Skip checkDebug()'s Python round-trip entirely in fast-continue mode

bindings/module.cpp

@@ -27,6 +27,7 @@
 
 #include "openscad_cpp_parser/api.hpp"
 
+#include <atomic>
 #include <cstdint>
 #include <exception>
 #include <memory>
@@ -362,17 +363,20 @@ using GetChildrenPositionsFn = std::function<const std::optional<std::vector<std
 // exactly where it already computes should_pause, without needing to
 // widen hook()'s existing (cmd, mods) return-tuple contract -- mirrors
 // generate_partial/get_children_positions's own "extra kwarg the Python
-// side may or may not call" shape.
-using SetFastContinueFn = std::function<void(std::optional<std::unordered_map<std::string, std::set<int>>>)>;
+// side may or may not call" shape. `hookSkippable` mirrors
+// setFastContinueBreakpoints's own second parameter exactly -- see its
+// doc comment for why this is narrower than "breakpoints is accurate."
+using SetFastContinueFn = std::function<void(std::optional<std::unordered_map<std::string, std::set<int>>>, bool)>;
 
 // Wraps a SetFastContinueFn as a Python callable taking either a
 // dict[str, set[int]]/dict[str, list[int]] (origin -> breakpoint lines) or
-// None. Mirrors generatePartialTrampoline's own shape.
+// None, plus a hook_skippable bool (default False). Mirrors
+// generatePartialTrampoline's own shape.
 nb::object setFastContinueTrampoline(const SetFastContinueFn& setFastContinue) {
     return nb::cpp_function(
-        [&setFastContinue](nb::object breakpoints) {
+        [&setFastContinue](nb::object breakpoints, bool hookSkippable) {
             if (breakpoints.is_none()) {
-                setFastContinue(std::nullopt);
+                setFastContinue(std::nullopt, hookSkippable);
                 return;
             }
             std::unordered_map<std::string, std::set<int>> bp;
@@ -381,7 +385,7 @@ nb::object setFastContinueTrampoline(const SetFastContinueFn& setFastContinue) {
                 for (nb::handle line : v) lines.insert(nb::cast<int>(line));
                 bp.emplace(nb::cast<std::string>(k), std::move(lines));
             }
-            setFastContinue(std::move(bp));
+            setFastContinue(std::move(bp), hookSkippable);
         },
         // Without .none(), nanobind rejects a Python `None` argument for an
         // `nb::object`-typed parameter registered this ad-hoc way (unlike a
@@ -391,9 +395,33 @@ nb::object setFastContinueTrampoline(const SetFastContinueFn& setFastContinue) {
         // was called with breakpoints=None (DebugSession's own
         // _apply_fast_continue calls it with None whenever fast-continue
         // mode isn't safe right now -- see debugger.py).
-        nb::arg("breakpoints").none());
+        nb::arg("breakpoints").none(), nb::arg("hook_skippable") = false);
 }
 
+// Python-visible handle onto a lock-free, GIL-free interrupt flag -- see
+// Evaluator::setFastContinueInterruptFlag's own doc comment (evaluator.hpp)
+// for why this exists at all: debug_evaluate() runs as one single blocking
+// call with the GIL released for its whole duration, so there is no live,
+// Python-callable Evaluator handle DebugSession.pause()/set_breakpoints()
+// (running on the MAIN/GUI thread) could otherwise invoke directly to
+// interrupt hook-skippable mode. A caller creates ONE of these before
+// calling debug_evaluate() (passing it as fast_continue_signal), keeps it
+// around for the whole debug session, and calls .request() from the main
+// thread any time a hook-skippable checkDebug() call needs to stop
+// skipping and actually consult Python again -- Pause, or a breakpoint
+// being toggled in an editor tab while a render is mid-flight. Nothing
+// else needs to read it back on the Python side; C++ test-and-clears it
+// (see checkDebug's own doc comment, debug_profile.cpp) the moment it acts
+// on it, so there's no separate "acknowledge" step.
+class FastContinueSignal {
+public:
+    void request() { flag_->store(true, std::memory_order_release); }
+    const std::shared_ptr<std::atomic<bool>>& flag() const { return flag_; }
+
+private:
+    std::shared_ptr<std::atomic<bool>> flag_ = std::make_shared<std::atomic<bool>>(false);
+};
+
 // GIL held (reacquired by the caller). `getFrame`/`callStack`/`generatePartial`/
 // `getChildrenPositions` are valid only for this synchronous call, so their
 // closures are only ever invoked from within the Python hook, before it returns.
@@ -470,7 +498,8 @@ void callPyReturnHook(nb::handle returnHook, const std::string& name, const osca
 // blocks on the Python side (threading.Event) so the GUI thread keeps running.
 nb::object debugEvaluate(const std::string& path, nb::dict viewportParams, nb::callable debugHook,
                           nb::callable errorBreak, nb::callable echoFn,
-                          std::shared_ptr<oscadeval::ManifoldCache> manifoldCache, nb::object returnHook) {
+                          std::shared_ptr<oscadeval::ManifoldCache> manifoldCache, nb::object returnHook,
+                          FastContinueSignal* fastContinueSignal) {
     std::unordered_map<std::string, oscadeval::Value> vp = toViewportParams(viewportParams);
 
     std::vector<oscadeval::ColoredBody> bodies;
@@ -501,8 +530,8 @@ nb::object debugEvaluate(const std::string& path, nb::dict viewportParams, nb::c
             return evPtr->lastChildrenPositions();
         };
         SetFastContinueFn setFastContinue =
-            [&evPtr](std::optional<std::unordered_map<std::string, std::set<int>>> breakpoints) {
-            evPtr->setFastContinueBreakpoints(std::move(breakpoints));
+            [&evPtr](std::optional<std::unordered_map<std::string, std::set<int>>> breakpoints, bool hookSkippable) {
+            evPtr->setFastContinueBreakpoints(std::move(breakpoints), hookSkippable);
         };
         oscadeval::DebugHooks hooks;
         hooks.debugHook = [&](int line, int depth, bool forced, bool exprLevel, const std::string& origin,
@@ -528,6 +557,7 @@ nb::object debugEvaluate(const std::string& path, nb::dict viewportParams, nb::c
             oscadeval::ResolvedUseScopes used = oscadeval::resolveUseScopes(ast, path, echoCpp);
             oscadeval::Evaluator ev(echoCpp, nullptr, manifoldCache, hooks, false);
             evPtr = &ev;
+            if (fastContinueSignal) ev.setFastContinueInterruptFlag(fastContinueSignal->flag());
             oscadeval::EvalContext ctx = oscadeval::EvalContext::makeRoot(used.rootScope.get());
             bodies = oscadeval::toRenderableBodies(ev.evaluate(used.processedNodes, ctx, vp));
             collectIdSpans(ev, idSpans);
@@ -602,13 +632,19 @@ NB_MODULE(_openscad_cpp_evaluator, m) {
         .def(nb::init<>())
         .def("clear", &oscadeval::ManifoldCache::clear);
 
+    // See FastContinueSignal's own doc comment, above -- only a
+    // constructor and request() are ever called from Python.
+    nb::class_<FastContinueSignal>(m, "FastContinueSignal")
+        .def(nb::init<>())
+        .def("request", &FastContinueSignal::request);
+
     m.def("evaluate", &evaluate, nb::arg("path"), nb::arg("viewport_params"), nb::arg("manifold_cache") = nullptr,
           nb::arg("profile") = false,
           "Evaluate a .scad file; return (bodies, echoes, id_to_node, csg_tree, profile_result, dyn, dyn_explicit).");
     m.def("parse_decls", &parseDecls, nb::arg("path"),
           "Parse a .scad file; return top-level declaration (namespace, name, start, end, line, column, origin) tuples.");
     m.def("debug_evaluate", &debugEvaluate, nb::arg("path"), nb::arg("viewport_params"), nb::arg("debug_hook"),
           nb::arg("error_break"), nb::arg("echo_fn"), nb::arg("manifold_cache") = nullptr,
-          nb::arg("return_hook") = nb::none(),
+          nb::arg("return_hook") = nb::none(), nb::arg("fast_continue_signal") = nullptr,
           "Evaluate with the debugger wired in; returns (bodies, [], id_to_node, dyn, dyn_explicit). Callbacks fire under the GIL.");
 }

include/openscad_cpp_evaluator/evaluator.hpp

@@ -15,6 +15,7 @@
 
 #include "openscad_cpp_parser/ast.hpp"
 
+#include <atomic>
 #include <chrono>
 #include <cstdint>
 #include <functional>
@@ -329,12 +330,61 @@ class Evaluator {
     // exception at all" (today's original, always-safe behavior) --
     // e.g. the instant a step command starts or a pause is requested.
     // Cheap to call as often as the debugger's own state changes: this
-    // just updates one member, no cache invalidation needed (chunkEligibleNow
-    // is re-checked on every lookupOrCompileChunk/lookupCompiledLiteralChunk
-    // call, not cached itself -- only the compiled bytecode ITSELF, which
-    // never depends on debugger state, is cached forever).
-    void setFastContinueBreakpoints(std::optional<std::unordered_map<std::string, std::set<int>>> breakpoints) {
+    // just updates two members, no cache invalidation needed (chunkEligibleNow/
+    // checkDebug are re-checked on every call, not cached themselves -- only
+    // the compiled bytecode ITSELF, which never depends on debugger state,
+    // is cached forever).
+    //
+    // `hookSkippable`: a STRICTLY narrower claim than "breakpoints is
+    // accurate" above -- true only when NOTHING needs to inspect a
+    // statement-level checkpoint's own line/depth at all, not just "nothing
+    // needs to inspect it unless it's a compiled function's entry." This is
+    // false for step_over/step_out even though THEY also pass a real
+    // breakpoints set (chunkEligibleNow/useBytecodeVm still apply for them):
+    // both need checkDebug() to keep calling into the debug hook on every
+    // statement so the caller's own step_hit logic (line/depth comparison
+    // against the step's own starting point) can run -- there is no way to
+    // decide that in advance the way a breakpoint LOCATION can be. Only a
+    // plain "Continue" with no step pending can safely skip the call
+    // entirely for a line with no breakpoint -- see checkDebug's own doc
+    // comment (debug_profile.cpp) for where this is actually consulted.
+    void setFastContinueBreakpoints(std::optional<std::unordered_map<std::string, std::set<int>>> breakpoints,
+                                     bool hookSkippable = false) {
         fastContinueBreakpoints_ = std::move(breakpoints);
+        fastContinueHookSkippable_ = hookSkippable;
+    }
+
+    // The other half of hook-skippable mode's safety net. checkDebug()'s
+    // whole premise (see its own doc comment, debug_profile.cpp) is that it
+    // can skip calling into Python for a line with no breakpoint -- but the
+    // caller that decided that (DebugSession, on the MAIN/GUI thread) needs
+    // a way to say "actually, don't skip the very next one" from OUTSIDE any
+    // hook call, since debug_evaluate() runs as one single blocking call
+    // with the GIL released for its whole duration: there is no live,
+    // Python-callable Evaluator handle to invoke setFastContinueBreakpoints
+    // on directly the way a synchronous API would allow. A user clicking
+    // Pause, or toggling a breakpoint in an editor tab, while a hook-
+    // skippable render is mid-flight needs to take effect on the very next
+    // checkpoint, not "whenever a breakpoint happens to be hit next" (which,
+    // in hook-skippable mode, could be never, for a script with none set).
+    //
+    // `flag` is a plain shared_ptr<atomic<bool>> -- lock-free and GIL-free
+    // by construction, so the caller (bindings/module.cpp wraps it in a
+    // small Python-visible class) can set it from the main thread at any
+    // moment with no synchronization needed beyond the atomic itself.
+    // checkDebug() atomically test-and-clears it (exchange) on every call
+    // that would otherwise skip: if it was set, this call falls through and
+    // actually invokes the Python hook instead, which re-derives and pushes
+    // fresh breakpoints/hookSkippable state via its own existing logic --
+    // clearing it here (not from Python) means there's no separate
+    // "acknowledge" round-trip needed. Never itself compared against
+    // anything else; nullptr (the default, and what a plain debugger
+    // attach that never wires this up leaves it at) simply means hook-
+    // skippable mode -- if ever engaged at all -- can't be interrupted this
+    // way, which is only actually reachable if a caller opts into
+    // hookSkippable=true above without also providing this.
+    void setFastContinueInterruptFlag(std::shared_ptr<std::atomic<bool>> flag) {
+        fastContinueInterrupt_ = std::move(flag);
     }
 
     // Checks whether a debug pause should happen at `node` (via the
@@ -952,6 +1002,16 @@ class Evaluator {
     // setter leaves it at) means "no fast-continue exception" -- exactly
     // today's original, always-safe behavior.
     std::optional<std::unordered_map<std::string, std::set<int>>> fastContinueBreakpoints_;
+    // See setFastContinueBreakpoints's own doc comment for why this is a
+    // separate, narrower flag from fastContinueBreakpoints_ itself -- false
+    // (the default) whenever a debugger is attached without explicitly
+    // opting in, matching today's always-safe "checkDebug always calls the
+    // hook" behavior.
+    bool fastContinueHookSkippable_ = false;
+    // See setFastContinueInterruptFlag's own doc comment. nullptr (the
+    // default) means hook-skippable mode, if ever engaged, can't be
+    // interrupted from outside a hook call.
+    std::shared_ptr<std::atomic<bool>> fastContinueInterrupt_;
     // The root EvalContext resolveTree() was called with -- lets
     // checkDebug()'s locals snapshot fall back to top-level script
     // variables when paused inside a nested user call, the same way the

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "scikit_build_core.build"
 
 [project]
 name = "openscad_cpp_evaluator"
-version = "0.6.1"
+version = "0.7.0"
 description = "C++ OpenSCAD evaluator with Python bindings"
 readme = "README.md"
 requires-python = ">=3.12"

python/openscad_cpp_evaluator/__init__.py

@@ -13,11 +13,12 @@
 from typing import Optional
 
 from . import _openscad_cpp_evaluator as _ext
-from ._openscad_cpp_evaluator import ManifoldCache
+from ._openscad_cpp_evaluator import FastContinueSignal, ManifoldCache
 
 __all__ = [
     "Evaluator", "ColoredBody", "EvalError", "ParseError", "OscObject", "parse", "to_renderable_bodies",
     "ManifoldCache", "CallSiteProfile", "ProfileResult", "format_csg_tree", "bodies_from_dicts",
+    "FastContinueSignal",
 ]
 
 
@@ -317,6 +318,16 @@ class Evaluator:
     `return_hook`: fires after a user function/function-literal call
     computes its result, before returning -- (name, value, depth). Only
     meaningful with debug_hook set (the plain evaluate() path never calls it).
+    `fast_continue_signal`: an optional FastContinueSignal the caller keeps
+    across the whole debug session and calls `.request()` on (e.g. from
+    DebugSession.pause()/set_breakpoints(), BelfrySCAD's own debugger.py)
+    any time a checkDebug() checkpoint skipped via hook-skippable fast-
+    continue mode (see debug_hook's own `set_fast_continue(breakpoints,
+    hook_skippable)` kwarg) needs to stop skipping and consult Python
+    again -- there is no other way to reach a running debug_evaluate() call
+    from outside a hook invocation, since it runs as one single blocking
+    call with the GIL released for its duration. Only meaningful with
+    debug_hook set; ignored otherwise.
 
     After evaluate() returns, `self.csg_tree` (list of _CSGNode, see
     format_csg_tree), `self.dyn` (dict of every currently-visible
@@ -328,13 +339,14 @@ class Evaluator:
     """
 
     def __init__(self, echo_fn=None, debug_hook=None, error_break_fn=None, return_hook=None,
-                 manifold_cache=None, profile=False):
+                 manifold_cache=None, profile=False, fast_continue_signal=None):
         self._echo_fn = echo_fn
         self._debug_hook = debug_hook
         self._error_break_fn = error_break_fn
         self._return_hook = return_hook
         self._manifold_cache = manifold_cache
         self._profile = profile
+        self._fast_continue_signal = fast_continue_signal
         self.csg_tree = []
         self.profile_result = None
         self.dyn = {}
@@ -350,7 +362,7 @@ def evaluate(self, source_path: str, viewport_params: Optional[dict] = None):
                     source_path, vp, self._debug_hook,
                     self._error_break_fn or (lambda *a, **k: None),
                     self._echo_fn or (lambda _m: None),
-                    self._manifold_cache, self._return_hook)
+                    self._manifold_cache, self._return_hook, self._fast_continue_signal)
                 self.csg_tree = []
                 self.profile_result = None
             else:

src/debug_profile.cpp

@@ -77,6 +77,36 @@ std::vector<DebugFrame> Evaluator::buildDebugFrames(const EvalContext* ctx) cons
 void Evaluator::checkDebug(const oscad::ASTNode& node, EvalContext& ctx, bool forced, bool exprLevel) {
     if (!debugHooks_.debugHook) return;
     const oscad::Position& pos = node.position();
+    // Fast-continue's hook-skippable mode (setFastContinueBreakpoints' own
+    // doc comment): a plain "Continue" with no step pending needs the debug
+    // hook called ONLY for a line that actually has a breakpoint -- every
+    // other checkpoint is guaranteed to do nothing (no forced/breakpoint/
+    // step_hit condition on the Python side can possibly fire), so skip the
+    // call (and the childStatementPositions/getFrame setup below, all
+    // wasted work otherwise) entirely rather than crossing into Python just
+    // to be told "continue". `forced` (the explicit breakpoint() builtin)
+    // always bypasses this, matching its own "bypasses nothing" contract.
+    // Never applies to step_over/step_out (hookSkippable is false for those
+    // even though they also set a real breakpoints set, see the setter's
+    // own doc comment) or step_into/step_to_child (fastContinueBreakpoints_
+    // itself is nullopt then) -- both need every statement inspected.
+    if (!forced && fastContinueHookSkippable_ && fastContinueBreakpoints_) {
+        // Test-and-clear: if the main thread requested an interrupt (Pause,
+        // or a breakpoint edit -- see setFastContinueInterruptFlag's own doc
+        // comment) since the last checkDebug() call, this call falls
+        // through and actually invokes the hook below instead of skipping,
+        // regardless of whether THIS specific line has a breakpoint --
+        // that's what lets the hook's own logic (pause_now, or a freshly
+        // updated breakpoints dict) run at all in hook-skippable mode.
+        const bool interrupted =
+            fastContinueInterrupt_ && fastContinueInterrupt_->exchange(false, std::memory_order_acq_rel);
+        if (!interrupted) {
+            auto originIt = fastContinueBreakpoints_->find(pos.origin);
+            if (originIt == fastContinueBreakpoints_->end() || !originIt->second.count(pos.line)) {
+                return;
+            }
+        }
+    }
     const int depth = static_cast<int>(callStack_.size());
     const DebugFramesFn getFrame = [this, &ctx]() { return buildDebugFrames(&ctx); };
     lastChildrenPositions_ = childStatementPositions(node);