrInst::Create(AsyncCtxAddrTy, AsyncCtxAddr, Indices, "", CurEntry.SaveAsyncCtxBlock);
new StoreInst(CurEntry.ContextVariables[i], AsyncVarAddr, CurEntry.SaveAsyncCtxBlock);
}
// to exit the block, we want to return without unwinding the stack frame
CallInst::Create(DoNotUnwindFunction, "", CurEntry.SaveAsyncCtxBlock);
ReturnInst::Create(TheModule->getContext(),
(F.getReturnType()->isVoidTy() ? 0 : Constant::getNullValue(F.getReturnType())),
CurEntry.SaveAsyncCtxBlock);
}
}
// Pass 3
// now all the SaveAsyncCtxBlock's have been constructed
// we can clone F and construct callback functions
// we could not construct the callbacks in Pass 2 because we need _all_ those SaveAsyncCtxBlock's appear in _each_ callback
for (std::vector<AsyncCallEntry>::iterator EI = AsyncCallEntries.begin(), EE = AsyncCallEntries.end(); EI != EE; ++EI) {
AsyncCallEntry & CurEntry = *EI;
Function *CurCallbackFunc = CurEntry.CallbackFunc;
ValueToValueMapTy VMap;
// Add the entry block
// load variables from the context
// also update VMap for CloneFunction
BasicBlock *EntryBlock = BasicBlock::Create(TheModule->getContext(), "AsyncCallbackEntry", CurCallbackFunc);
std::vector<LoadInst *> LoadedAsyncVars;
{
Type *AsyncCtxAddrTy = CurEntry.ContextStructType->getPointerTo();
BitCastInst *AsyncCtxAddr = new BitCastInst(CurCallbackFunc->arg_begin(), AsyncCtxAddrTy, "AsyncCtx", EntryBlock);
SmallVector<Value*, 2> Indices;
for (size_t i = 0; i < CurEntry.ContextVariables.size(); ++i) {
Indices.clear();
Indices.push_back(ConstantInt::get(I32, 0));
Indices.push_back(ConstantInt::get(I32, i + 1)); // the 0th element of AsyncCtx is the callback function
GetElementPtrInst *AsyncVarAddr = GetElementPtrInst::Create(AsyncCtxAddrTy, AsyncCtxAddr, Indices, "", EntryBlock);
LoadedAsyncVars.push_back(new LoadInst(AsyncVarAddr, "", EntryBlock));
// we want the argument to be replaced by the loaded value
if (isa<Argument>(CurEntry.ContextVariables[i]))
VMap[CurEntry.ContextVariables[i]] = LoadedAsyncVars.back();
}
}
// we don't need any argument, just leave dummy entries there to cheat CloneFunctionInto
for (Function::const_arg_iterator AI = F.arg_begin(), AE = F.arg_end(); AI != AE; ++AI) {
if (VMap.count(AI) == 0)
VMap[AI] = Constant::getNullValue(AI->getType());
}
// Clone the function
{
SmallVector<ReturnInst*, 8> Returns;
CloneFunctionInto(CurCallbackFunc, &F, VMap, false, Returns);
// return type of the callback functions is always void
// need to fix the return type
if (!F.getReturnType()->isVoidTy()) {
// for those return instructions that are from the original function
// it means we are 'truly' leaving this function
// need to store the return value right before ruturn
for (size_t i = 0; i < OrigReturns.size(); ++i) {
ReturnInst *RI = cast<ReturnInst>(VMap[OrigReturns[i]]);
// Need to store the return value into the global area
CallInst *RawRetValAddr = CallInst::Create(GetAsyncReturnValueAddrFunction, "", RI);
BitCastInst *RetValAddr = new BitCastInst(RawRetValAddr, F.getReturnType()->getPointerTo(), "AsyncRetValAddr", RI);
new StoreInst(RI->getOperand(0), RetValAddr, RI);
}
// we want to unwind the stack back to where it was before the original function as called
// but we don't actually need to do this here
// at this point it must be true that no callback is pended
// so the scheduler will correct the stack pointer and pop the frame
// here we just fix the return type
for (size_t i = 0; i < Returns.size(); ++i) {
ReplaceInstWithInst(Returns[i], ReturnInst::Create(TheModule->getContext()));
}
}
}
// the callback function does not have any return value
// so clear all the attributes for return
{
AttributeSet Attrs = CurCallbackFunc->getAttributes();
CurCallbackFunc->setAttributes(
Attrs.removeAttributes(TheModule->getContext(), AttributeSet::ReturnIndex, Attrs.getRetAttributes())
);
}
// in the callback function, we never allocate a new async frame
// instead we reuse the existing one
for (std::vector<AsyncCallEntry>::iterator EI = AsyncCallEntries.begin(), EE = AsyncCallEntries.end(); EI != EE; ++EI) {
Instruction *I = cast<Instruction>(VMap[EI->AllocAsyncCtxInst]);
ReplaceInstWithInst(I, CallInst::Create(ReallocAsyncCtxFunction, I->getOperand(0), "ReallocAsyncCtx"));
}
// mapped entry point & async call
BasicBlock *ResumeBlock = cast<BasicBlock>(VMap[CurEntry.AfterCallBlock]);
Instruction *MappedAsyncCall = cast<Instruction>(VMap[CurEntry.AsyncCallInst]);
// To save space, for each async call in the callback function, we just ignore the sync case, and leave it to the scheduler
// TODO need an option for this
{
for (std::vector<AsyncCallEntry>::iterator EI = AsyncCallEntries.begin(), EE = AsyncCallEntries.end(); EI != EE; ++EI) {
AsyncCallEntry & CurEntry = *EI;
Instruction *MappedAsyncCallInst = cast<Instruction>(VMap[CurEntry.AsyncCallInst]);
BasicBlock *MappedAsyncCallBlock = MappedAsyncCallInst->getParent();
BasicBlock *MappedAfterCallBlock = cast<BasicBlock>(VMap[CurEntry.AfterCallBlock]);
// for the sync case of the call, go to NewBlock (instead of MappedAfterCallBlock)
BasicBlock *NewBlock = BasicBlock::Create(TheModule->getContext(), "", CurCallbackFunc, MappedAfterCallBlock);
MappedAsyncCallBlock->getTerminator()->setSuccessor(1, NewBlock);
// store the return value
if (!MappedAsyncCallInst->use_empty()) {
CallInst *RawRetValAddr = CallInst::Create(GetAsyncReturnValueAddrFunction, "", NewBlock);
BitCastInst *RetValAddr = new BitCastInst(RawRetValAddr, MappedAsyncCallInst->getType()->getPointerTo(), "AsyncRetValAddr", NewBlock);
new StoreInst(MappedAsyncCallInst, RetValAddr, NewBlock);
}
// tell the scheduler that we want to keep the current async stack frame
CallInst::Create(DoNotUnwindAsyncFunction, "", NewBlock);
// finally we go to the SaveAsyncCtxBlock, to register the callbac, save the local variables and leave
BasicBlock *MappedSaveAsyncCtxBlock = cast<BasicBlock>(VMap[CurEntry.SaveAsyncCtxBlock]);
BranchInst::Create(MappedSaveAsyncCtxBlock, NewBlock);
}
}
std::vector<AllocaInst*> ToPromote;
// applying loaded variables in the entry block
{
BasicBlockSet ReachableBlocks = FindReachableBlocksFrom(ResumeBlock);
for (size_t i = 0; i < CurEntry.ContextVariables.size(); ++i) {
Value *OrigVar = CurEntry.ContextVariables[i];
if (isa<Argument>(OrigVar)) continue; // already processed
Value *CurVar = VMap[OrigVar];
assert(CurVar != MappedAsyncCall);
if (Instruction *Inst = dyn_cast<Instruction>(CurVar)) {
if (ReachableBlocks.count(Inst->getParent())) {
// Inst could be either defined or loaded from the async context
// Do the dirty works in memory
// TODO: might need to check the safety first
// TODO: can we create phi directly?
AllocaInst *Addr = DemoteRegToStack(*Inst, false);
new StoreInst(LoadedAsyncVars[i], Addr, EntryBlock);
ToPromote.push_back(Addr);
} else {
// The parent block is not reachable, which means there is no confliction
// it's safe to replace Inst with the loaded value
assert(Inst != LoadedAsyncVars[i]); // this should only happen when OrigVar is an Argument
Inst->replaceAllUsesWith(LoadedAsyncVars[i]);
}
}
}
}
// resolve the return value of the previous async function
// it could be the value just loaded from the global area
// or directly returned by the function (in its sync case)
if (!CurEntry.AsyncCallInst->use_empty()) {
// load the async return value
CallInst *RawRetValAddr = CallInst::Create(GetAsyncReturnValueAddrFunction, "", EntryBlock);
BitCastInst *RetValAddr = new BitCastInst(RawRetValAddr, MappedAsyncCall->getType()->getPointerTo(), "AsyncRetValAddr", EntryBlock);
LoadInst *RetVal = new LoadInst(RetValAddr, "AsyncRetVal", EntryBlock);
AllocaInst *Addr = DemoteRegToStack(*MappedAsyncCall, false);
new StoreInst(RetVal, Addr, EntryBlock);
ToPromote.push_back(Addr);
}
// TODO remove unreachable blocks before creating phi
// We go right to ResumeBlock from the EntryBlock
BranchInst::Create(ResumeBlock, EntryBlock);
/*
* Creating phi's
* Normal stack frames and async stack frames are interleaving with each other.
* In a callback function, if we call an async function, we might need to realloc the async ctx.
* at this point we don't want anything stored after the ctx,
* such that we can free and extend the ctx by simply update STACKTOP.
* Therefore we don't want any alloca's in callback functions.
*
*/
if (!ToPromote.empty()) {
DominatorTreeWrapperPass DTW;
DTW.runOnFunction(*CurCallbackFunc);
PromoteMemToReg(ToPromote, DTW.getDomTree());
}
removeUnreachableBlocks(*CurCallbackFunc);
}
// Pass 4
// Here are modifications to the original function, which we won't want to be cloned into the callback functions
for (std::vector<AsyncCallEntry>::iterator EI = AsyncCallEntries.begin(), EE = AsyncCallEntries.end(); EI != EE; ++EI) {
AsyncCallEntry & CurEntry = *EI;
// remove the frame if no async functinon has been called
CallInst::Create(FreeAsyncCtxFunction, CurEntry.AllocAsyncCtxInst, "", CurEntry.AfterCallBlock->getFirstNonPHI());
}
}
void linkInstructionBranches(Instructions &instructions) {
/* Go through all instructions and link them according to the flow graph.
*
* In specifics, link each instruction's follower, the instruction that
* naturally follows if no branches are taken. Also fill in the branches
* array, which contains all branches an instruction can take. This
* directly creates an address type for each instruction: does it start
* a subroutine, is it a jump destination, is it a tail of a jump or none
* of these?
*/
for (Instructions::iterator i = instructions.begin(); i != instructions.end(); ++i) {
// If this is an instruction that has a natural follower, link it
if ((i->opcode != kOpcodeJMP) && (i->opcode != kOpcodeRETN)) {
Instructions::iterator follower = i + 1;
i->follower = (follower != instructions.end()) ? &*follower : 0;
if (follower != instructions.end())
follower->predecessors.push_back(&*i);
}
// Link destinations of unconditional branches
if ((i->opcode == kOpcodeJMP) || (i->opcode == kOpcodeJSR) || (i->opcode == kOpcodeSTORESTATE)) {
assert(((i->opcode == kOpcodeSTORESTATE) && (i->argCount == 3)) || (i->argCount == 1));
Instruction *branch = findInstruction(instructions, i->address + i->args[0]);
if (!branch)
throw Common::Exception("Can't find destination of unconditional branch");
i->branches.push_back(branch);
if (i->opcode == kOpcodeJSR)
setAddressType(branch, kAddressTypeSubRoutine);
else if (i->opcode == kOpcodeSTORESTATE)
setAddressType(branch, kAddressTypeStoreState);
else {
setAddressType(branch, kAddressTypeJumpLabel);
branch->predecessors.push_back(&*i);
}
setAddressType(const_cast<Instruction *>(i->follower), kAddressTypeTail);
}
// Link destinations of conditional branches
if ((i->opcode == kOpcodeJZ) || (i->opcode == kOpcodeJNZ)) {
assert(i->argCount == 1);
if (!i->follower)
throw Common::Exception("Conditional branch has no false destination");
Instruction *branch = findInstruction(instructions, i->address + i->args[0]);
if (!branch)
throw Common::Exception("Can't find destination of conditional branch");
setAddressType(branch, kAddressTypeJumpLabel);
setAddressType(const_cast<Instruction *>(i->follower), kAddressTypeTail);
i->branches.push_back(branch); // True branch
i->branches.push_back(i->follower); // False branch
branch->predecessors.push_back(&*i);
}
}
}
