speedups for compilation of initializers in -O gen-C++ generated code

src/script_opt/CPP/Compile.h

@@ -143,6 +143,7 @@ public:
     // cohort associated with a given type.
     int TypeOffset(const TypePtr& t) { return GI_Offset(RegisterType(t)); }
     int TypeCohort(const TypePtr& t) { return GI_Cohort(RegisterType(t)); }
+    int TypeFinalCohort(const TypePtr& t) { return GI_FinalCohort(RegisterType(t)); }
 
     // Tracks a Zeek ValPtr used as a constant value.  These occur
     // in two contexts: directly as constant expressions, and indirectly
@@ -963,6 +964,7 @@ private:
     // associated with an initialization.
     int GI_Offset(const std::shared_ptr<CPP_InitInfo>& gi) const { return gi ? gi->Offset() : -1; }
     int GI_Cohort(const std::shared_ptr<CPP_InitInfo>& gi) const { return gi ? gi->InitCohort() : 0; }
+    int GI_FinalCohort(const std::shared_ptr<CPP_InitInfo>& gi) const { return gi ? gi->FinalInitCohort() : 0; }
 
     // Generate code to initialize the mappings for record field
     // offsets for field accesses into regions of records that

src/script_opt/CPP/Inits.cc

@@ -166,7 +166,7 @@ void CPPCompile::InitializeConsts() {
     StartBlock();
 
     for ( const auto& c : consts )
-        Emit("CPP_ValElem(%s, %s),", TypeTagName(c.first), Fmt(c.second));
+        Emit("{%s, %s},", TypeTagName(c.first), Fmt(c.second));
 
     EndBlock(true);
 }

src/script_opt/CPP/InitsInfo.cc

@@ -7,6 +7,7 @@
 #include "zeek/ZeekString.h"
 #include "zeek/script_opt/CPP/Attrs.h"
 #include "zeek/script_opt/CPP/Compile.h"
+#include "zeek/script_opt/CPP/RuntimeInits.h"
 
 using namespace std;
 
@@ -38,6 +39,13 @@ void CPP_InitsInfo::GenerateInitializers(CPPCompile* c) {
     c->Emit("%s %s = %s(%s, %s,", gt, InitializersName(), gt, base_name, Fmt(offset_set));
 
     c->IndentUp();
+    GenerateCohorts(c);
+    c->IndentDown();
+
+    c->Emit(");");
+}
+
+void CPP_InitsInfo::GenerateCohorts(CPPCompile* c) {
     c->Emit("{");
 
     int n = 0;
@@ -47,7 +55,7 @@ void CPP_InitsInfo::GenerateInitializers(CPPCompile* c) {
         if ( ++n > 1 )
             c->Emit("");
 
-        if ( cohort.size() == 1 && ! IsCompound() )
+        if ( cohort.size() == 1 && ! UsesCompoundVectors() )
             BuildCohort(c, cohort);
         else {
             c->Emit("{");
@@ -57,8 +65,6 @@ void CPP_InitsInfo::GenerateInitializers(CPPCompile* c) {
     }
 
     c->Emit("}");
-    c->IndentDown();
-    c->Emit(");");
 }
 
 void CPP_InitsInfo::BuildOffsetSet(CPPCompile* c) {
@@ -80,25 +86,25 @@ void CPP_InitsInfo::BuildOffsetSet(CPPCompile* c) {
     offset_set = c->IndMgr().AddIndices(offsets_vec);
 }
 
-void CPP_InitsInfo::BuildCohort(CPPCompile* c, std::vector<std::shared_ptr<CPP_InitInfo>>& cohort) {
-    int n = 0;
+static std::string describe_initializer(const Obj* o) {
+    auto od = obj_desc(o);
 
+    // Escape any embedded comment characters.
+    od = regex_replace(od, std::regex("/\\*"), "<<SLASH-STAR>>");
+    od = regex_replace(od, std::regex("\\*/"), "<<STAR-SLASH>>");
+
+    return od;
+}
+
+void CPP_InitsInfo::BuildCohort(CPPCompile* c, std::vector<std::shared_ptr<CPP_InitInfo>>& cohort) {
     for ( auto& co : cohort ) {
         vector<string> ivs;
         auto o = co->InitObj();
-        if ( o ) {
-            auto od = obj_desc(o);
-
-            // Escape any embedded comment characters.
-            od = regex_replace(od, std::regex("/\\*"), "<<SLASH-STAR>>");
-            od = regex_replace(od, std::regex("\\*/"), "<<STAR-SLASH>>");
-
-            c->Emit("/* #%s: Initializing %s: */", Fmt(co->Offset()), od);
-        }
+        if ( o )
+            c->Emit("/* #%s: Initializing %s: */", Fmt(co->Offset()), describe_initializer(o));
 
         co->InitializerVals(ivs);
         BuildCohortElement(c, co->InitializerType(), ivs);
-        ++n;
     }
 }
 
@@ -117,12 +123,50 @@ void CPP_InitsInfo::BuildCohortElement(CPPCompile* c, string init_type, vector<s
     c->Emit("std::make_shared<%s>(%s),", init_type, full_init);
 }
 
+void CPP_CompoundInitsInfo::GenerateInitializers(CPPCompile* c) {
+    c->Emit("");
+    c->Emit("static int %s_init[] = {", tag);
+    int n = 0;
+
+    c->IndentUp();
+
+    for ( auto& cohort : instances ) {
+        if ( ++n > 1 )
+            c->Emit("");
+
+        // Figure out the size of the cohort.
+        for ( auto& co : cohort ) {
+            auto o = co->InitObj();
+            if ( o )
+                c->Emit("/* #%s: Initializing %s: */", Fmt(co->Offset()), describe_initializer(o));
+
+            vector<string> ivs;
+            co->InitializerVals(ivs);
+            c->Emit(Fmt(int(ivs.size())) + ",");
+            BuildCohortElement(c, co->InitializerType(), ivs);
+        }
+
+        static const auto end_of_vv = Fmt(END_OF_VEC_VEC) + ",";
+        c->Emit(end_of_vv);
+    }
+
+    static const auto end_of_vvv = Fmt(END_OF_VEC_VEC_VEC) + ",";
+    c->Emit(end_of_vvv);
+
+    c->IndentDown();
+    c->Emit("};");
+
+    CPP_InitsInfo::GenerateInitializers(c);
+}
+
+void CPP_CompoundInitsInfo::GenerateCohorts(CPPCompile* c) { c->Emit("%s_init", tag); }
+
 void CPP_CompoundInitsInfo::BuildCohortElement(CPPCompile* c, string init_type, vector<string>& ivs) {
     string init_line;
     for ( auto& iv : ivs )
-        init_line += iv + ", ";
+        init_line += iv + ",";
 
-    c->Emit("{ %s},", init_line);
+    c->Emit("%s", init_line);
 }
 
 void CPP_BasicConstInitsInfo::BuildCohortElement(CPPCompile* c, string init_type, vector<string>& ivs) {
@@ -174,7 +218,7 @@ PatternConstInfo::PatternConstInfo(CPPCompile* c, ValPtr v) : CPP_InitInfo(v) {
 CompoundItemInfo::CompoundItemInfo(CPPCompile* _c, ValPtr v) : CPP_InitInfo(v), c(_c) {
     auto& t = v->GetType();
     type = c->TypeOffset(t);
-    init_cohort = c->TypeCohort(t) + 1;
+    init_cohort = c->TypeFinalCohort(t) + 1;
 }
 
 ListConstInfo::ListConstInfo(CPPCompile* _c, ValPtr v) : CompoundItemInfo(_c) {
@@ -400,7 +444,11 @@ void TypeTypeInfo::AddInitializerVals(std::vector<std::string>& ivs) const {
 }
 
 VectorTypeInfo::VectorTypeInfo(CPPCompile* _c, TypePtr _t) : AbstractTypeInfo(_c, std::move(_t)) {
-    yield = t->Yield();
+    auto vt = t->AsVectorType();
+    if ( vt->IsUnspecifiedVector() )
+        yield = base_type(TYPE_VOID);
+    else
+        yield = t->Yield();
     auto gi = c->RegisterType(yield);
     if ( gi )
         init_cohort = gi->InitCohort();
@@ -552,7 +600,8 @@ void IndicesManager::Generate(CPPCompile* c) {
             c->Emit(line);
     }
 
-    c->Emit("-1");
+    static const auto end_of_vv = Fmt(END_OF_VEC_VEC);
+    c->Emit(end_of_vv);
     c->EndBlock(true);
 }
 

src/script_opt/CPP/InitsInfo.h

@@ -133,10 +133,10 @@ public:
     // Sets the associated C++ type.
     virtual void SetCPPType(std::string ct) { CPP_type = std::move(ct); }
 
-    // Whether this initializer is in terms of compound objects. Used
+    // Whether this initializer is in terms of compound vectors. Used
     // for avoiding compiler warnings about singleton initializations in
     // braces.
-    virtual bool IsCompound() const { return false; }
+    virtual bool UsesCompoundVectors() const { return false; }
 
     // Returns the type associated with the table used for initialization
     // (i.e., this is the type of the global returned by InitializersName()).
@@ -146,9 +146,11 @@ public:
     void AddInstance(std::shared_ptr<CPP_InitInfo> g);
 
     // Emit code to populate the table used to initialize this collection.
-    void GenerateInitializers(CPPCompile* c);
+    virtual void GenerateInitializers(CPPCompile* c);
 
 protected:
+    virtual void GenerateCohorts(CPPCompile* c);
+
     // Computes offset_set - see below.
     void BuildOffsetSet(CPPCompile* c);
 
@@ -214,7 +216,7 @@ public:
         BuildInitType();
     }
 
-    bool IsCompound() const override { return true; }
+    bool UsesCompoundVectors() const override { return true; }
 
 private:
     void BuildInitType() { inits_type = std::string("CPP_CustomInits<") + CPPType() + ">"; }
@@ -236,7 +238,7 @@ public:
             inits_type = std::string("CPP_BasicConsts<") + CPP_type + ", " + c_type + ", " + tag + "Val>";
     }
 
-    bool IsCompound() const override { return false; }
+    bool UsesCompoundVectors() const override { return false; }
 
     void BuildCohortElement(CPPCompile* c, std::string init_type, std::vector<std::string>& ivs) override;
 };
@@ -254,7 +256,12 @@ public:
             inits_type = std::string("CPP_IndexedInits<") + CPPType() + ">";
     }
 
-    bool IsCompound() const override { return true; }
+    // This isn't true (anymore) because we separately build up the compound
+    // vectors needed for the initialization.
+    bool UsesCompoundVectors() const override { return false; }
+
+    void GenerateInitializers(CPPCompile* c) override;
+    void GenerateCohorts(CPPCompile* c) override;
 
     void BuildCohortElement(CPPCompile* c, std::string init_type, std::vector<std::string>& ivs) override;
 };

src/script_opt/CPP/RuntimeInits.cc

@@ -465,12 +465,12 @@ void CPP_GlobalInit::Generate(InitsManager* im, std::vector<void*>& /* inits_vec
         global->SetAttrs(im->Attributes(attrs));
 }
 
-void generate_indices_set(int* inits, std::vector<std::vector<int>>& indices_set) {
+size_t generate_indices_set(int* inits, std::vector<std::vector<int>>& indices_set) {
     // First figure out how many groups of indices there are, so we
     // can pre-allocate the outer vector.
     auto i_ptr = inits;
     int num_inits = 0;
-    while ( *i_ptr >= 0 ) {
+    while ( *i_ptr != END_OF_VEC_VEC && *i_ptr != END_OF_VEC_VEC_VEC ) {
         ++num_inits;
         int n = *i_ptr;
         i_ptr += n + 1; // skip over vector elements
@@ -479,7 +479,7 @@ void generate_indices_set(int* inits, std::vector<std::vector<int>>& indices_set
     indices_set.reserve(num_inits);
 
     i_ptr = inits;
-    while ( *i_ptr >= 0 ) {
+    while ( *i_ptr != END_OF_VEC_VEC ) {
         int n = *i_ptr;
         ++i_ptr;
         std::vector<int> indices;
@@ -490,6 +490,20 @@ void generate_indices_set(int* inits, std::vector<std::vector<int>>& indices_set
 
         indices_set.emplace_back(std::move(indices));
     }
+
+    return i_ptr - inits + 1;
+}
+
+std::vector<std::vector<std::vector<int>>> generate_indices_set(int* inits) {
+    std::vector<std::vector<std::vector<int>>> indices_set;
+
+    while ( *inits != END_OF_VEC_VEC_VEC ) {
+        std::vector<std::vector<int>> cohort_inits;
+        inits += generate_indices_set(inits, cohort_inits);
+        indices_set.push_back(std::move(cohort_inits));
+    }
+
+    return indices_set;
 }
 
 } // namespace zeek::detail

src/script_opt/CPP/RuntimeInits.h

@@ -19,6 +19,28 @@ using FuncValPtr = IntrusivePtr<FuncVal>;
 
 class InitsManager;
 
+// Helper function that takes a (large) array of int's and from them
+// constructs the corresponding vector-of-vector-of-indices.  Each
+// vector-of-indices is represented first by an int specifying its
+// size, and then that many int's for its values.  We recognize the
+// end of the array upon encountering a "size" entry of END_OF_VEC_VEC.
+//
+// Returns how many elements were processed out of "inits", including its
+// terminator.
+extern size_t generate_indices_set(int* inits, std::vector<std::vector<int>>& indices_set);
+
+// The same but for one more level of vector construction. The source array
+// has sub-arrays terminated with END_OF_VEC_VEC per the above, and the whole
+// shebang is terminated with END_OF_VEC_VEC_VEC.
+//
+// Returns the vector construction.
+extern std::vector<std::vector<std::vector<int>>> generate_indices_set(int* inits);
+
+// These need to be distinct from any values that can appear, which means
+// they should be negative, and not -1, which is used as a "N/A" value.
+#define END_OF_VEC_VEC -100
+#define END_OF_VEC_VEC_VEC -200
+
 // An abstract helper class used to access elements of an initialization vector.
 // We need the abstraction because InitsManager below needs to be able to refer
 // to any of a range of templated classes.
@@ -29,7 +51,12 @@ public:
 };
 
 // Convenient way to refer to an offset associated with a particular Zeek type.
-using CPP_ValElem = std::pair<TypeTag, int>;
+// A "struct" rather than a std::pair because C++ compilers are terribly slow
+// at initializing large numbers of the latter.
+struct CPP_ValElem {
+    TypeTag tag;
+    int offset;
+};
 
 // This class groups together all of the vectors needed for run-time
 // initialization.  We gather them together into a single object so as
@@ -57,7 +84,7 @@ public:
     // index.
     ValPtr ConstVals(int offset) const {
         auto& cv = const_vals[offset];
-        return Consts(cv.first, cv.second);
+        return Consts(cv.tag, cv.offset);
     }
 
     // Retrieves the Zeek constant value for a particular Zeek type.
@@ -157,9 +184,6 @@ protected:
     // Pre-initialize all elements requiring it.
     virtual void DoPreInits(InitsManager* im, const std::vector<int>& offsets_vec) {}
 
-    // Generate a single element.
-    virtual void GenerateElement(InitsManager* im, T2& init, int offset) {}
-
     // The initialization vector in its entirety.
     std::vector<T1>& inits_vec;
 
@@ -221,16 +245,16 @@ using ValElemVecVec = std::vector<ValElemVec>;
 template<class T>
 class CPP_IndexedInits : public CPP_AbstractInits<T, ValElemVecVec> {
 public:
-    CPP_IndexedInits(std::vector<T>& _inits_vec, int _offsets_set, std::vector<ValElemVecVec> _inits)
-        : CPP_AbstractInits<T, ValElemVecVec>(_inits_vec, _offsets_set, std::move(_inits)) {}
+    CPP_IndexedInits(std::vector<T>& _inits_vec, int _offsets_set, int* raw_inits)
+        : CPP_AbstractInits<T, ValElemVecVec>(_inits_vec, _offsets_set, generate_indices_set(raw_inits)) {}
 
 protected:
     void InitializeCohortWithOffsets(InitsManager* im, int cohort, const std::vector<int>& cohort_offsets) override;
 
-    // Note, in the following we pass in the inits_vec, even though
-    // the method will have direct access to it, because we want to
-    // use overloading to dispatch to custom generation for different
-    // types of values.
+    // Note, in the following we pass in the inits_vec ("ivec"), even though
+    // the method will have direct access to it, because we want to use
+    // overloading to dispatch to custom generation for different types of
+    // values.
     void Generate(InitsManager* im, std::vector<EnumValPtr>& ivec, int offset, ValElemVec& init_vals);
     void Generate(InitsManager* im, std::vector<StringValPtr>& ivec, int offset, ValElemVec& init_vals);
     void Generate(InitsManager* im, std::vector<PatternValPtr>& ivec, int offset, ValElemVec& init_vals);
@@ -254,8 +278,8 @@ protected:
 // on subclasses of TypePtr.
 class CPP_TypeInits : public CPP_IndexedInits<TypePtr> {
 public:
-    CPP_TypeInits(std::vector<TypePtr>& _inits_vec, int _offsets_set, std::vector<std::vector<ValElemVec>> _inits)
-        : CPP_IndexedInits<TypePtr>(_inits_vec, _offsets_set, _inits) {}
+    CPP_TypeInits(std::vector<TypePtr>& _inits_vec, int _offsets_set, int* raw_inits)
+        : CPP_IndexedInits<TypePtr>(_inits_vec, _offsets_set, raw_inits) {}
 
 protected:
     void DoPreInits(InitsManager* im, const std::vector<int>& offsets_vec) override;
@@ -504,11 +528,4 @@ struct CPP_RegisterBody {
     std::vector<std::string> events;
 };
 
-// Helper function that takes a (large) array of int's and from them
-// constructs the corresponding vector-of-vector-of-indices.  Each
-// vector-of-indices is represented first by an int specifying its
-// size, and then that many int's for its values.  We recognize the
-// end of the array upon encountering a "size" entry of -1.
-extern void generate_indices_set(int* inits, std::vector<std::vector<int>>& indices_set);
-
 } // namespace zeek::detail