more extensive ZAM inlining & compilation of lambdas

src/Expr.cc

@@ -4825,10 +4825,19 @@ ScopePtr LambdaExpr::GetScope() const
 	return ingredients->Scope();
 	}
 
+void LambdaExpr::ReplaceBody(StmtPtr new_body)
+	{
+	ingredients->ReplaceBody(std::move(new_body));
+	}
+
 ValPtr LambdaExpr::Eval(Frame* f) const
 	{
 	auto lamb = make_intrusive<ScriptFunc>(ingredients->GetID());
 
+	// Use the primary function as the source of the frame size
+	// and function body, rather than the ingredients, since script
+	// optimization might have changed the former but not the latter.
+	lamb->SetFrameSize(primary_func->FrameSize());
 	StmtPtr body = primary_func->GetBodies()[0].stmts;
 
 	if ( run_state::is_parsing )

src/Expr.h

@@ -1482,6 +1482,8 @@ public:
 
 	const FunctionIngredientsPtr& Ingredients() const { return ingredients; }
 
+	void ReplaceBody(StmtPtr new_body);
+
 	bool IsReduced(Reducer* c) const override;
 	bool HasReducedOps(Reducer* c) const override;
 	ExprPtr Reduce(Reducer* c, StmtPtr& red_stmt) override;

src/Func.h

@@ -375,6 +375,7 @@ public:
 	const IDPtr& GetID() const { return id; }
 
 	const StmtPtr& Body() const { return body; }
+	void ReplaceBody(StmtPtr new_body) { body = std::move(new_body); }
 
 	const auto& Inits() const { return inits; }
 	void ClearInits() { inits.clear(); }

src/script_opt/Expr.cc

@@ -2717,6 +2717,7 @@ ExprPtr InlineExpr::Reduce(Reducer* c, StmtPtr& red_stmt)
 	red_stmt = nullptr;
 
 	auto args_list = args->Exprs();
+	auto ret_val = c->PushInlineBlock(type);
 
 	loop_over_list(args_list, i)
 		{
@@ -2730,7 +2731,6 @@ ExprPtr InlineExpr::Reduce(Reducer* c, StmtPtr& red_stmt)
 		red_stmt = MergeStmts(red_stmt, arg_red_stmt, assign_stmt);
 		}
 
-	auto ret_val = c->PushInlineBlock(type);
 	body = body->Reduce(c);
 	c->PopInlineBlock();
 

src/script_opt/Inline.cc

@@ -170,11 +170,6 @@ ExprPtr Inliner::CheckForInlining(CallExprPtr c)
 		// We don't inline indirect calls.
 		return c;
 
-	if ( c->IsInWhen() )
-		// Don't inline these, as doing so requires propagating
-		// the in-when attribute to the inlined function body.
-		return c;
-
 	auto n = f->AsNameExpr();
 	auto func = n->Id();
 
@@ -190,22 +185,38 @@ ExprPtr Inliner::CheckForInlining(CallExprPtr c)
 	if ( function->GetKind() != Func::SCRIPT_FUNC )
 		return c;
 
-	// Check for mismatches in argument count due to single-arg-of-type-any
-	// loophole used for variadic BiFs.
-	if ( function->GetType()->Params()->NumFields() == 1 && c->Args()->Exprs().size() != 1 )
-		return c;
-
 	auto func_vf = static_cast<ScriptFunc*>(function);
 
 	if ( inline_ables.count(func_vf) == 0 )
 		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);
+		return c;
+		}
+
+	// Check for mismatches in argument count due to single-arg-of-type-any
+	// loophole used for variadic BiFs.  (The issue isn't calls to the
+	// 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);
+		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 oi = body->GetOptInfo();
 
 	if ( num_stmts + oi->num_stmts + num_exprs + oi->num_exprs > MAX_INLINE_SIZE )
-		return nullptr;
+		{
+		skipped_inlining.insert(func_vf);
+		return nullptr; // signals "stop inlining"
+		}
 
 	num_stmts += oi->num_stmts;
 	num_exprs += oi->num_exprs;

src/script_opt/Inline.h

@@ -32,7 +32,10 @@ public:
 	ExprPtr CheckForInlining(CallExprPtr c);
 
 	// True if the given function has been inlined.
-	bool WasInlined(const Func* f) { return inline_ables.count(f) > 0; }
+	bool WasInlined(const Func* f)
+		{
+		return inline_ables.count(f) > 0 && skipped_inlining.count(f) == 0;
+		}
 
 protected:
 	// Driver routine that analyzes all of the script functions and
@@ -49,6 +52,10 @@ protected:
 	// Functions that we've determined to be suitable for inlining.
 	std::unordered_set<const Func*> inline_ables;
 
+	// Functions that we didn't fully inline, so require separate
+	// compilation.
+	std::unordered_set<const Func*> skipped_inlining;
+
 	// As we do inlining for a given function, this tracks the
 	// largest frame size of any inlined function.
 	int max_inlined_frame_size;

src/script_opt/Reduce.cc

@@ -4,19 +4,36 @@
 
 #include "zeek/Desc.h"
 #include "zeek/Expr.h"
+#include "zeek/Func.h"
 #include "zeek/ID.h"
 #include "zeek/Reporter.h"
 #include "zeek/Scope.h"
 #include "zeek/Stmt.h"
 #include "zeek/Var.h"
 #include "zeek/script_opt/ExprOptInfo.h"
-#include "zeek/script_opt/ProfileFunc.h"
 #include "zeek/script_opt/StmtOptInfo.h"
 #include "zeek/script_opt/TempVar.h"
 
 namespace zeek::detail
 	{
 
+Reducer::Reducer(const ScriptFunc* func)
+	{
+	auto& ft = func->GetType();
+
+	// Track the parameters so we don't remap them.
+	int num_params = ft->Params()->NumFields();
+	auto& scope_vars = current_scope()->OrderedVars();
+
+	for ( auto i = 0; i < num_params; ++i )
+		tracked_ids.insert(scope_vars[i].get());
+
+	// Now include any captures.
+	if ( ft->GetCaptures() )
+		for ( auto& c : *ft->GetCaptures() )
+			tracked_ids.insert(c.Id().get());
+	}
+
 StmtPtr Reducer::Reduce(StmtPtr s)
 	{
 	reduction_root = std::move(s);
@@ -57,10 +74,9 @@ NameExprPtr Reducer::UpdateName(NameExprPtr n)
 	return ne;
 	}
 
-bool Reducer::NameIsReduced(const NameExpr* n) const
+bool Reducer::NameIsReduced(const NameExpr* n)
 	{
-	auto id = n->Id();
-	return inline_block_level == 0 || id->IsGlobal() || IsTemporary(id) || IsNewLocal(n);
+	return ID_IsReducedOrTopLevel(n->Id());
 	}
 
 void Reducer::UpdateIDs(IDPList* ids)
@@ -69,7 +85,7 @@ void Reducer::UpdateIDs(IDPList* ids)
 		{
 		IDPtr id = {NewRef{}, (*ids)[i]};
 
-		if ( ! ID_IsReduced(id) )
+		if ( ! ID_IsReducedOrTopLevel(id) )
 			{
 			Unref((*ids)[i]);
 			(*ids)[i] = UpdateID(id).release();
@@ -80,7 +96,7 @@ void Reducer::UpdateIDs(IDPList* ids)
 void Reducer::UpdateIDs(std::vector<IDPtr>& ids)
 	{
 	for ( auto& id : ids )
-		if ( ! ID_IsReduced(id) )
+		if ( ! ID_IsReducedOrTopLevel(id) )
 			id = UpdateID(id);
 	}
 
@@ -104,15 +120,27 @@ bool Reducer::IDsAreReduced(const std::vector<IDPtr>& ids) const
 
 IDPtr Reducer::UpdateID(IDPtr id)
 	{
-	if ( ID_IsReduced(id) )
+	if ( ID_IsReducedOrTopLevel(id) )
 		return id;
 
 	return FindNewLocal(id);
 	}
 
+bool Reducer::ID_IsReducedOrTopLevel(const ID* id)
+	{
+	if ( inline_block_level == 0 )
+		{
+		tracked_ids.insert(id);
+		return true;
+		}
+
+	return ID_IsReduced(id);
+	}
+
 bool Reducer::ID_IsReduced(const ID* id) const
 	{
-	return inline_block_level == 0 || id->IsGlobal() || IsTemporary(id) || IsNewLocal(id);
+	return inline_block_level == 0 || tracked_ids.count(id) > 0 || id->IsGlobal() ||
+	       IsTemporary(id);
 	}
 
 NameExprPtr Reducer::GenInlineBlockName(const IDPtr& id)
@@ -126,6 +154,8 @@ NameExprPtr Reducer::PushInlineBlock(TypePtr type)
 	{
 	++inline_block_level;
 
+	block_locals.emplace_back(std::unordered_map<const ID*, IDPtr>());
+
 	if ( ! type || type->Tag() == TYPE_VOID )
 		return nullptr;
 
@@ -133,11 +163,13 @@ NameExprPtr Reducer::PushInlineBlock(TypePtr type)
 	ret_id->SetType(type);
 	ret_id->GetOptInfo()->SetTemp();
 
+	ret_vars.insert(ret_id.get());
+
 	// Track this as a new local *if* we're in the outermost inlining
 	// block.  If we're recursively deeper into inlining, then this
 	// variable will get mapped to a local anyway, so no need.
 	if ( inline_block_level == 1 )
-		new_locals.insert(ret_id.get());
+		AddNewLocal(ret_id);
 
 	return GenInlineBlockName(ret_id);
 	}
@@ -145,6 +177,18 @@ NameExprPtr Reducer::PushInlineBlock(TypePtr type)
 void Reducer::PopInlineBlock()
 	{
 	--inline_block_level;
+
+	for ( auto& l : block_locals.back() )
+		{
+		auto key = l.first;
+		auto prev = l.second;
+		if ( prev )
+			orig_to_new_locals[key] = prev;
+		else
+			orig_to_new_locals.erase(key);
+		}
+
+	block_locals.pop_back();
 	}
 
 bool Reducer::SameVal(const Val* v1, const Val* v2) const
@@ -750,6 +794,12 @@ IDPtr Reducer::FindNewLocal(const IDPtr& id)
 	return GenLocal(id);
 	}
 
+void Reducer::AddNewLocal(const IDPtr& l)
+	{
+	new_locals.insert(l.get());
+	tracked_ids.insert(l.get());
+	}
+
 IDPtr Reducer::GenLocal(const IDPtr& orig)
 	{
 	if ( Optimizing() )
@@ -758,6 +808,9 @@ IDPtr Reducer::GenLocal(const IDPtr& orig)
 	if ( omitted_stmts.size() > 0 )
 		reporter->InternalError("Generating a new local while pruning statements");
 
+	// Make sure the identifier is not being re-re-mapped.
+	ASSERT(strchr(orig->Name(), '.') == nullptr);
+
 	char buf[8192];
 	int n = new_locals.size();
 	snprintf(buf, sizeof buf, "%s.%d", orig->Name(), n);
@@ -769,9 +822,16 @@ IDPtr Reducer::GenLocal(const IDPtr& orig)
 	if ( orig->GetOptInfo()->IsTemp() )
 		local_id->GetOptInfo()->SetTemp();
 
-	new_locals.insert(local_id.get());
+	IDPtr prev;
+	if ( orig_to_new_locals.count(orig.get()) )
+		prev = orig_to_new_locals[orig.get()];
+
+	AddNewLocal(local_id);
 	orig_to_new_locals[orig.get()] = local_id;
 
+	if ( ! block_locals.empty() && ret_vars.count(orig.get()) == 0 )
+		block_locals.back()[orig.get()] = prev;
+
 	return local_id;
 	}
 

src/script_opt/Reduce.h

@@ -16,7 +16,7 @@ class TempVar;
 class Reducer
 	{
 public:
-	Reducer() { }
+	Reducer(const ScriptFunc* func);
 
 	StmtPtr Reduce(StmtPtr s);
 
@@ -27,7 +27,7 @@ public:
 	ExprPtr GenTemporaryExpr(const TypePtr& t, ExprPtr rhs);
 
 	NameExprPtr UpdateName(NameExprPtr n);
-	bool NameIsReduced(const NameExpr* n) const;
+	bool NameIsReduced(const NameExpr* n);
 
 	void UpdateIDs(IDPList* ids);
 	bool IDsAreReduced(const IDPList* ids) const;
@@ -39,6 +39,10 @@ public:
 	bool ID_IsReduced(const IDPtr& id) const { return ID_IsReduced(id.get()); }
 	bool ID_IsReduced(const ID* id) const;
 
+	// A version of ID_IsReduced() that tracks top-level variables, too.
+	bool ID_IsReducedOrTopLevel(const IDPtr& id) { return ID_IsReducedOrTopLevel(id.get()); }
+	bool ID_IsReducedOrTopLevel(const ID* id);
+
 	// This is called *prior* to pushing a new inline block, in
 	// order to generate the equivalent of function parameters.
 	NameExprPtr GenInlineBlockName(const IDPtr& id);
@@ -205,6 +209,8 @@ protected:
 	IDPtr FindNewLocal(const IDPtr& id);
 	IDPtr FindNewLocal(const NameExprPtr& n) { return FindNewLocal(n->IdPtr()); }
 
+	void AddNewLocal(const IDPtr& l);
+
 	// Generate a new local to use in lieu of the original (seen
 	// in an inlined block).  The difference is that the new
 	// version has a distinct name and has a correct frame offset
@@ -228,6 +234,9 @@ protected:
 	// variable, if it corresponds to one.
 	std::unordered_map<const ID*, std::shared_ptr<TempVar>> ids_to_temps;
 
+	// Identifiers that we're tracking (and don't want to replace).
+	IDSet tracked_ids;
+
 	// Local variables created during reduction/optimization.
 	IDSet new_locals;
 
@@ -250,10 +259,18 @@ protected:
 	// Maps statements to replacements constructed during optimization.
 	std::unordered_map<const Stmt*, StmtPtr> replaced_stmts;
 
+	// Tracks return variables we've created.
+	IDSet ret_vars;
+
 	// Tracks whether we're inside an inline block, and if so then
 	// how deeply.
 	int inline_block_level = 0;
 
+	// Tracks locals introduced in the current block, remembering
+	// their previous replacement value (per "orig_to_new_locals"),
+	// if any.  When we pop the block, we restore the previous mapping.
+	std::vector<std::unordered_map<const ID*, IDPtr>> block_locals;
+
 	// Tracks how deeply we are in "bifurcation", i.e., duplicating
 	// code for if-else cascades.  We need to cap this at a certain
 	// depth or else we can get functions whose size blows up
@@ -265,7 +282,7 @@ protected:
 	IDSet constant_vars;
 
 	// Statement at which the current reduction started.
-	StmtPtr reduction_root = nullptr;
+	StmtPtr reduction_root;
 
 	// Statement we're currently working on.
 	const Stmt* curr_stmt = nullptr;

src/script_opt/ScriptOpt.cc

@@ -36,7 +36,7 @@ static ZAMCompiler* ZAM = nullptr;
 static bool generating_CPP = false;
 static std::string CPP_dir; // where to generate C++ code
 
-static std::unordered_set<const ScriptFunc*> lambdas;
+static std::unordered_map<const ScriptFunc*, LambdaExpr*> lambdas;
 static std::unordered_set<const ScriptFunc*> when_lambdas;
 static ScriptFuncPtr global_stmts;
 
@@ -52,7 +52,7 @@ void analyze_lambda(LambdaExpr* l)
 	{
 	auto& pf = l->PrimaryFunc();
 	analyze_func(pf);
-	lambdas.insert(pf.get());
+	lambdas[pf.get()] = l;
 	}
 
 void analyze_when_lambda(LambdaExpr* l)
@@ -185,7 +185,7 @@ static void optimize_func(ScriptFunc* f, std::shared_ptr<ProfileFunc> pf, ScopeP
 
 	push_existing_scope(scope);
 
-	auto rc = std::make_shared<Reducer>();
+	auto rc = std::make_shared<Reducer>(f);
 	auto new_body = rc->Reduce(body);
 
 	if ( reporter->Errors() > 0 )
@@ -496,9 +496,7 @@ static void analyze_scripts_for_ZAM(std::unique_ptr<ProfileFuncs>& pfs)
 
 	if ( inl )
 		{
-		for ( auto& f : funcs )
-			{
-			for ( const auto& g : f.Profile()->Globals() )
+		for ( auto& g : pfs->Globals() )
 			{
 			if ( g->GetType()->Tag() != TYPE_FUNC )
 				continue;
@@ -513,14 +511,14 @@ static void analyze_scripts_for_ZAM(std::unique_ptr<ProfileFuncs>& pfs)
 				func_used_indirectly.insert(func);
 			}
 		}
-		}
 
 	bool did_one = false;
 
 	for ( auto& f : funcs )
 		{
 		auto func = f.Func();
-		bool is_lambda = lambdas.count(func) > 0;
+		auto l = lambdas.find(func);
+		bool is_lambda = l != lambdas.end();
 
 		if ( ! analysis_options.only_funcs.empty() || ! analysis_options.only_files.empty() )
 			{
@@ -539,6 +537,9 @@ static void analyze_scripts_for_ZAM(std::unique_ptr<ProfileFuncs>& pfs)
 		optimize_func(func, f.ProfilePtr(), f.Scope(), new_body);
 		f.SetBody(new_body);
 
+		if ( is_lambda )
+			l->second->ReplaceBody(new_body);
+
 		did_one = true;
 		}
 

src/script_opt/ZAM/Ops.in

@@ -1594,7 +1594,13 @@ macro WhenCall(func)
 		for ( auto i = 0; i < n; ++i )
 			args.push_back(aux->ToVal(frame, i));
 		f->SetCall(z.call_expr);
+		/* It's possible that this function will call another that
+		 * itself returns null because *it* is the actual blocker.
+		 * That will set ZAM_error, which we need to ignore.
+		 */
+		auto hold_ZAM_error = ZAM_error;
 		vp = func->Invoke(&args, f);
+		ZAM_error = hold_ZAM_error;
 		f->SetTriggerAssoc(current_assoc);
 		if ( ! vp )
 			throw ZAMDelayedCallException();