Skip to content

Commit

Permalink
Optimize exception frames allocation
Browse files Browse the repository at this point in the history 
* Make llvm-gcroot a ModulePass
* Use llvm-gcroot to calculate the minimum number of exception frames necessary.
  • Loading branch information
@yuyichao
yuyichao committed Aug 2, 2016
1 parent 82dafeb commit 15082a9
Show file tree
Hide file tree
Showing 3 changed files with 222 additions and 52 deletions.
35 changes: 8 additions & 27 deletions src/codegen.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -381,6 +381,7 @@ static Function *gcroot_func;
static Function *gckill_func;
static Function *jlcall_frame_func;
static Function *gcroot_flush_func;
static Function *except_enter_func;

static std::vector<Type *> two_pvalue_llvmt;
static std::vector<Type *> three_pvalue_llvmt;
Expand Down Expand Up @@ -505,7 +506,6 @@ typedef struct {
std::vector<bool> ssavalue_assigned;
std::map<int, jl_arrayvar_t> *arrayvars;
std::map<int, BasicBlock*> *labels;
std::map<int, Value*> *handlers;
jl_module_t *module;
jl_lambda_info_t *linfo;
const char *name;
Expand Down Expand Up @@ -3259,18 +3259,7 @@ static jl_cgval_t emit_expr(jl_value_t *expr, jl_codectx_t *ctx)
else if (head == enter_sym) {
assert(jl_is_long(args[0]));
int labl = jl_unbox_long(args[0]);
Value *jbuf = builder.CreateGEP((*ctx->handlers)[labl],
ConstantInt::get(T_size,0));
builder.CreateCall(prepare_call(jlenter_func), jbuf);
#ifndef _OS_WINDOWS_
#ifdef LLVM37
CallInst *sj = builder.CreateCall(prepare_call(setjmp_func), { jbuf, ConstantInt::get(T_int32,0) });
#else
CallInst *sj = builder.CreateCall2(prepare_call(setjmp_func), jbuf, ConstantInt::get(T_int32,0));
#endif
#else
CallInst *sj = builder.CreateCall(prepare_call(setjmp_func), jbuf);
#endif
CallInst *sj = builder.CreateCall(prepare_call(except_enter_func));
// We need to mark this on the call site as well. See issue #6757
sj->setCanReturnTwice();
Value *isz = builder.CreateICmpEQ(sj, ConstantInt::get(T_int32,0));
Expand Down Expand Up @@ -3892,11 +3881,9 @@ static std::unique_ptr<Module> emit_function(jl_lambda_info_t *lam, jl_llvm_func
//jl_printf(JL_STDOUT, "\n");
std::map<int, jl_arrayvar_t> arrayvars;
std::map<int, BasicBlock*> labels;
std::map<int, Value*> handlers;
jl_codectx_t ctx = {};
ctx.arrayvars = &arrayvars;
ctx.labels = &labels;
ctx.handlers = &handlers;
ctx.module = lam->def ? lam->def->module : ptls->current_module;
ctx.linfo = lam;
ctx.name = jl_symbol_name(lam->def ? lam->def->name : anonymous_sym);
Expand Down Expand Up @@ -4329,18 +4316,6 @@ static std::unique_ptr<Module> emit_function(jl_lambda_info_t *lam, jl_llvm_func
allocate_gc_frame(b0, &ctx);

// step 8. allocate space for exception handler contexts
for(i=0; i < stmtslen; i++) {
jl_value_t *stmt = jl_array_ptr_ref(stmts,i);
if (jl_is_expr(stmt) && ((jl_expr_t*)stmt)->head == enter_sym) {
int labl = jl_unbox_long(jl_exprarg(stmt,0));
AllocaInst *handlr =
builder.CreateAlloca(T_int8,
ConstantInt::get(T_int32,
sizeof(jl_handler_t)));
handlr->setAlignment(16);
handlers[labl] = handlr;
}
}

// step 9. allocate local variables slots
// must be in the first basic block for the llvm mem2reg pass to work
Expand Down Expand Up @@ -5533,6 +5508,12 @@ static void init_julia_llvm_env(Module *m)
"julia.gcroot_flush", m);
add_named_global(gcroot_flush_func, (void*)NULL, /*dllimport*/false);

except_enter_func = Function::Create(FunctionType::get(T_int32, false),
Function::ExternalLinkage,
"julia.except_enter", m);
except_enter_func->addFnAttr(Attribute::ReturnsTwice);
add_named_global(except_enter_func, (void*)NULL, /*dllimport*/false);

// set up optimization passes
#ifdef LLVM37
// No DataLayout pass needed anymore.
Expand Down
226 changes: 214 additions & 12 deletions src/llvm-gcroot.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -130,6 +130,8 @@ class JuliaGCAllocator {
gckill_func(M.getFunction("julia.gc_root_kill")),
jlcall_frame_func(M.getFunction("julia.jlcall_frame_decl")),
gcroot_flush_func(M.getFunction("julia.gcroot_flush")),
except_enter_func(M.getFunction("julia.except_enter")),
jlleave_func(M.getFunction("jl_pop_handler")),
tbaa_gcframe(tbaa)
{
/* Algorithm sketch:
Expand Down Expand Up @@ -158,6 +160,8 @@ class JuliaGCAllocator {
Function *const gckill_func;
Function *const jlcall_frame_func;
Function *const gcroot_flush_func;
Function *const except_enter_func;
Function *const jlleave_func;
MDNode *const tbaa_gcframe;

Instruction *get_pgcstack(Instruction *ptlsStates);
Expand All @@ -171,10 +175,156 @@ class JuliaGCAllocator {
std::map<BasicBlock*, std::map<frame_register, liveness::id> > &bb_uses,
std::map<BasicBlock*, SmallBitVector> &regs_used);
void rearrangeRoots();
void lowerHandlers();
public:
void allocate_frame();
};

struct HandlerData {
std::vector<std::pair<CallInst*,uint64_t>> leaves;
std::set<CallInst*> nested;
bool processed{false};
};

void JuliaGCAllocator::lowerHandlers()
{
if (!except_enter_func)
return;
auto jlenter_func = M.getNamedValue("jl_enter_handler");
auto setjmp_func = M.getNamedValue(jl_setjmp_name);
// Collect all exception enters
std::map<CallInst*,HandlerData> handlers;
for (auto &BB: F) {
for (auto &I: BB) {
auto call = dyn_cast<CallInst>(&I);
if (call && call->getCalledValue() == except_enter_func) {
handlers[call] = HandlerData();
}
}
}

if (handlers.empty())
return;

typedef std::map<CallInst*,HandlerData>::iterator hdlr_iter_t;
// For each exception enter, compute the life time of the enter, find
// the corresponding leaves and collect a list of nested exception frames.
// This assumes the exception frames has simple structure. E.g.
// there's no recursion and different frames do no share the same leave.
std::function<void(hdlr_iter_t)> process_handler = [&] (hdlr_iter_t hdlr) {
auto enter = hdlr->first;
auto &data = hdlr->second;
if (data.processed)
return;
data.processed = true;
std::vector<BasicBlock::iterator> frontier = {
++BasicBlock::iterator(enter)
};
// Start from the enter, the enter is live until it hits a leave
// or the control flow terminates.
std::set<Instruction*> visited;
auto add_inst = [&] (BasicBlock::iterator it) {
auto I = &*it;
if (visited.find(I) == visited.end()) {
visited.insert(I);
frontier.push_back(it);
}
};
while (!frontier.empty()) {
BasicBlock::iterator IT = frontier.back();
Instruction *I = &*IT;
++IT;
frontier.pop_back();
if (I->isTerminator()) {
BasicBlock *bb = I->getParent();
for (auto S = succ_begin(bb), E = succ_end(bb); S != E; S++)
add_inst(S->getFirstInsertionPt());
continue;
}
auto call = dyn_cast<CallInst>(I);
if (!call) {
add_inst(IT);
continue;
}
auto callee = call->getCalledValue();
if (callee == except_enter_func) {
// Nested frame, make sure the frame is processed and skip to
// its leaves.
auto other_it = handlers.find(call);
assert(other_it != handlers.end());
process_handler(other_it);
auto &other_data = other_it->second;
data.nested.insert(other_data.nested.begin(),
other_data.nested.end());
data.nested.insert(call);
for (auto leave: other_it->second.leaves) {
if (leave.second > 0) {
data.leaves.push_back(std::make_pair(leave.first,
leave.second - 1));
}
else {
add_inst(++BasicBlock::iterator(leave.first));
}
}
}
else if (callee == jlleave_func) {
// Leave, record how many pops are left on this leave.
assert(call->getNumArgOperands() == 1);
auto arg = cast<ConstantInt>(call->getArgOperand(0));
uint64_t pop_count = arg->getLimitedValue();
assert(pop_count >= 1);
data.leaves.push_back(std::make_pair(call, pop_count - 1));
}
else {
// Otherwise, go to the next instruction.
add_inst(IT);
}
}
};

for (auto IT = handlers.begin(), E = handlers.end(); IT != E; ++IT)
process_handler(IT);

std::set<CallInst*> processing;
Value *handler_sz = ConstantInt::get(T_int32, sizeof(jl_handler_t));
// Now allocate the stack slots.
while (!handlers.empty()) {
processing.clear();
auto buff = new AllocaInst(T_int8, handler_sz, "", ptlsStates);
buff->setAlignment(16);
for (auto &hdlr: handlers) {
auto enter = hdlr.first;
auto &nested = hdlr.second.nested;
if (nested.empty()) {
processing.insert(enter);
}
}
for (auto processed: processing) {
handlers.erase(processed);
CallInst::Create(jlenter_func, {buff}, "", processed);
#ifndef _OS_WINDOWS_
auto sj = CallInst::Create(setjmp_func,
{ buff, ConstantInt::get(T_int32,0) },
"", processed);
#else
auto sj = CallInst::Create(setjmp_func, { buff }, "", processed);
#endif
// We need to mark this on the call site as well. See issue #6757
sj->setCanReturnTwice();
for (auto &hdlr: handlers) {
auto &nested = hdlr.second.nested;
auto it = nested.find(processed);
if (it != nested.end()) {
nested.erase(it);
}
}
processed->replaceAllUsesWith(sj);
processed->eraseFromParent();
}
assert(!processing.empty());
}
}

Instruction *JuliaGCAllocator::get_pgcstack(Instruction *ptlsStates)
{
Constant *offset = ConstantInt::getSigned(T_int32, offsetof(jl_tls_states_t, pgcstack) / sizeof(void*));
Expand Down Expand Up @@ -528,6 +678,7 @@ void JuliaGCAllocator::rearrangeRoots()

void JuliaGCAllocator::allocate_frame()
{
lowerHandlers();
Instruction *last_gcframe_inst = gcframe;
collapseRedundantRoots();
rearrangeRoots();
Expand Down Expand Up @@ -916,26 +1067,82 @@ void JuliaGCAllocator::allocate_frame()
#endif
}

struct LowerGCFrame: public FunctionPass {
struct LowerGCFrame: public ModulePass {
static char ID;
LowerGCFrame(MDNode *_tbaa_gcframe=nullptr)
: FunctionPass(ID),
: ModulePass(ID),
tbaa_gcframe(_tbaa_gcframe)
{}

private:
MDNode *tbaa_gcframe; // One `LLVMContext` only
bool runOnFunction(Function &F) override;
void runOnFunction(Module *M, Function &F, Function *ptls_getter,
Type *T_pjlvalue);
bool runOnModule(Module &M) override;
};

bool LowerGCFrame::runOnFunction(Function &F)
static void eraseFunction(Module &M, const char *name)
{
if (Function *f = M.getFunction(name)) {
f->eraseFromParent();
}
}

static void add_function(Module &M, const char *name, FunctionType *ft)
{
Function::Create(ft, Function::ExternalLinkage, name, &M);
}

static void ensure_enter_function(Module &M)
{
Module *M = F.getParent();
auto T_int8 = Type::getInt8Ty(M.getContext());
auto T_pint8 = PointerType::get(T_int8, 0);
auto T_void = Type::getVoidTy(M.getContext());
auto T_int32 = Type::getInt32Ty(M.getContext());
if (!M.getNamedValue("jl_enter_handler")) {
std::vector<Type*> ehargs(0);
ehargs.push_back(T_pint8);
add_function(M, "jl_enter_handler",
FunctionType::get(T_void, ehargs, false));
}
if (!M.getNamedValue(jl_setjmp_name)) {
std::vector<Type*> args2(0);
args2.push_back(T_pint8);
#ifndef _OS_WINDOWS_
args2.push_back(T_int32);
#endif
add_function(M, jl_setjmp_name,
FunctionType::get(T_int32, args2, false));
}
}

Function *ptls_getter = M->getFunction("jl_get_ptls_states");
bool LowerGCFrame::runOnModule(Module &M)
{
Function *ptls_getter = M.getFunction("jl_get_ptls_states");
if (!ptls_getter)
return true;
ensure_enter_function(M);
FunctionType *functype = ptls_getter->getFunctionType();
auto T_ppjlvalue =
cast<PointerType>(functype->getReturnType())->getElementType();
auto T_pjlvalue = cast<PointerType>(T_ppjlvalue)->getElementType();
for (auto F = M.begin(), E = M.end(); F != E; ++F) {
if (F->isDeclaration())
continue;
runOnFunction(&M, *F, ptls_getter, T_pjlvalue);
}

// Cleanup for GC frame lowering.
eraseFunction(M, "julia.gc_root_decl");
eraseFunction(M, "julia.gc_root_kill");
eraseFunction(M, "julia.jlcall_frame_decl");
eraseFunction(M, "julia.gcroot_flush");
return true;
}

void LowerGCFrame::runOnFunction(Module *M, Function &F, Function *ptls_getter,
Type *T_pjlvalue)
{
CallInst *ptlsStates = NULL;
for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end();
I != E; ++I) {
Expand All @@ -947,15 +1154,10 @@ bool LowerGCFrame::runOnFunction(Function &F)
}
}
if (!ptlsStates)
return true;
return;

FunctionType *functype = ptls_getter->getFunctionType();
auto T_ppjlvalue =
cast<PointerType>(functype->getReturnType())->getElementType();
auto T_pjlvalue = cast<PointerType>(T_ppjlvalue)->getElementType();
JuliaGCAllocator allocator(ptlsStates, T_pjlvalue, tbaa_gcframe);
allocator.allocate_frame();
return true;
}

char LowerGCFrame::ID = 0;
Expand Down
13 changes: 0 additions & 13 deletions src/llvm-ptls.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -140,21 +140,8 @@ void LowerPTLS::runOnFunction(LLVMContext &ctx, Module &M, Function *F,
#endif
}

static void eraseFunction(Module &M, const char *name)
{
if (Function *f = M.getFunction(name)) {
f->eraseFromParent();
}
}

bool LowerPTLS::runOnModule(Module &M)
{
// Cleanup for GC frame lowering.
eraseFunction(M, "julia.gc_root_decl");
eraseFunction(M, "julia.gc_root_kill");
eraseFunction(M, "julia.jlcall_frame_decl");
eraseFunction(M, "julia.gcroot_flush");

Function *ptls_getter = M.getFunction("jl_get_ptls_states");
if (!ptls_getter)
return true;
Expand Down

0 comments on commit 15082a9

Please sign in to comment.