coalescing of event handlers (ZAM optimization)

This commit is contained in:
Vern Paxson 2023-11-16 12:58:28 -08:00
parent a488dcffb6
commit 3d21d80dac
6 changed files with 267 additions and 46 deletions

View file

@ -172,11 +172,20 @@ void EventGroup::UpdateFuncBodies() {
static auto is_group_disabled = [](const auto& g) { return g->IsDisabled(); };
for ( auto& func : funcs ) {
for ( auto& b : func->bodies )
bool func_changed = false;
for ( auto& b : func->bodies ) {
auto prev = b.disabled;
b.disabled = std::any_of(b.groups.cbegin(), b.groups.cend(), is_group_disabled);
if ( prev != b.disabled )
func_changed = true;
}
if ( ! func_changed )
continue;
static auto is_body_enabled = [](const auto& b) { return ! b.disabled; };
func->has_enabled_bodies = std::any_of(func->bodies.cbegin(), func->bodies.cend(), is_body_enabled);
func->all_bodies_enabled = std::all_of(func->bodies.cbegin(), func->bodies.cend(), is_body_enabled);
}
}

View file

@ -84,6 +84,13 @@ public:
*/
bool HasEnabledBodies() const { return ! bodies.empty() && has_enabled_bodies; };
/**
* Is every body enabled?
*
* @return true if all bodies are enabled. (If no bodies, then true.)
*/
bool HasAllBodiesEnabled() const { return all_bodies_enabled; };
/**
* Calls a Zeek function.
* @param args the list of arguments to the function call.
@ -152,6 +159,7 @@ private:
// expose accessors in the zeek:: public interface.
friend class EventGroup;
bool has_enabled_bodies = true;
bool all_bodies_enabled = true;
};
namespace detail {

View file

@ -3,6 +3,8 @@
#include "zeek/script_opt/Inline.h"
#include "zeek/Desc.h"
#include "zeek/EventRegistry.h"
#include "zeek/module_util.h"
#include "zeek/script_opt/ProfileFunc.h"
#include "zeek/script_opt/ScriptOpt.h"
#include "zeek/script_opt/StmtOptInfo.h"
@ -99,13 +101,13 @@ void Inliner::Analyze() {
}
for ( auto& f : funcs ) {
if ( f.ShouldSkip() )
continue;
const auto& func_ptr = f.FuncPtr();
const auto& func = func_ptr.get();
const auto& body = f.Body();
if ( ! should_analyze(func_ptr, body) )
continue;
// Candidates are non-event, non-hook, non-recursive,
// non-compiled functions ...
if ( func->Flavor() != FUNC_FLAVOR_FUNCTION )
@ -120,33 +122,171 @@ void Inliner::Analyze() {
inline_ables[func] = f.Profile();
}
CoalesceEventHandlers();
for ( auto& f : funcs )
if ( should_analyze(f.FuncPtr(), f.Body()) )
if ( f.ShouldAnalyze() )
InlineFunction(&f);
}
void Inliner::CoalesceEventHandlers() {
std::unordered_map<ScriptFunc*, size_t> event_handlers;
BodyInfo body_to_info;
for ( size_t i = 0U; i < funcs.size(); ++i ) {
auto& f = funcs[i];
if ( ! f.ShouldAnalyze() )
continue;
auto& func_ptr = f.FuncPtr();
const auto& func = func_ptr.get();
const auto& func_type = func->GetType();
if ( func_type->AsFuncType()->Flavor() != FUNC_FLAVOR_EVENT )
continue;
// Special-case: zeek_init both has tons of event handlers (even
// with -b), such that it inevitably blows out the inlining budget,
// *and* only runs once, such that even if we could inline it, if
// it takes more time to compile it than to just run it via the
// interpreter, it's a lose.
static std::string zeek_init_name = "zeek_init";
if ( func->Name() == zeek_init_name )
continue;
const auto& body = f.Body();
if ( func->GetKind() == Func::SCRIPT_FUNC && func->GetBodies().size() > 1 ) {
if ( event_handlers.count(func) == 0 )
event_handlers[func] = 1;
else
++event_handlers[func];
ASSERT(body_to_info.count(body.get()) == 0);
body_to_info[body.get()] = i;
}
}
for ( auto& e : event_handlers ) {
auto func = e.first;
auto& bodies = func->GetBodies();
if ( bodies.size() != e.second )
// It's potentially unsound to inline some-but-not-all event
// handlers, because doing so may violate &priority's. We
// could do the work of identifying such instances and only
// skipping those, but given that ZAM is feature-complete
// the mismatch here should only arise when using restrictions
// like --optimize-file, which likely aren't the common case.
continue;
CoalesceEventHandlers({NewRef{}, func}, bodies, body_to_info);
}
}
void Inliner::CoalesceEventHandlers(ScriptFuncPtr func, const std::vector<Func::Body>& bodies,
const BodyInfo& body_to_info) {
auto merged_body = make_intrusive<StmtList>();
auto oi = merged_body->GetOptInfo();
auto& params = func->GetType()->Params();
auto nparams = params->NumFields();
size_t init_frame_size = static_cast<size_t>(nparams);
PreInline(oi, init_frame_size);
// We pattern the new (alternate) body off of the first body.
auto& b0 = func->GetBodies()[0].stmts;
auto b0_info = body_to_info.find(b0.get());
ASSERT(b0_info != body_to_info.end());
auto& info0 = funcs[b0_info->second];
auto& scope0 = info0.Scope();
auto& vars = scope0->OrderedVars();
// We need to create a new Scope. Otherwise, when inlining the first
// body the analysis of identifiers gets confused regarding whether
// a given identifier represents the outer instance or the inner.
auto empty_attrs = std::make_unique<std::vector<AttrPtr>>();
push_scope(scope0->GetID(), std::move(empty_attrs));
std::vector<IDPtr> param_ids;
for ( auto i = 0; i < nparams; ++i ) {
auto& vi = vars[i];
// We use a special scope name so that when debugging issues we can
// see that a given variable came from coalescing event handlers.
auto p = install_ID(vi->Name(), "<event>", false, false);
p->SetType(vi->GetType());
param_ids.push_back(std::move(p));
}
auto new_scope = pop_scope();
auto orig_scope = func->GetScope();
func->SetScope(new_scope);
// Build up the calling arguments.
auto args = make_intrusive<ListExpr>();
for ( auto& p : param_ids )
args->Append(make_intrusive<NameExpr>(p));
for ( auto& b : bodies ) {
auto bp = b.stmts;
auto bi_find = body_to_info.find(bp.get());
ASSERT(bi_find != body_to_info.end());
auto& bi = funcs[bi_find->second];
auto ie = DoInline(func, bp, args, bi.Scope(), bi.Profile());
if ( ! ie ) {
// Failure presumably occurred due to hitting the maximum
// AST complexity for inlining. We can give up by simply
// returning, as at this point we haven't made any actual
// changes other than the function's scope.
func->SetScope(orig_scope);
return;
}
merged_body->Stmts().push_back(make_intrusive<ExprStmt>(ie));
}
auto inlined_func = make_intrusive<CoalescedScriptFunc>(merged_body, new_scope, func);
// Replace the function for that EventHandler with the delegating one.
auto* eh = event_registry->Lookup(func->Name());
ASSERT(eh);
eh->SetFunc(inlined_func);
// Likewise, replace the value of the identifier.
auto fid = lookup_ID(func->Name(), GLOBAL_MODULE_NAME, false, false, false);
ASSERT(fid);
fid->SetVal(make_intrusive<FuncVal>(inlined_func));
PostInline(oi, inlined_func);
funcs.emplace_back(inlined_func, new_scope, merged_body, 0);
auto pf = std::make_shared<ProfileFunc>(inlined_func.get(), merged_body, true);
funcs.back().SetProfile(std::move(pf));
}
void Inliner::InlineFunction(FuncInfo* f) {
max_inlined_frame_size = 0;
// It's important that we take the current frame size from the
// *scope* and not f->Func(). The latter tracks the maximum required
// across all bodies, but we want to track the size for this
// particular body.
curr_frame_size = f->Scope()->Length();
auto oi = f->Body()->GetOptInfo();
PreInline(oi, f->Scope()->Length());
f->Body()->Inline(this);
PostInline(oi, f->FuncPtr());
}
void Inliner::PreInline(StmtOptInfo* oi, size_t frame_size) {
max_inlined_frame_size = 0;
curr_frame_size = frame_size;
num_stmts = oi->num_stmts;
num_exprs = oi->num_exprs;
}
f->Body()->Inline(this);
void Inliner::PostInline(StmtOptInfo* oi, ScriptFuncPtr f) {
oi->num_stmts = num_stmts;
oi->num_exprs = num_exprs;
int new_frame_size = curr_frame_size + max_inlined_frame_size;
if ( new_frame_size > f->Func()->FrameSize() )
f->Func()->SetFrameSize(new_frame_size);
if ( new_frame_size > f->FrameSize() )
f->SetFrameSize(new_frame_size);
}
ExprPtr Inliner::CheckForInlining(CallExprPtr c) {
@ -166,21 +306,21 @@ ExprPtr Inliner::CheckForInlining(CallExprPtr c) {
if ( ! func_v )
return c;
auto function = func_v->AsFunc();
auto function = func_v->AsFuncVal()->AsFuncPtr();
if ( function->GetKind() != Func::SCRIPT_FUNC )
return c;
auto func_vf = static_cast<ScriptFunc*>(function);
auto func_vf = cast_intrusive<ScriptFunc>(function);
auto ia = inline_ables.find(func_vf);
auto ia = inline_ables.find(func_vf.get());
if ( ia == inline_ables.end() )
return c;
if ( c->IsInWhen() ) {
// Don't inline these, as doing so requires propagating
// the in-when attribute to the inlined function body.
skipped_inlining.insert(func_vf);
skipped_inlining.insert(func_vf.get());
return c;
}
@ -189,21 +329,32 @@ ExprPtr Inliner::CheckForInlining(CallExprPtr c) {
// BiFs, which won't happen here, but instead to script functions that
// are misusing/abusing the loophole.)
if ( function->GetType()->Params()->NumFields() == 1 && c->Args()->Exprs().size() != 1 ) {
skipped_inlining.insert(func_vf);
skipped_inlining.insert(func_vf.get());
return c;
}
// We're going to inline the body, unless it's too large.
auto body = func_vf->GetBodies()[0].stmts; // there's only 1 body
auto scope = func_vf->GetScope();
auto ie = DoInline(func_vf, body, c->ArgsPtr(), scope, ia->second);
if ( ie ) {
ie->SetOriginal(c);
did_inline.insert(func_vf.get());
}
return ie;
}
ExprPtr Inliner::DoInline(ScriptFuncPtr sf, StmtPtr body, ListExprPtr args, ScopePtr scope, const ProfileFunc* pf) {
// Inline the body, unless it's too large.
auto oi = body->GetOptInfo();
if ( num_stmts + oi->num_stmts + num_exprs + oi->num_exprs > MAX_INLINE_SIZE ) {
skipped_inlining.insert(func_vf);
skipped_inlining.insert(sf.get());
return nullptr; // signals "stop inlining"
}
did_inline.insert(func_vf);
num_stmts += oi->num_stmts;
num_exprs += oi->num_exprs;
@ -220,10 +371,8 @@ ExprPtr Inliner::CheckForInlining(CallExprPtr c) {
// with the scope, which gives us all the variables declared in
// the function, *using the knowledge that the parameters are
// declared first*.
auto scope = func_vf->GetScope();
auto& pf = ia->second;
auto& vars = scope->OrderedVars();
int nparam = func_vf->GetType()->Params()->NumFields();
int nparam = sf->GetType()->Params()->NumFields();
std::vector<IDPtr> params;
std::vector<bool> param_is_modified;
@ -237,7 +386,7 @@ ExprPtr Inliner::CheckForInlining(CallExprPtr c) {
// Recursively inline the body. This is safe to do because we've
// ensured there are no recursive loops ... but we have to be
// careful in accounting for the frame sizes.
int frame_size = func_vf->FrameSize();
int frame_size = sf->FrameSize();
int hold_curr_frame_size = curr_frame_size;
curr_frame_size = frame_size;
@ -255,12 +404,10 @@ ExprPtr Inliner::CheckForInlining(CallExprPtr c) {
else
max_inlined_frame_size = hold_max_inlined_frame_size;
auto t = c->GetType();
auto ie = make_intrusive<InlineExpr>(c->ArgsPtr(), std::move(params), std::move(param_is_modified), body_dup,
curr_frame_size, t);
ie->SetOriginal(c);
auto t = scope->GetReturnType();
return ie;
ASSERT(params.size() == args->Exprs().size());
return make_intrusive<InlineExpr>(args, params, param_is_modified, body_dup, curr_frame_size, t);
}
} // namespace zeek::detail

View file

@ -37,9 +37,34 @@ protected:
// recursively inlines eligible ones.
void Analyze();
// Maps an event handler body to its corresponding FuncInfo. For the
// latter we use a cursor rather than a direct reference or pointer
// because the collection of FuncInfo's are maintained in a vector and
// can wind up moving around in memory.
using BodyInfo = std::unordered_map<const Stmt*, size_t>;
// Goes through all the event handlers and coalesces those with
// multiple bodies into a single "alternative" body.
void CoalesceEventHandlers();
// For a given event handler, collection of bodies, and associated
// function information, creates a new FuncInfo entry reflecting an
// alternative body for the event handler with all of the bodies
// coalesced into a single new body.
void CoalesceEventHandlers(ScriptFuncPtr sf, const std::vector<Func::Body>& bodies, const BodyInfo& body_to_info);
// Recursively inlines any calls associated with the given function.
void InlineFunction(FuncInfo* f);
// Performs common functionality prior to inlining a call body.
void PreInline(StmtOptInfo* oi, size_t frame_size);
// Performs common functionality that comes after inlining a call body.
void PostInline(StmtOptInfo* oi, ScriptFuncPtr f);
// Inlines the given body using the given arguments.
ExprPtr DoInline(ScriptFuncPtr sf, StmtPtr body, ListExprPtr args, ScopePtr scope, const ProfileFunc* pf);
// Information about all of the functions (and events/hooks) in
// the full set of scripts.
std::vector<FuncInfo>& funcs;

View file

@ -473,17 +473,15 @@ static void analyze_scripts_for_ZAM(std::unique_ptr<ProfileFuncs>& pfs) {
bool did_one = false;
for ( auto& f : funcs ) {
if ( ! f.ShouldAnalyze() )
continue;
auto func = f.Func();
auto l = lambdas.find(func);
bool is_lambda = l != lambdas.end();
if ( ! analysis_options.only_funcs.empty() || ! analysis_options.only_files.empty() ) {
if ( ! should_analyze(f.FuncPtr(), f.Body()) )
continue;
}
else if ( ! analysis_options.compile_all && ! is_lambda && inl && inl->WasFullyInlined(func) &&
func_used_indirectly.count(func) == 0 ) {
if ( ! analysis_options.compile_all && ! is_lambda && inl && inl->WasFullyInlined(func) &&
func_used_indirectly.count(func) == 0 ) {
// No need to compile as it won't be called directly.
// We'd like to zero out the body to recover the
// memory, but a *few* such functions do get called,
@ -563,7 +561,7 @@ void analyze_scripts(bool no_unused_warnings) {
if ( should_analyze(func.FuncPtr(), func.Body()) )
have_one_to_do = true;
else
func.SetSkip(true);
func.SetShouldNotAnalyze();
if ( ! have_one_to_do )
reporter->FatalError("no matching functions/files for C++ compilation");

View file

@ -134,13 +134,18 @@ public:
const ProfileFunc* Profile() const { return pf.get(); }
std::shared_ptr<ProfileFunc> ProfilePtr() const { return pf; }
void SetScope(ScopePtr new_scope) { scope = std::move(new_scope); }
void SetBody(StmtPtr new_body) { body = std::move(new_body); }
void SetProfile(std::shared_ptr<ProfileFunc> _pf) { pf = std::move(_pf); }
bool ShouldAnalyze() const { return should_analyze; }
void SetShouldNotAnalyze() {
should_analyze = false;
skip = true;
}
// The following provide a way of marking FuncInfo's as
// should-be-skipped for script optimization, generally because
// the function body has a property that a given script optimizer
// doesn't know how to deal with. Defaults to don't-skip.
// should-be-skipped for a given phase of script optimization.
bool ShouldSkip() const { return skip; }
void SetSkip(bool should_skip) { skip = should_skip; }
@ -151,10 +156,39 @@ protected:
std::shared_ptr<ProfileFunc> pf;
int priority;
// Whether to skip optimizing this function.
// Whether to analyze this function at all, per optimization selection
// via --optimize-file/--optimize-func. If those flags aren't used,
// then this will remain true, given that both ZAM and -O gen-C++ are
// feature-complete.
bool should_analyze = true;
// Whether to skip optimizing this function in a given context. May be
// altered during optimization.
bool skip = false;
};
// ScriptFunc subclass that runs a single (coalesced) body if possible,
// otherwise delegates to the original function with multiple bodies.
class CoalescedScriptFunc : public ScriptFunc {
public:
CoalescedScriptFunc(StmtPtr merged_body, ScopePtr scope, ScriptFuncPtr orig_func)
: ScriptFunc(orig_func->Name(), orig_func->GetType(), {merged_body}, {0}), orig_func(orig_func) {
SetScope(scope);
};
ValPtr Invoke(zeek::Args* args, Frame* parent) const override {
// If the original function has all bodies enabled, run our
// coalesced one, otherwise delegate.
if ( orig_func->HasAllBodiesEnabled() )
return ScriptFunc::Invoke(args, parent);
return orig_func->Invoke(args, parent);
}
private:
ScriptFuncPtr orig_func;
};
// We track which functions are definitely not recursive. We do this
// as the negative, rather than tracking functions known to be recursive,
// so that if we don't do the analysis at all (it's driven by inlining),