Defer graph compilation in atomic bundles

This avoids situations like compiling a graph hundreds of times when it
is loaded because it has hundreds of nodes and each event triggers a
re-compile.

This speeds things up dramatically, but exacerbates the theoretical
problem of there not being enough time in a cycle to execute a bundle.
As far as I can tell, the execute phase of events is very fast, so
hundreds or thousands can easily run in a tiny fraction of the process
cycle, but this still needs resolution to be truly hard real-time.  What
probably needs to happen is that all context and state used to process
is moved to CompiledGraph and nodes do not access their own fields at
all, but have some references into the CompiledGraph.  This way, a
compiled graph is separate from its "source code", and an old one could
continue to be run while a new one is beng applied across several
cycles.

1 files changed, 25 insertions, 2 deletions

diff --git a/src/server/events/Mark.cpp b/src/server/events/Mark.cpp
index 11690487..c72cc14f 100644
--- a/src/server/events/Mark.cpp
+++ b/src/server/events/Mark.cpp
@@ -15,6 +15,7 @@
 */
 
 #include "Engine.hpp"
+#include "PreProcessContext.hpp"
 #include "UndoStack.hpp"
 #include "events/Mark.hpp"
 
@@ -32,8 +33,15 @@ Mark::Mark(Engine&         engine,
 	, _depth(0)
 {}
 
+Mark::~Mark()
+{
+	for (const auto& g : _compiled_graphs) {
+		delete g.second;
+	}
+}
+
 bool
-Mark::pre_process()
+Mark::pre_process(PreProcessContext& ctx)
 {
 	UndoStack* const stack = ((_mode == Mode::UNDO)
 	                          ? _engine.redo_stack()
@@ -41,10 +49,21 @@ Mark::pre_process()
 
 	switch (_type) {
 	case Type::BUNDLE_START:
+		ctx.set_in_bundle(true);
 		_depth = stack->start_entry();
 		break;
 	case Type::BUNDLE_END:
 		_depth = stack->finish_entry();
+		ctx.set_in_bundle(false);
+		if (!ctx.dirty_graphs().empty()) {
+			for (GraphImpl* g : ctx.dirty_graphs()) {
+				CompiledGraph* cg = CompiledGraph::compile(g);
+				if (cg) {
+					_compiled_graphs.insert(std::make_pair(g, cg));
+				}
+			}
+			ctx.dirty_graphs().clear();
+		}
 		break;
 	}
 
@@ -53,7 +72,11 @@ Mark::pre_process()
 
 void
 Mark::execute(RunContext& context)
-{}
+{
+	for (auto& g : _compiled_graphs) {
+		g.second = g.first->swap_compiled_graph(g.second);
+	}
+}
 
 void
 Mark::post_process()