Include what the new code uses

The Windows build stopped on test_kdtree.cpp: it calls std::iota without
including <numeric>, which libstdc++ happens to pull in anyway. Added
there, and the same for <limits> and <algorithm>/<cmath> where the
recent changes rely on them being included by something else.

src/libslic3r/ClipperUtils.cpp

@@ -1,7 +1,11 @@
+#include <algorithm>
+#include <cmath>
 #include <limits>
 #include <numeric>
 #include <unordered_map>
 
+#include <tbb/parallel_for.h>
+
 #include "ClipperUtils.hpp"
 #include "Geometry.hpp"
 #include "ShortestPath.hpp"
@@ -813,6 +817,69 @@ Slic3r::ExPolygons intersection_ex(const Slic3r::Surfaces &subject, const Slic3r
     { return _clipper_ex(ClipperLib::ctIntersection, ClipperUtils::SurfacesProvider(subject), ClipperUtils::SurfacesProvider(clip), do_safety_offset); }
 Slic3r::ExPolygons intersection_ex(const Slic3r::SurfacesPtr &subject, const Slic3r::ExPolygons &clip, ApplySafetyOffset do_safety_offset)
     { return _clipper_ex(ClipperLib::ctIntersection, ClipperUtils::SurfacesPtrProvider(subject), ClipperUtils::ExPolygonsProvider(clip), do_safety_offset); }
+
+namespace ClipperUtils {
+    std::vector<ExPolygonsTile> tile_expolygons(const ExPolygons &expolygons, size_t per_tile)
+    {
+        BoundingBox              extent;
+        std::vector<BoundingBox> bboxes;
+        bboxes.reserve(expolygons.size());
+        for (const ExPolygon &expoly : expolygons) {
+            bboxes.emplace_back(get_extents(expoly));
+            extent.merge(bboxes.back());
+        }
+        if (! extent.defined)
+            return {};
+        const int     tiles  = std::clamp(int(std::sqrt(double(expolygons.size()) / double(std::max<size_t>(per_tile, 1)))), 1, 32);
+        const Point   size   = extent.size();
+        const coord_t tile_w = std::max<coord_t>(1, size.x() / tiles + 1), tile_h = std::max<coord_t>(1, size.y() / tiles + 1);
+        std::vector<ExPolygonsTile> out(size_t(tiles * tiles));
+        for (size_t i = 0; i < expolygons.size(); ++ i) {
+            const Point c = bboxes[i].center();
+            ExPolygonsTile &tile = out[size_t(std::clamp(int((c.y() - extent.min.y()) / tile_h), 0, tiles - 1) * tiles +
+                                              std::clamp(int((c.x() - extent.min.x()) / tile_w), 0, tiles - 1))];
+            tile.members.emplace_back(i);
+            tile.bbox.merge(bboxes[i]);
+        }
+        out.erase(std::remove_if(out.begin(), out.end(), [](const ExPolygonsTile &tile) { return tile.members.empty(); }), out.end());
+        return out;
+    }
+}
+
+static Slic3r::ExPolygons clipper_ex_by_piece(ClipperLib::ClipType clipType, const Slic3r::ExPolygons &subject, const Slic3r::Polygons &clip, ApplySafetyOffset do_safety_offset)
+{
+    // A few dozen subject ExPolygons to a tile, each tile one ClipperLib call with the clip cut to the tile's box.
+    const std::vector<ClipperUtils::ExPolygonsTile> tiles = ClipperUtils::tile_expolygons(subject, 32);
+    std::vector<BoundingBox> clip_bboxes;
+    clip_bboxes.reserve(clip.size());
+    for (const Polygon &polygon : clip)
+        clip_bboxes.emplace_back(get_extents(polygon));
+
+    std::vector<Slic3r::ExPolygons> out_tiles(tiles.size());
+    tbb::parallel_for(size_t(0), tiles.size(), [&](size_t tile_idx) {
+        const ClipperUtils::ExPolygonsTile &tile = tiles[tile_idx];
+        Slic3r::ExPolygons local_subject;
+        local_subject.reserve(tile.members.size());
+        for (size_t i : tile.members)
+            local_subject.emplace_back(subject[i]);
+        // Grown so that the cut edges of the clip stay clear of the subject, also after the safety offset.
+        const BoundingBox bbox = tile.bbox.inflated(SCALED_EPSILON);
+        Polygons local_clip;
+        for (size_t i = 0; i < clip.size(); ++i)
+            if (clip_bboxes[i].overlap(bbox))
+                if (Polygon clipped = ClipperUtils::clip_clipper_polygon_with_subject_bbox(clip[i], bbox); ! clipped.empty())
+                    local_clip.emplace_back(std::move(clipped));
+        out_tiles[tile_idx] = _clipper_ex(clipType, ClipperUtils::ExPolygonsProvider(local_subject), ClipperUtils::PolygonsProvider(local_clip), do_safety_offset);
+    });
+    Slic3r::ExPolygons out;
+    for (Slic3r::ExPolygons &out_tile : out_tiles)
+        append(out, std::move(out_tile));
+    return out;
+}
+Slic3r::ExPolygons diff_ex_by_piece(const Slic3r::ExPolygons &subject, const Slic3r::Polygons &clip, ApplySafetyOffset do_safety_offset)
+    { return clipper_ex_by_piece(ClipperLib::ctDifference, subject, clip, do_safety_offset); }
+Slic3r::ExPolygons intersection_ex_by_piece(const Slic3r::ExPolygons &subject, const Slic3r::Polygons &clip, ApplySafetyOffset do_safety_offset)
+    { return clipper_ex_by_piece(ClipperLib::ctIntersection, subject, clip, do_safety_offset); }
 // May be used to "heal" unusual models (3DLabPrints etc.) by providing fill_type (pftEvenOdd, pftNonZero, pftPositive, pftNegative).
 Slic3r::ExPolygons union_ex(const Slic3r::Polygons &subject, ClipperLib::PolyFillType fill_type)
     { return _clipper_ex(ClipperLib::ctUnion, ClipperUtils::PolygonsProvider(subject), ClipperUtils::EmptyPathsProvider(), ApplySafetyOffset::No, fill_type); }

src/libslic3r/ClipperUtils.hpp

@@ -2,6 +2,7 @@
 #define slic3r_ClipperUtils_hpp_
 
 #include "libslic3r.h"
+#include "BoundingBox.hpp"
 #include "clipper.hpp"
 #include "ExPolygon.hpp"
 #include "Polygon.hpp"
@@ -321,6 +322,15 @@ namespace ClipperUtils {
     [[nodiscard]] Polygons clip_clipper_polygons_with_subject_bbox(const ExPolygon &src, const BoundingBox &bbox, const bool get_entire_polygons = false);
     [[nodiscard]] Polygons clip_clipper_polygons_with_subject_bbox(const ExPolygons &src, const BoundingBox &bbox, const bool get_entire_polygons = false);
 
+    // Splits ExPolygons into tiles by the centres of their boxes, about `per_tile` of them to a tile, to run ClipperLib on a
+    // layer of many pieces tile by tile. Returns the non-empty tiles, each with the indices of its ExPolygons and their box.
+    struct ExPolygonsTile
+    {
+        BoundingBox         bbox;
+        std::vector<size_t> members;
+    };
+    [[nodiscard]] std::vector<ExPolygonsTile> tile_expolygons(const ExPolygons &expolygons, size_t per_tile);
+
     }
 
 // Perform union of input polygons using the non-zero rule, convert to ExPolygons.
@@ -518,6 +528,11 @@ Slic3r::ExPolygons intersection_ex(const Slic3r::Surfaces &subject, const Slic3r
 Slic3r::ExPolygons intersection_ex(const Slic3r::Surfaces &subject, const Slic3r::ExPolygons &clip, ApplySafetyOffset do_safety_offset = ApplySafetyOffset::No);
 Slic3r::ExPolygons intersection_ex(const Slic3r::Surfaces &subject, const Slic3r::Surfaces &clip, ApplySafetyOffset do_safety_offset = ApplySafetyOffset::No);
 Slic3r::ExPolygons intersection_ex(const Slic3r::SurfacesPtr &subject, const Slic3r::ExPolygons &clip, ApplySafetyOffset do_safety_offset = ApplySafetyOffset::No);
+// diff_ex() / intersection_ex() of the subject split into tiles, each against only the part of the clip near it, the tiles in
+// parallel. The same area as the operation on the whole subject when its ExPolygons do not overlap, and much faster for a
+// subject of thousands of pieces spread over a layer: ClipperLib slows down with the number of edges crossing a scan line.
+Slic3r::ExPolygons diff_ex_by_piece(const Slic3r::ExPolygons &subject, const Slic3r::Polygons &clip, ApplySafetyOffset do_safety_offset = ApplySafetyOffset::No);
+Slic3r::ExPolygons intersection_ex_by_piece(const Slic3r::ExPolygons &subject, const Slic3r::Polygons &clip, ApplySafetyOffset do_safety_offset = ApplySafetyOffset::No);
 Slic3r::Polylines  intersection_pl(const Slic3r::Polylines &subject, const Slic3r::Polygon &clip);
 Slic3r::Polylines  intersection_pl(const Slic3r::Polyline &subject, const Slic3r::ExPolygon &clip);
 Slic3r::Polylines  intersection_pl(const Slic3r::Polylines &subject, const Slic3r::ExPolygon &clip);

src/libslic3r/Fill/Fill.cpp

@@ -9,6 +9,8 @@
 #include "../PrintConfig.hpp"
 #include "../Surface.hpp"
 
+#include <tbb/parallel_for.h>
+
 #include "AABBTreeLines.hpp"
 #include "ExtrusionEntity.hpp"
 #include "Fill.hpp"
@@ -630,24 +632,28 @@ void split_solid_surface(size_t layer_id, const SurfaceFill &fill, ExPolygons &n
     if (!line_based_pattern) {
         const coord_t scaled_spacing = scaled<coord_t>(fill.params.spacing);
 
-        for (const ExPolygon &expolygon : fill.expolygons) {
+        // Each expolygon is split on its own, so they run in parallel and are collected in their original order.
+        std::vector<std::pair<ExPolygons, ExPolygons>> split_parts(fill.expolygons.size()); // normal, narrow
+        tbb::parallel_for(size_t(0), fill.expolygons.size(), [&](size_t idx) {
+            const ExPolygon &expolygon = fill.expolygons[idx];
             Polygons filled_area = to_polygons(expolygon);
 
             // "Core" area: open (erode+dilate) to drop thin features, then clamp back to the original polygon.
             Polygons inner_area  = intersection(filled_area, opening(filled_area, scaled_spacing, scaled_spacing));
 
             if (inner_area.empty()) {
-                narrow_infill.emplace_back(expolygon);
-                continue;
+                split_parts[idx].second.emplace_back(expolygon);
+                return;
             }
 
             ExPolygons inner_ex = union_ex(inner_area);
             ExPolygons expolys{expolygon};
-            ExPolygons narrow_ex = diff_ex(expolys, inner_ex);
-            ExPolygons normal_ex = intersection_ex(expolys, inner_ex);
-
-            append(normal_infill, normal_ex); // normal infill area
-            append(narrow_infill, narrow_ex); // narrow infill area
+            split_parts[idx].second = diff_ex(expolys, inner_ex);         // narrow infill area
+            split_parts[idx].first  = intersection_ex(expolys, inner_ex); // normal infill area
+        });
+        for (auto &[normal_ex, narrow_ex] : split_parts) {
+            append(normal_infill, std::move(normal_ex));
+            append(narrow_infill, std::move(narrow_ex));
         }
 
         return;
@@ -669,7 +675,10 @@ void split_solid_surface(size_t layer_id, const SurfaceFill &fill, ExPolygons &n
     }
     const double aligning_angle = -base_angle + PI;
 
-	for (const ExPolygon &expolygon : fill.expolygons) {
+    // Each expolygon is reconstructed on its own, so they run in parallel and are collected in their original order.
+    std::vector<Polygons> split_reconstructed(fill.expolygons.size());
+    tbb::parallel_for(size_t(0), fill.expolygons.size(), [&](size_t expolygon_idx) {
+        const ExPolygon &expolygon = fill.expolygons[expolygon_idx];
         Polygons filled_area = to_polygons(expolygon);
         polygons_rotate(filled_area, aligning_angle);
         BoundingBox bb = get_extents(filled_area);
@@ -800,8 +809,10 @@ void split_solid_surface(size_t layer_id, const SurfaceFill &fill, ExPolygons &n
             }
         }
 
-        polygons_append(normal_fill_areas, reconstructed_area);
-    }
+        split_reconstructed[expolygon_idx] = std::move(reconstructed_area);
+    });
+    for (Polygons &reconstructed_area : split_reconstructed)
+        polygons_append(normal_fill_areas, std::move(reconstructed_area));
 
     polygons_rotate(normal_fill_areas, -aligning_angle);
 
@@ -1409,7 +1420,15 @@ void Layer::make_fills(FillAdaptive::Octree* adaptive_fill_octree, FillAdaptive:
             // Orca: Reuse the body origin used for bridge anchoring, resetting it for each surface.
             f->set_bounding_box(infill_bounding_box(*this, surface_fill, expoly, bbox));
 
-            f->no_overlap_expolygons = intersection_ex(surface_fill.no_overlap_expolygons, ExPolygons() = {expoly}, ApplySafetyOffset::Yes);
+            // Only the part of the layer-wide no-overlap area under this expolygon matters, so clip it to the
+            // expolygon's box first (padded past the safety offset, which grows the clip side). The result is
+            // identical; the cost is not: a layer split into many small fills, e.g. by colour painting,
+            // otherwise intersects every one of them with the whole layer.
+            BoundingBox no_overlap_bbox = get_extents(expoly);
+            no_overlap_bbox.offset(SCALED_EPSILON);
+            f->no_overlap_expolygons = intersection_ex(
+                ClipperUtils::clip_clipper_polygons_with_subject_bbox(surface_fill.no_overlap_expolygons, no_overlap_bbox),
+                ExPolygons() = {expoly}, ApplySafetyOffset::Yes);
             if (params.symmetric_infill_y_axis) {
                 params.symmetric_y_axis = f->extended_object_bounding_box().center().x();
                 expoly.symmetric_y(params.symmetric_y_axis);

src/libslic3r/GCode/OrderingStrategies.cpp

@@ -10,6 +10,7 @@
 #include <limits>
 #include <numeric>
 #include <unordered_map>
+#include <unordered_set>
 #include <utility>
 #include <vector>
 
@@ -134,15 +135,79 @@ bool tsp_remove_crossings(std::vector<size_t>& path, const Points& centers)
         return {std::numeric_limits<size_t>::max(), std::numeric_limits<size_t>::max()};
     };
 
+    // For many islands, the same scan with the edges binned in a uniform grid over their boxes, so each edge is only tested against the edges sharing a
+    // cell with it - two edges whose boxes overlap always do. It returns the same crossing as the all-pairs scan
+    // (smallest i, then smallest j), so the result is unchanged; with thousands of islands on a layer the all-pairs
+    // scan, repeated after every reversal, never finished. Rebuilding the grid costs more than it saves on small inputs.
+    constexpr size_t grid_min_size = 500;
+    BoundingBox extent;
+    for (size_t idx : path)
+        extent.merge(centers[idx]);
+    const int     grid_n = std::clamp(int(std::sqrt(double(pn))), 1, 256);
+    const coord_t cell_w = std::max<coord_t>(1, (extent.max.x() - extent.min.x()) / grid_n + 1);
+    const coord_t cell_h = std::max<coord_t>(1, (extent.max.y() - extent.min.y()) / grid_n + 1);
+    const auto    for_cells = [&](const Point& a, const Point& b, auto&& fn) {
+        const int x0 = int((std::min(a.x(), b.x()) - extent.min.x()) / cell_w), x1 = int((std::max(a.x(), b.x()) - extent.min.x()) / cell_w);
+        const int y0 = int((std::min(a.y(), b.y()) - extent.min.y()) / cell_h), y1 = int((std::max(a.y(), b.y()) - extent.min.y()) / cell_h);
+        for (int y = y0; y <= y1; ++y)
+            for (int x = x0; x <= x1; ++x)
+                fn(y * grid_n + x);
+    };
+    std::vector<std::vector<size_t>> edge_cells(size_t(grid_n) * grid_n);
+
+    auto find_crossing_grid = [&]() -> std::pair<size_t, size_t> {
+        for (std::vector<size_t>& cell : edge_cells)
+            cell.clear();
+        for (size_t j = 0; j < n_edges; ++j)
+            for_cells(centers[path[j]], centers[path[(j + 1) % pn]], [&](int cell) { edge_cells[cell].emplace_back(j); });
+
+        for (size_t i = 0; i < n_edges; ++i) {
+            const Point& ai = centers[path[i]];
+            const Point& bi = centers[path[(i + 1) % pn]];
+
+            size_t first_j = std::numeric_limits<size_t>::max();
+            for_cells(ai, bi, [&](int cell) {
+                for (size_t j : edge_cells[cell]) {
+                    if (j < i + 2 || j >= first_j) continue;
+                    // Skip the (0, pn-1) pair: edges (0,1) and (pn-1,0) share node 0.
+                    if (i == 0 && j == pn - 1) continue;
+
+                    const Point& aj = centers[path[j]];
+                    const Point& bj = centers[path[(j + 1) % pn]];
+
+                    if (!bboxes_overlap(ai, bi, aj, bj)) continue;
+                    if (Geometry::segments_intersect(ai, bi, aj, bj))
+                        first_j = j;
+                }
+            });
+            if (first_j != std::numeric_limits<size_t>::max())
+                return {i, first_j};
+        }
+        return {std::numeric_limits<size_t>::max(), std::numeric_limits<size_t>::max()};
+    };
+
     // Process crossings one at a time: find first, reverse it, restart scan.
     // Cap iterations to prevent infinite loops on collinear/overlapping segments.
     int max_iters = static_cast<int>(pn * pn);
     bool improved = false;
+    // Reversing between two segments that only touch or overlap along a line need not remove the intersection, and on
+    // islands laid out on a regular grid (a tiled texture, an array of parts) the loop cycled through the same orderings
+    // until the pn * pn cap - effectively forever. Stop as soon as an ordering repeats: until then this is the same loop.
+    std::unordered_set<uint64_t> seen_paths;
+    const auto path_hash = [&path]() {
+        uint64_t h = 1469598103934665603ull; // FNV-1a
+        for (size_t idx : path)
+            h = (h ^ uint64_t(idx)) * 1099511628211ull;
+        return h;
+    };
+    seen_paths.insert(path_hash());
     while (max_iters-- > 0) {
-        auto [ci, cj] = find_crossing();
+        auto [ci, cj] = pn >= grid_min_size ? find_crossing_grid() : find_crossing();
         if (ci == std::numeric_limits<size_t>::max()) break;
         improved = true;
         std::reverse(path.begin() + ci + 1, path.begin() + cj + 1);
+        if (!seen_paths.insert(path_hash()).second)
+            break;
     }
     return improved;
 }

src/libslic3r/GCode/SeamPlacer.cpp

@@ -8,6 +8,7 @@
 #include <boost/log/trivial.hpp>
 #include <random>
 #include <algorithm>
+#include <limits>
 #include <queue>
 #include <unordered_map>
 
@@ -1178,21 +1179,21 @@ std::optional<std::pair<size_t, size_t>> SeamPlacer::find_next_seam_in_layer(
     const size_t layer_idx, const float max_distance,
     const SeamPlacerImpl::SeamComparator &comparator) const {
   using namespace SeamPlacerImpl;
-  std::vector<size_t> nearby_points_indices = find_nearby_points(*layers[layer_idx].points_tree, projected_position,
-                                                                 max_distance);
-
-  if (nearby_points_indices.empty()) {
-    return {};
+  // Find the best nearby point and the nearest one. A layer of a fine relief has tens of thousands of candidates within
+  // the radius, so they are looked at as the search finds them rather than collected into a vector first.
+  constexpr size_t none = std::numeric_limits<size_t>::max();
+  size_t best_nearby_point_index = none;
+  size_t nearest_point_index = none;
+  visit_nearby_points(*layers[layer_idx].points_tree, projected_position, max_distance,
+                      [&layers, &comparator, &projected_position, layer_idx, &best_nearby_point_index, &nearest_point_index]
+                      (size_t nearby_point_index) {
+    if (best_nearby_point_index == none) {
+      // The first point found starts both, as the first of the collected ones did.
+      best_nearby_point_index = nearest_point_index = nearby_point_index;
     }
-
-  size_t best_nearby_point_index = nearby_points_indices[0];
-  size_t nearest_point_index = nearby_points_indices[0];
-
-  // Now find best nearby point, nearest point, and corresponding indices
-  for (const size_t &nearby_point_index : nearby_points_indices) {
     const SeamCandidate &point = layers[layer_idx].points[nearby_point_index];
     if (point.perimeter.finalized) {
-      continue; // skip over finalized perimeters, try to find some that is not finalized
+      return; // skip over finalized perimeters, try to find some that is not finalized
     }
     if (comparator.is_first_better(point, layers[layer_idx].points[best_nearby_point_index],
                                    projected_position.head<2>())
@@ -1204,6 +1205,10 @@ std::optional<std::pair<size_t, size_t>> SeamPlacer::find_next_seam_in_layer(
         || layers[layer_idx].points[nearest_point_index].perimeter.finalized) {
       nearest_point_index = nearby_point_index;
     }
+  });
+
+  if (best_nearby_point_index == none) {
+    return {};
   }
 
   const SeamCandidate &best_nearby_point = layers[layer_idx].points[best_nearby_point_index];

src/libslic3r/KDTreeIndirect.hpp

@@ -313,6 +313,36 @@ std::vector<size_t> find_nearby_points(const KDTreeIndirectType &kdtree, const P
     return visitor.result;
 }
 
+// Visits the points within max_distance of center, in the order find_nearby_points() would collect them, and hands
+// each of them to `visitor_fn` instead of returning them all: a search over a dense set spends more on collecting the
+// points into a vector than on the search itself, and its caller usually keeps only a few of them.
+template<typename KDTreeIndirectType, typename PointType, typename VisitorFn>
+void visit_nearby_points(const KDTreeIndirectType &kdtree, const PointType &center,
+                         const typename KDTreeIndirectType::CoordType &max_distance, VisitorFn visitor_fn)
+{
+    using CoordType = typename KDTreeIndirectType::CoordType;
+
+    struct Visitor {
+        const KDTreeIndirectType &kdtree;
+        const PointType           center;
+        const CoordType           max_distance_squared;
+        VisitorFn                 visitor_fn;
+
+        unsigned int operator()(size_t idx, size_t dimension) {
+            auto dist = CoordType(0);
+            for (size_t i = 0; i < KDTreeIndirectType::NumDimensions; ++i) {
+                CoordType d = center[i] - kdtree.coordinate(idx, i);
+                dist += d * d;
+            }
+            if (dist < max_distance_squared)
+                visitor_fn(idx);
+            return kdtree.descent_mask(center[dimension], max_distance_squared, idx, dimension);
+        }
+    } visitor { kdtree, center, max_distance * max_distance, visitor_fn };
+
+    kdtree.visit(visitor);
+}
+
 template<typename KDTreeIndirectType, typename PointType>
 std::vector<size_t> find_nearby_points(const KDTreeIndirectType &kdtree, const PointType &center,
                                        const typename KDTreeIndirectType::CoordType& max_distance)

src/libslic3r/LayerRegion.cpp

@@ -72,10 +72,11 @@ void LayerRegion::slices_to_fill_surfaces_clipped()
         by_surface[size_t(surface.surface_type)].emplace_back(&surface);
     // Trim surfaces by the fill_boundaries.
     this->fill_surfaces.surfaces.clear();
+    const Polygons fill_boundaries = to_polygons(this->fill_expolygons);
     for (size_t surface_type = 0; surface_type < size_t(stCount); ++ surface_type) {
         const SurfacesPtr &this_surfaces = by_surface[surface_type];
         if (! this_surfaces.empty())
-            this->fill_surfaces.append(intersection_ex(this_surfaces, this->fill_expolygons), SurfaceType(surface_type));
+            this->fill_surfaces.append(intersection_ex_by_piece(to_expolygons(this_surfaces), fill_boundaries), SurfaceType(surface_type));
     }
 }
 

src/libslic3r/MultiMaterialSegmentation.cpp

@@ -8,6 +8,7 @@
 #include "MutablePolygon.hpp"
 #include "format.hpp"
 
+#include <numeric>
 #include <utility>
 #include <unordered_set>
 
@@ -1311,10 +1312,15 @@ static inline std::vector<std::vector<ExPolygons>> segmentation_top_and_bottom_l
     }
 #endif // MM_SEGMENTATION_DEBUG_TOP_BOTTOM
 
-    // When the upper surface of an object is occluded, it should no longer be considered the upper surface
+    // When the upper surface of an object is occluded, it should no longer be considered the upper surface.
+    // Every (colour, layer) pair is trimmed on its own, so they all run at once: the painted faces of a finely
+    // textured part project hundreds of thousands of triangles onto one layer, which used to be trimmed serially.
     {
-        for (size_t extruder_idx = 0; extruder_idx < num_facets_states; ++extruder_idx) {
-            for (size_t layer_idx = 0; layer_idx < layers.size(); ++layer_idx) {
+        const size_t occluded_pairs = num_facets_states * layers.size();
+        tbb::parallel_for(tbb::blocked_range<size_t>(0, occluded_pairs), [&](const tbb::blocked_range<size_t> &range) {
+            for (size_t pair_idx = range.begin(); pair_idx < range.end(); ++pair_idx) {
+                const size_t extruder_idx = pair_idx / layers.size();
+                const size_t layer_idx    = pair_idx % layers.size();
                 if (!top_raw[extruder_idx].empty() && !top_raw[extruder_idx][layer_idx].empty() && layer_idx + 1 < layers.size()) {
                     top_raw[extruder_idx][layer_idx] = diff(top_raw[extruder_idx][layer_idx], input_expolygons[layer_idx + 1]);
                 }
@@ -1322,7 +1328,7 @@ static inline std::vector<std::vector<ExPolygons>> segmentation_top_and_bottom_l
                     bottom_raw[extruder_idx][layer_idx] = diff(bottom_raw[extruder_idx][layer_idx], input_expolygons[layer_idx - 1]);
                 }
             }
-        }
+        });
     }
 
     std::vector<std::vector<ExPolygons>> triangles_by_color_bottom(num_facets_states);
@@ -1378,13 +1384,62 @@ static inline std::vector<std::vector<ExPolygons>> segmentation_top_and_bottom_l
         return out;
     };
 
+    // Projects a painted top or bottom face `ex` of layer `layer_idx` onto the shell layers below or above it (in
+    // `shell_layers`, nearest first), one more perimeter in on each, stopping at the first layer where nothing is left.
+    // Only the slices within the deepest offset of `ex` (three times that with the miter joins) decide the result, so the
+    // work is done per tile of `ex`'s ExPolygons on the slices cut to the tile's box grown by that much: the same result, but
+    // each ClipperLib call stays the size of a tile rather than of a layer cut through a fine relief, and the tiles run in
+    // parallel.
+    const auto project_to_shells = [&input_expolygons](const ExPolygons &ex, size_t layer_idx, const std::vector<size_t> &shell_layers,
+                                                       const LayerColorStat &stat, std::vector<ExPolygons> &dst, size_t dst_offset) {
+        std::vector<float> offsets(shell_layers.size());
+        float              offset = 0.f;
+        for (size_t i = 0; i < shell_layers.size(); ++i) {
+            //BBS: offset width should be 2*spacing to avoid too narrow area which has overlap of wall line
+            offset -= (stat.extrusion_spacing + stat.extrusion_width);
+            offsets[i] = offset;
+        }
+        if (offsets.empty())
+            return;
+        const coord_t reach = coord_t(std::ceil(DefaultMiterLimit * std::abs(offsets.back()))) + 10 * SCALED_EPSILON;
+        const std::vector<ClipperUtils::ExPolygonsTile> tiles = ClipperUtils::tile_expolygons(ex, 16);
+        // [shell layer][tile]
+        std::vector<std::vector<ExPolygons>> shells(shell_layers.size(), std::vector<ExPolygons>(tiles.size()));
+        tbb::parallel_for(size_t(0), tiles.size(), [&](size_t tile_idx) {
+            const ClipperUtils::ExPolygonsTile &tile = tiles[tile_idx];
+            const BoundingBox                   bbox = tile.bbox.inflated(reach);
+            ExPolygons                          tile_ex;
+            tile_ex.reserve(tile.members.size());
+            for (size_t i : tile.members)
+                tile_ex.emplace_back(ex[i]);
+            Polygons layer_slices_trimmed = ClipperUtils::clip_clipper_polygons_with_subject_bbox(input_expolygons[layer_idx], bbox);
+            for (size_t i = 0; i < shell_layers.size() && ! layer_slices_trimmed.empty(); ++i) {
+                const ExPolygons trimmed = intersection_ex(layer_slices_trimmed, ClipperUtils::clip_clipper_polygons_with_subject_bbox(input_expolygons[shell_layers[i]], bbox));
+                shells[i][tile_idx]  = opening_ex(intersection_ex(tile_ex, offset_ex(trimmed, offsets[i])), stat.small_region_threshold);
+                layer_slices_trimmed = to_polygons(trimmed);
+            }
+        });
+        for (size_t i = 0; i < shell_layers.size(); ++i) {
+            bool empty = true;
+            for (ExPolygons &shell : shells[i])
+                if (! shell.empty()) {
+                    append(dst[shell_layers[i] + dst_offset], std::move(shell));
+                    empty = false;
+                }
+            if (empty)
+                break;
+        }
+    };
+
     tbb::parallel_for(tbb::blocked_range<size_t>(0, num_layers, granularity), [&granularity, &num_layers, &num_facets_states, &layer_color_stat, &top_raw, &triangles_by_color_top,
-                                                                               &throw_on_cancel_callback, &input_expolygons, &bottom_raw, &triangles_by_color_bottom,
+                                                                               &throw_on_cancel_callback, &bottom_raw, &triangles_by_color_bottom, &project_to_shells,
                                                                                &shell_triangles_by_color_top, &shell_triangles_by_color_bottom](const tbb::blocked_range<size_t> &range) {
         size_t group_idx   = range.begin() / granularity;
         size_t layer_idx_offset = (group_idx & 1) * num_layers;
         for (size_t layer_idx = range.begin(); layer_idx < range.end(); ++ layer_idx) {
-            for (size_t color_idx = 0; color_idx < num_facets_states; ++color_idx) {
+            // Each colour writes only its own vectors, so the colours run in parallel: a painted top or bottom face
+            // projects onto a single layer, which otherwise did all of its colours on one thread.
+            tbb::parallel_for(size_t(0), size_t(num_facets_states), [&](size_t color_idx) {
                 throw_on_cancel_callback();
                 LayerColorStat stat = layer_color_stat(layer_idx, color_idx);
                 if (std::vector<Polygons> &top = top_raw[color_idx]; ! top.empty() && ! top[layer_idx].empty())
@@ -1393,18 +1448,10 @@ static inline std::vector<std::vector<ExPolygons>> segmentation_top_and_bottom_l
                         top_ex = opening_ex(top_ex, stat.small_region_threshold);
                         if (! top_ex.empty()) {
                             append(triangles_by_color_top[color_idx][layer_idx + layer_idx_offset], top_ex);
-                            float offset = 0.f;
-                            ExPolygons layer_slices_trimmed = input_expolygons[layer_idx];
-                            for (int last_idx = int(layer_idx) - 1; last_idx > std::max(int(layer_idx - stat.top_shell_layers), int(0)); --last_idx) {
-                                //BBS: offset width should be 2*spacing to avoid too narrow area which has overlap of wall line
-                                //offset -= stat.extrusion_width ;
-                                offset -= (stat.extrusion_spacing + stat.extrusion_width);
-                                layer_slices_trimmed = intersection_ex(layer_slices_trimmed, input_expolygons[last_idx]);
-                                ExPolygons last = opening_ex(intersection_ex(top_ex, offset_ex(layer_slices_trimmed, offset)), stat.small_region_threshold);
-                                if (last.empty())
-                                    break;
-                                append(shell_triangles_by_color_top[color_idx][last_idx + layer_idx_offset], std::move(last));
-                            }
+                            std::vector<size_t> shell_layers;
+                            for (int last_idx = int(layer_idx) - 1; last_idx > std::max(int(layer_idx - stat.top_shell_layers), int(0)); --last_idx)
+                                shell_layers.emplace_back(size_t(last_idx));
+                            project_to_shells(top_ex, layer_idx, shell_layers, stat, shell_triangles_by_color_top[color_idx], layer_idx_offset);
                         }
                     }
                 if (std::vector<Polygons> &bottom = bottom_raw[color_idx]; ! bottom.empty() && ! bottom[layer_idx].empty())
@@ -1413,21 +1460,13 @@ static inline std::vector<std::vector<ExPolygons>> segmentation_top_and_bottom_l
                         bottom_ex = opening_ex(bottom_ex, stat.small_region_threshold);
                         if (! bottom_ex.empty()) {
                             append(triangles_by_color_bottom[color_idx][layer_idx + layer_idx_offset], bottom_ex);
-                            float offset = 0.f;
-                            ExPolygons layer_slices_trimmed = input_expolygons[layer_idx];
-                            for (size_t last_idx = layer_idx + 1; last_idx < std::min(layer_idx + stat.bottom_shell_layers, num_layers); ++last_idx) {
-                                //BBS: offset width should be 2*spacing to avoid too narrow area which has overlap of wall line
-                                //offset -= stat.extrusion_width;
-                                offset -= (stat.extrusion_spacing + stat.extrusion_width);
-                                layer_slices_trimmed = intersection_ex(layer_slices_trimmed, input_expolygons[last_idx]);
-                                ExPolygons last = opening_ex(intersection_ex(bottom_ex, offset_ex(layer_slices_trimmed, offset)), stat.small_region_threshold);
-                                if (last.empty())
-                                    break;
-                                append(shell_triangles_by_color_bottom[color_idx][last_idx + layer_idx_offset], std::move(last));
-                            }
-                        }
+                            std::vector<size_t> shell_layers;
+                            for (size_t last_idx = layer_idx + 1; last_idx < std::min(layer_idx + stat.bottom_shell_layers, num_layers); ++last_idx)
+                                shell_layers.emplace_back(last_idx);
+                            project_to_shells(bottom_ex, layer_idx, shell_layers, stat, shell_triangles_by_color_bottom[color_idx], layer_idx_offset);
                         }
                     }
+            });
         }
     });
 
@@ -1437,22 +1476,25 @@ static inline std::vector<std::vector<ExPolygons>> segmentation_top_and_bottom_l
                                                                   &shell_triangles_by_color_top, &shell_triangles_by_color_bottom](const tbb::blocked_range<size_t> &range) {
         for (size_t layer_idx = range.begin(); layer_idx < range.end(); ++ layer_idx) {
             throw_on_cancel_callback();
-            ExPolygons painted_exploys;
-            for (size_t color_idx = 0; color_idx < triangles_by_color_merged.size(); ++color_idx) {
+            // The per-colour unions below are independent of each other, so they run in parallel (a painted top or
+            // bottom face puts all of its colours on one layer); whatever combines the colours stays in colour order.
+            const auto merge_colour_union = [&](size_t color_idx) {
                 auto &self = triangles_by_color_merged[color_idx][layer_idx];
                 append(self, std::move(triangles_by_color_bottom[color_idx][layer_idx]));
                 append(self, std::move(triangles_by_color_bottom[color_idx][layer_idx + num_layers]));
                 append(self, std::move(triangles_by_color_top[color_idx][layer_idx]));
                 append(self, std::move(triangles_by_color_top[color_idx][layer_idx + num_layers]));
                 self = union_ex(self);
+            };
+            tbb::parallel_for(size_t(0), triangles_by_color_merged.size(), merge_colour_union);
 
-                append(painted_exploys, self);
-            }
-
+            ExPolygons painted_exploys;
+            for (size_t color_idx = 0; color_idx < triangles_by_color_merged.size(); ++color_idx)
+                append(painted_exploys, triangles_by_color_merged[color_idx][layer_idx]);
             painted_exploys = union_ex(painted_exploys);
 
             //BBS: merge the top and bottom shell layers
-            for (size_t color_idx = 0; color_idx < triangles_by_color_merged.size(); ++color_idx) {
+            tbb::parallel_for(size_t(0), triangles_by_color_merged.size(), [&](size_t color_idx) {
                 auto &self = triangles_by_color_merged[color_idx][layer_idx];
 
                 auto top_area = diff_ex(union_ex(shell_triangles_by_color_top[color_idx][layer_idx],
@@ -1466,7 +1508,7 @@ static inline std::vector<std::vector<ExPolygons>> segmentation_top_and_bottom_l
                 append(self, top_area);
                 append(self, bottom_area);
                 self = union_ex(self);
-            }
+            });
             // Trim one region by the other if some of the regions overlap.
             ExPolygons painted_regions;
             for (size_t color_idx = 1; color_idx < triangles_by_color_merged.size(); ++color_idx) {
@@ -1833,7 +1875,69 @@ static void remove_multiple_edges_in_vertices(MMU_Graph &graph, const std::vecto
     }
 }
 
-static std::vector<std::vector<ExPolygons>> merge_segmented_layers(const std::vector<std::vector<ExPolygons>> &segmented_regions,
+
+// Finds the islands (layer ExPolygons) a region piece overlaps. A top or bottom region is projected from the neighbouring
+// layers and may reach past the island it belongs to, or over several islands.
+class IslandLocator
+{
+public:
+    explicit IslandLocator(const ExPolygons &islands) : m_islands(islands)
+    {
+        m_bboxes.reserve(islands.size());
+        for (const ExPolygon &island : islands) {
+            m_bboxes.emplace_back(get_extents(island));
+            m_extent.merge(m_bboxes.back());
+        }
+        if (!m_extent.defined)
+            return;
+        const Point size = m_extent.size();
+        m_cell_w = std::max<coord_t>(1, size.x() / GRID + 1);
+        m_cell_h = std::max<coord_t>(1, size.y() / GRID + 1);
+        m_grid.assign(GRID * GRID, {});
+        for (size_t i = 0; i < m_bboxes.size(); ++i)
+            for_cells(m_bboxes[i], [&](int cell) { m_grid[cell].emplace_back(i); });
+    }
+
+    void find(const ExPolygon &piece, std::vector<size_t> &out) const
+    {
+        out.clear();
+        const BoundingBox bbox = get_extents(piece);
+        if (!m_extent.defined || !m_extent.overlap(bbox))
+            return;
+        for_cells(bbox, [&](int cell) {
+            for (size_t i : m_grid[cell])
+                if (m_bboxes[i].overlap(bbox))
+                    out.emplace_back(i);
+        });
+        sort_remove_duplicates(out);
+        if (out.size() > 1)
+            out.erase(std::remove_if(out.begin(), out.end(), [&](size_t i) {
+                const BoundingBox common(m_bboxes[i].min.cwiseMax(bbox.min), m_bboxes[i].max.cwiseMin(bbox.max));
+                return intersection(ClipperUtils::clip_clipper_polygons_with_subject_bbox(piece, common.inflated(SCALED_EPSILON)),
+                                    ClipperUtils::clip_clipper_polygons_with_subject_bbox(m_islands[i], common.inflated(SCALED_EPSILON))).empty();
+            }), out.end());
+    }
+
+private:
+    static constexpr int GRID = 64;
+    template<typename Fn> void for_cells(const BoundingBox &bb, Fn &&fn) const
+    {
+        const int x0 = std::clamp(int((bb.min.x() - m_extent.min.x()) / m_cell_w), 0, GRID - 1), x1 = std::clamp(int((bb.max.x() - m_extent.min.x()) / m_cell_w), 0, GRID - 1);
+        const int y0 = std::clamp(int((bb.min.y() - m_extent.min.y()) / m_cell_h), 0, GRID - 1), y1 = std::clamp(int((bb.max.y() - m_extent.min.y()) / m_cell_h), 0, GRID - 1);
+        for (int y = y0; y <= y1; ++y)
+            for (int x = x0; x <= x1; ++x)
+                fn(y * GRID + x);
+    }
+
+    const ExPolygons                &m_islands;
+    std::vector<BoundingBox>         m_bboxes;
+    BoundingBox                      m_extent;
+    coord_t                          m_cell_w = 1, m_cell_h = 1;
+    std::vector<std::vector<size_t>> m_grid;
+};
+
+static std::vector<std::vector<ExPolygons>> merge_segmented_layers(const std::vector<ExPolygons>              &input_expolygons,
+                                                                   const std::vector<std::vector<ExPolygons>> &segmented_regions,
                                                                    std::vector<std::vector<ExPolygons>>      &&top_and_bottom_layers,
                                                                    const size_t                                num_facets_states,
                                                                    const std::function<void()>                &throw_on_cancel_callback)
@@ -1844,33 +1948,91 @@ static std::vector<std::vector<ExPolygons>> merge_segmented_layers(const std::ve
     assert(!top_and_bottom_layers.size() || num_facets_states == top_and_bottom_layers.size());
 
     BOOST_LOG_TRIVIAL(debug) << "Print object segmentation - Merging segmented layers in parallel - Begin";
-    tbb::parallel_for(tbb::blocked_range<size_t>(0, num_layers), [&segmented_regions, &top_and_bottom_layers, &segmented_regions_merged, &num_facets_states, &throw_on_cancel_callback](const tbb::blocked_range<size_t> &range) {
+    // Every region of a layer is merged together with the regions of the islands it overlaps, and the islands are further
+    // apart than the dimple removal below reaches, so this gives the same result as merging the layer at once. On a layer
+    // cut through a fine relief every region shares thousands of hole contours with every other, and ClipperLib, splitting
+    // and re-linking one huge polygon over and over, took anything up to half an hour for a layer; per island each operation
+    // stays the size of the island, and the islands run in parallel.
+    tbb::parallel_for(tbb::blocked_range<size_t>(0, num_layers), [&](const tbb::blocked_range<size_t> &range) {
         for (size_t layer_idx = range.begin(); layer_idx < range.end(); ++layer_idx) {
             assert(segmented_regions[layer_idx].size() == num_facets_states);
-            // Zero is skipped because it is the default color of the volume
-            for (size_t extruder_id = 1; extruder_id < num_facets_states; ++extruder_id) {
             throw_on_cancel_callback();
-                if (!segmented_regions[layer_idx][extruder_id].empty()) {
-                    ExPolygons segmented_regions_trimmed = segmented_regions[layer_idx][extruder_id];
-                    if (!top_and_bottom_layers.empty()) {
-                        for (const std::vector<ExPolygons> &top_and_bottom_by_extruder : top_and_bottom_layers) {
-                            if (!top_and_bottom_by_extruder[layer_idx].empty() && !segmented_regions_trimmed.empty()) {
-                                segmented_regions_trimmed = diff_ex(segmented_regions_trimmed, top_and_bottom_by_extruder[layer_idx]);
-                            }
-                        }
+            // Group the islands joined by a region overlapping several of them; the last group takes the regions lying
+            // outside every island.
+            const ExPolygons   &islands = input_expolygons[layer_idx];
+            const IslandLocator locator(islands);
+            std::vector<size_t> parent(islands.size() + 1);
+            std::iota(parent.begin(), parent.end(), 0);
+            const auto root = [&parent](size_t i) {
+                while (parent[i] != i)
+                    i = parent[i] = parent[parent[i]];
+                return i;
+            };
+            // Islands of every piece: side regions of colours 1.., then top/bottom regions of colours 0..
+            std::vector<const ExPolygon *> pieces;
+            for (size_t extruder_id = 1; extruder_id < num_facets_states; ++extruder_id)
+                for (const ExPolygon &piece : segmented_regions[layer_idx][extruder_id])
+                    pieces.emplace_back(&piece);
+            if (!top_and_bottom_layers.empty())
+                for (size_t color_idx = 0; color_idx < num_facets_states; ++color_idx)
+                    for (const ExPolygon &piece : top_and_bottom_layers[color_idx][layer_idx])
+                        pieces.emplace_back(&piece);
+            std::vector<std::vector<size_t>> overlapped(pieces.size());
+            tbb::parallel_for(size_t(0), pieces.size(), [&](size_t i) { locator.find(*pieces[i], overlapped[i]); });
+            std::vector<size_t> piece_island(pieces.size());
+            for (size_t i = 0; i < pieces.size(); ++i) {
+                piece_island[i] = overlapped[i].empty() ? islands.size() : overlapped[i].front();
+                for (size_t island : overlapped[i])
+                    parent[root(island)] = root(piece_island[i]);
             }
+            std::vector<size_t> bucket_of(parent.size(), size_t(-1));
+            size_t              num_buckets = 0;
+            for (size_t i = 0; i < parent.size(); ++i)
+                if (size_t &b = bucket_of[root(i)]; b == size_t(-1))
+                    b = num_buckets++;
 
-                    segmented_regions_merged[layer_idx][extruder_id - 1] = std::move(segmented_regions_trimmed);
-                }
+            // [bucket][colour]
+            std::vector<std::vector<ExPolygons>> sides(num_buckets, std::vector<ExPolygons>(num_facets_states));
+            std::vector<std::vector<ExPolygons>> tops(num_buckets, std::vector<ExPolygons>(num_facets_states));
+            size_t                               piece_idx = 0;
+            for (size_t extruder_id = 1; extruder_id < num_facets_states; ++extruder_id)
+                for (const ExPolygon &piece : segmented_regions[layer_idx][extruder_id])
+                    sides[bucket_of[root(piece_island[piece_idx++])]][extruder_id].emplace_back(piece);
+            if (!top_and_bottom_layers.empty())
+                for (size_t color_idx = 0; color_idx < num_facets_states; ++color_idx)
+                    for (const ExPolygon &piece : top_and_bottom_layers[color_idx][layer_idx])
+                        tops[bucket_of[root(piece_island[piece_idx++])]][color_idx].emplace_back(piece);
 
-                if (!top_and_bottom_layers.empty() && !top_and_bottom_layers[extruder_id][layer_idx].empty()) {
-                    bool was_top_and_bottom_empty = segmented_regions_merged[layer_idx][extruder_id - 1].empty();
-                    append(segmented_regions_merged[layer_idx][extruder_id - 1], top_and_bottom_layers[extruder_id][layer_idx]);
+            // Side regions minus the top/bottom regions of every colour.
+            std::vector<std::vector<ExPolygons>> merged(num_buckets, std::vector<ExPolygons>(num_facets_states));
+            tbb::parallel_for(size_t(0), num_buckets, [&](size_t bucket) {
+                Polygons tops_all;
+                for (const ExPolygons &t : tops[bucket])
+                    polygons_append(tops_all, t);
+                for (size_t extruder_id = 1; extruder_id < num_facets_states; ++extruder_id)
+                    if (!sides[bucket][extruder_id].empty())
+                        merged[bucket][extruder_id] = tops_all.empty() ? std::move(sides[bucket][extruder_id]) :
+                                                                         diff_ex_by_piece(sides[bucket][extruder_id], tops_all);
+            });
 
-                    // Remove dimples (#7235) appearing after merging side segmentation of the model with tops and bottoms painted layers.
-                    if (!was_top_and_bottom_empty)
-                        segmented_regions_merged[layer_idx][extruder_id - 1] = offset2_ex(union_ex(segmented_regions_merged[layer_idx][extruder_id - 1]), float(SCALED_EPSILON), -float(SCALED_EPSILON));
+            // Then this colour's top/bottom regions, with the dimples removed (#7235) when the layer has side regions left.
+            for (size_t extruder_id = 1; extruder_id < num_facets_states; ++extruder_id) {
+                if (top_and_bottom_layers.empty() || top_and_bottom_layers[extruder_id][layer_idx].empty()) {
+                    for (size_t bucket = 0; bucket < num_buckets; ++bucket)
+                        append(segmented_regions_merged[layer_idx][extruder_id - 1], std::move(merged[bucket][extruder_id]));
+                    continue;
                 }
+                bool was_top_and_bottom_empty = true;
+                for (size_t bucket = 0; bucket < num_buckets && was_top_and_bottom_empty; ++bucket)
+                    was_top_and_bottom_empty = merged[bucket][extruder_id].empty();
+                tbb::parallel_for(size_t(0), num_buckets, [&](size_t bucket) {
+                    ExPolygons &region = merged[bucket][extruder_id];
+                    append(region, tops[bucket][extruder_id]);
+                    if (!was_top_and_bottom_empty && !region.empty())
+                        region = offset2_ex(union_ex(region), float(SCALED_EPSILON), -float(SCALED_EPSILON));
+                });
+                for (size_t bucket = 0; bucket < num_buckets; ++bucket)
+                    append(segmented_regions_merged[layer_idx][extruder_id - 1], std::move(merged[bucket][extruder_id]));
             }
         }
     }); // end of parallel_for
@@ -2157,16 +2319,56 @@ std::vector<std::vector<ExPolygons>> segmentation_by_painting(const PrintObject
 
                 assert(!color_poly.empty());
                 assert(!color_poly.front().empty());
-                if (has_layer_only_one_color(color_poly)) {
-                    // If the whole layer is painted using the same color, it is not needed to construct a Voronoi diagram for the segmentation of this layer.
-                    segmented_regions[layer_idx][size_t(color_poly.front().front().color)] = input_expolygons[layer_idx];
-                } else {
-                    MMU_Graph graph = build_graph(layer_idx, color_poly);
-                    remove_multiple_edges_in_vertices(graph, color_poly);
-                    graph.remove_nodes_with_one_arc();
-                    segmented_regions[layer_idx] = extract_colored_segments(graph, num_facets_states);
-                    //segmented_regions[layer_idx] = extract_colored_segments(color_poly, num_extruders, layer_idx);
+                // Each island (an ExPolygon with its holes) is segmented on its own. Any point of an island is closer to
+                // that island's contours than to any other island's - the way out crosses its own boundary first - so its
+                // Voronoi cells, and with them its colour regions, depend on nothing else. A layer cut through a fine relief
+                // has thousands of islands, and one Voronoi diagram over all of them degenerated into overlapping regions
+                // that every boolean afterwards had to untangle. Per island the diagrams stay small and the islands run in
+                // parallel; an island in a single colour needs no diagram at all.
+                const ExPolygons                      &islands = input_expolygons[layer_idx];
+                std::vector<std::pair<size_t, size_t>> island_contours(islands.size()); // [first, last) into color_poly
+                {
+                    // The same order EdgeGrid::Grid::create() lists the contours in, and so colorize_contours().
+                    size_t idx = 0;
+                    for (size_t island_idx = 0; island_idx < islands.size(); ++island_idx) {
+                        const size_t first = idx;
+                        if (!islands[island_idx].contour.empty())
+                            ++idx;
+                        for (const Polygon &hole : islands[island_idx].holes)
+                            if (!hole.empty())
+                                ++idx;
+                        island_contours[island_idx] = {first, idx};
                     }
+                    assert(idx == color_poly.size());
+                }
+                std::vector<std::vector<ExPolygons>> island_regions(islands.size());
+                tbb::parallel_for(size_t(0), islands.size(), [&](size_t island_idx) {
+                    const auto [first, last] = island_contours[island_idx];
+                    if (first == last)
+                        return;
+                    const std::vector<ColoredLines> island_poly(color_poly.begin() + first, color_poly.begin() + last);
+                    std::vector<ExPolygons>        &regions = island_regions[island_idx];
+                    if (has_layer_only_one_color(island_poly)) {
+                        regions.assign(num_facets_states, ExPolygons());
+                        regions[size_t(island_poly.front().front().color)].emplace_back(islands[island_idx]);
+                    } else {
+                        MMU_Graph graph = build_graph(layer_idx, island_poly);
+                        remove_multiple_edges_in_vertices(graph, island_poly);
+                        graph.remove_nodes_with_one_arc();
+                        regions = extract_colored_segments(graph, num_facets_states);
+                        // The faces of one colour tile it without overlapping; merged here, where an island is small,
+                        // every later boolean gets a few regions instead of thousands of faces sharing their edges. An
+                        // island with many holes keeps its faces: merged, each colour would be one region with thousands
+                        // of holes, and subtracting from that is far slower than from the faces one at a time.
+                        if (island_poly.size() <= 64)
+                            for (ExPolygons &faces : regions)
+                                if (faces.size() > 1)
+                                    faces = union_ex(faces);
+                    }
+                });
+                for (std::vector<ExPolygons> &regions : island_regions)
+                    for (size_t color_idx = 0; color_idx < regions.size(); ++color_idx)
+                        append(segmented_regions[layer_idx][color_idx], std::move(regions[color_idx]));
 
 #ifdef MM_SEGMENTATION_DEBUG_REGIONS
                 export_regions_to_svg(debug_out_path("3-mm-regions-sides-%d-%d.svg", layer_idx, iRun), segmented_regions[layer_idx], input_expolygons[layer_idx]);
@@ -2189,7 +2391,7 @@ std::vector<std::vector<ExPolygons>> segmentation_by_painting(const PrintObject
         throw_on_cancel_callback();
     }
 
-    std::vector<std::vector<ExPolygons>> segmented_regions_merged = merge_segmented_layers(segmented_regions, std::move(top_and_bottom_layers), num_facets_states, throw_on_cancel_callback);
+    std::vector<std::vector<ExPolygons>> segmented_regions_merged = merge_segmented_layers(input_expolygons, segmented_regions, std::move(top_and_bottom_layers), num_facets_states, throw_on_cancel_callback);
     throw_on_cancel_callback();
 
 #ifdef MM_SEGMENTATION_DEBUG_REGIONS

src/libslic3r/MultiPoint.hpp

@@ -19,11 +19,13 @@ public:
     
     MultiPoint() {}
     MultiPoint(const MultiPoint &other) : points(other.points) {}
-    MultiPoint(MultiPoint &&other) : points(std::move(other.points)) {}
+    MultiPoint(MultiPoint &&other) noexcept : points(std::move(other.points)) {}
     MultiPoint(std::initializer_list<Point> list) : points(list) {}
     explicit MultiPoint(const Points &_points) : points(_points) {}
+    // Without it, the derived classes' move constructors passing std::move(points) here copied them.
+    explicit MultiPoint(Points &&_points) noexcept : points(std::move(_points)) {}
     MultiPoint& operator=(const MultiPoint &other) { points = other.points; return *this; }
-    MultiPoint& operator=(MultiPoint &&other) { points = std::move(other.points); return *this; }
+    MultiPoint& operator=(MultiPoint &&other) noexcept { points = std::move(other.points); return *this; }
     virtual ~MultiPoint() = default;
     void scale(double factor);
     void scale(double factor_x, double factor_y);

src/libslic3r/PerimeterGenerator.cpp

@@ -17,6 +17,8 @@
 #include <cassert>
 #include <unordered_set>
 #include <thread>
+#include <tbb/blocked_range.h>
+#include <tbb/parallel_for.h>
 #include "libslic3r/AABBTreeLines.hpp"
 #include "Print.hpp"
 static const int overhang_sampling_number = 6;
@@ -2481,7 +2483,22 @@ void PerimeterGenerator::process_arachne()
     const bool only_one_wall_first_layer = this->config->only_one_wall_first_layer && has_bottom_shell_layers(*this->config);
     // we need to process each island separately because we might have different
     // extra perimeters for each one
-    for (const Surface& surface : all_surfaces) {
+    // Each island is independent up to its outputs, so they are generated in parallel - a layer split into thousands
+    // of islands (e.g. by colour painting) otherwise ran on one thread - and the outputs are then committed in the
+    // original island order, which is what the extra overhang perimeters (applied to the last island's loops and to
+    // all fill surfaces so far) depend on.
+    struct ArachneSurfaceResult
+    {
+        ExtrusionEntityCollection loops;
+        bool                      has_loops = false;
+        ExPolygons                infill;
+        ExPolygons                no_overlap;
+    };
+    std::vector<ArachneSurfaceResult> results(all_surfaces.size());
+    tbb::parallel_for(tbb::blocked_range<size_t>(0, all_surfaces.size()), [&](const tbb::blocked_range<size_t> &range) {
+      for (size_t surface_idx = range.begin(); surface_idx < range.end(); ++surface_idx) {
+        const Surface        &surface = all_surfaces[surface_idx];
+        ArachneSurfaceResult &result  = results[surface_idx];
             coord_t bead_width_0 = ext_perimeter_spacing;
             // detect how many perimeters must be generated for this island
             int loop_number = this->config->wall_loops + surface.extra_perimeters - 1; // 0-indexed loops
@@ -2838,7 +2855,8 @@ void PerimeterGenerator::process_arachne()
                                         this->config->overhang_reverse_internal_only);
                 }
                 defer_unsupported_loops(*this, extrusion_coll);
-            this->loops->append(extrusion_coll);
+                result.loops     = std::move(extrusion_coll);
+                result.has_loops = true;
             }
 
             const coord_t spacing = (perimeters.size() == 1) ? ext_perimeter_spacing2 : perimeter_spacing;
@@ -2881,9 +2899,7 @@ void PerimeterGenerator::process_arachne()
             if (!top_expolygons.empty()) {
                 infill_exp = union_ex(infill_exp, offset_ex(top_expolygons, double(top_inset)));
             }
-        this->fill_surfaces->append(infill_exp, stInternal);
-
-        apply_extra_perimeters(infill_exp);
+            result.infill = std::move(infill_exp);
 
             // BBS: get the no-overlap infill expolygons
             {
@@ -2894,9 +2910,17 @@ void PerimeterGenerator::process_arachne()
                     float(+min_perimeter_infill_spacing / 2.));
                 if (!top_expolygons.empty())
                     polyWithoutOverlap = union_ex(polyWithoutOverlap, top_expolygons);
-            this->fill_no_overlap->insert(this->fill_no_overlap->end(), polyWithoutOverlap.begin(), polyWithoutOverlap.end());
+                result.no_overlap = std::move(polyWithoutOverlap);
             }
       }
+    });
+    for (ArachneSurfaceResult &result : results) {
+        if (result.has_loops)
+            this->loops->append(result.loops);
+        this->fill_surfaces->append(result.infill, stInternal);
+        apply_extra_perimeters(result.infill);
+        this->fill_no_overlap->insert(this->fill_no_overlap->end(), result.no_overlap.begin(), result.no_overlap.end());
+    }
 }
 
 bool PerimeterGeneratorLoop::is_internal_contour() const

src/libslic3r/Polygon.hpp

@@ -27,7 +27,7 @@ public:
     explicit Polygon(const Points &points) : MultiPoint(points) {}
 	Polygon(std::initializer_list<Point> points) : MultiPoint(points) {}
     Polygon(const Polygon &other) : MultiPoint(other.points) {}
-    Polygon(Polygon &&other) : MultiPoint(std::move(other.points)) {}
+    Polygon(Polygon &&other) noexcept : MultiPoint(std::move(other.points)) {}
 	static Polygon new_scale(const std::vector<Vec2d> &points) { 
         Polygon pgn;
         pgn.points.reserve(points.size());
@@ -36,7 +36,7 @@ public:
 		return pgn;
 	}
     Polygon& operator=(const Polygon &other) { points = other.points; return *this; }
-    Polygon& operator=(Polygon &&other) { points = std::move(other.points); return *this; }
+    Polygon& operator=(Polygon &&other) noexcept { points = std::move(other.points); return *this; }
 
     Point& operator[](Points::size_type idx) { return this->points[idx]; }
     const Point& operator[](Points::size_type idx) const { return this->points[idx]; }

src/libslic3r/Polyline.hpp

@@ -20,7 +20,7 @@ class Polyline : public MultiPoint {
 public:
     Polyline() {};
     Polyline(const Polyline& other) : MultiPoint(other.points), fitting_result(other.fitting_result) {}
-    Polyline(Polyline &&other) : MultiPoint(std::move(other.points)), fitting_result(std::move(other.fitting_result))  {}
+    Polyline(Polyline &&other) noexcept : MultiPoint(std::move(other.points)), fitting_result(std::move(other.fitting_result))  {}
     Polyline(std::initializer_list<Point> list) : MultiPoint(list) { 
         fitting_result.clear();
     }
@@ -41,7 +41,7 @@ public:
         fitting_result = other.fitting_result;
         return *this;
     }
-    Polyline& operator=(Polyline&& other) {
+    Polyline& operator=(Polyline&& other) noexcept {
         points = std::move(other.points);
         fitting_result = std::move(other.fitting_result);
         return *this;

src/libslic3r/PrintObject.cpp

@@ -30,6 +30,7 @@
 
 #include <cstddef>
 #include <float.h>
+#include <array>
 #include <iterator>
 #include <mutex>
 #include <string>
@@ -42,6 +43,7 @@
 #include <boost/log/trivial.hpp>
 
 #include <tbb/parallel_for.h>
+#include <tbb/parallel_invoke.h>
 #include <tbb/spin_mutex.h>
 #include <tbb/concurrent_unordered_set.h>
 
@@ -1664,7 +1666,9 @@ void PrintObject::detect_surfaces_type()
     bool interface_shells = ! spiral_mode && m_config.interface_shells.value;
     size_t num_layers     = spiral_mode ? std::min(size_t(this->printing_region(0).config().bottom_shell_layers), m_layers.size()) : m_layers.size();
 
-    for (size_t region_id = 0; region_id < this->num_printing_regions(); ++ region_id) {
+    // The regions of a layer do not see each other here, and a layer cut through a fine relief takes far longer than the
+    // others, so the regions run next to each other instead of one after another, each still over all layers.
+    tbb::parallel_for(size_t(0), this->num_printing_regions(), [&](size_t region_id) {
         BOOST_LOG_TRIVIAL(debug) << "Detecting solid surfaces for region " << region_id << " in parallel - start";
 #ifdef SLIC3R_DEBUG_SLICE_PROCESSING
         for (Layer *layer : m_layers)
@@ -1722,7 +1726,7 @@ void PrintObject::detect_surfaces_type()
                     if (upper_layer) {
                         ExPolygons upper_slices = interface_shells ?
                             diff_ex(layerm_slices_surfaces, upper_layer->m_regions[region_id]->slices.surfaces, ApplySafetyOffset::Yes) :
-                            diff_ex(layerm_slices_surfaces, upper_layer->lslices, ApplySafetyOffset::Yes);
+                            diff_ex_by_piece(layerm_slices_surfaces, to_polygons(upper_layer->lslices), ApplySafetyOffset::Yes);
                         surfaces_append(top, opening_ex(upper_slices, offset), stTop);
                     } else {
                         // if no upper layer, all surfaces of this one are solid
@@ -1748,7 +1752,7 @@ void PrintObject::detect_surfaces_type()
                         surfaces_append(
                             bottom,
                             opening_ex(
-                                diff_ex(layerm_slices_surfaces, lower_layer->lslices, ApplySafetyOffset::Yes),
+                                diff_ex_by_piece(layerm_slices_surfaces, to_polygons(lower_layer->lslices), ApplySafetyOffset::Yes),
                                 offset),
                             surface_type_bottom_other);
                         // if user requested internal shells, we need to identify surfaces
@@ -1779,34 +1783,44 @@ void PrintObject::detect_surfaces_type()
                     // and top surfaces; let's do an intersection to discover them and consider them
                     // as bottom surfaces (to allow for bridge detection)
                     if (! top.empty() && ! bottom.empty()) {
-                        const auto cracks = intersection_ex(top, bottom);
+                        const auto cracks = intersection_ex_by_piece(to_expolygons(top), to_polygons(bottom));
                         if (!cracks.empty()) {
                             if (lower_layer) { // Only detect small cracks for non-first layer, because first layer should always be bottom
                                 const float small_crack_threshold = -layerm->flow(frExternalPerimeter).scaled_width() * 1.5;
                                 
+                                // Only the bottom surfaces near a crack can take part: one that contains it must contain its box,
+                                // and one whose box misses the grown crack is left unchanged by removing it. A layer cut through
+                                // a fine relief has thousands of both, which made this loop quadratic.
                                 for (const auto& crack : cracks) {
                                     if (offset_ex(crack, small_crack_threshold).empty()) {
                                         // For small cracks, if it's part of a large bottom surface, then it should be added to bottom as well
-                                        if (std::any_of(bottom.begin(), bottom.end(), [&crack, small_crack_threshold](const Surface& s) {
+                                        const BoundingBox crack_bbox = get_extents(crack);
+                                        if (std::any_of(bottom.begin(), bottom.end(), [&crack, &crack_bbox, small_crack_threshold](const Surface& s) {
                                                 const auto& se = s.expolygon;
-                                                return diff_ex(crack, se, ApplySafetyOffset::Yes).empty()
+                                                return get_extents(se).inflated(SCALED_EPSILON).contains(crack_bbox)
+                                                    && diff_ex(crack, se, ApplySafetyOffset::Yes).empty()
                                                     && se.area() > crack.area() * 2
                                                     && !offset_ex(diff_ex(se, crack), small_crack_threshold).empty();
                                         })) continue;
 
                                         // Crack too small, leave it as part of the top surface, remove it from bottom surfaces
+                                        const ExPolygons  grown_crack = offset_ex(crack, -small_crack_threshold);
+                                        const BoundingBox grown_bbox  = get_extents(grown_crack);
                                         Surfaces bot_tmp;
                                         for (auto& b : bottom) {
-                                            surfaces_append(bot_tmp, diff_ex(b.expolygon, offset_ex(crack, -small_crack_threshold)), b.surface_type);
+                                            if (get_extents(b.expolygon).overlap(grown_bbox))
+                                                surfaces_append(bot_tmp, diff_ex(b.expolygon, grown_crack), b.surface_type);
+                                            else
+                                                bot_tmp.emplace_back(std::move(b));
                                         }
                                         bottom = std::move(bot_tmp);
                                     }
                                 }
                             }
 
-                            Polygons top_polygons = to_polygons(std::move(top));
+                            ExPolygons top_expolygons = to_expolygons(std::move(top));
                             top.clear();
-                            surfaces_append(top, diff_ex(top_polygons, bottom), stTop);
+                            surfaces_append(top, diff_ex_by_piece(top_expolygons, to_polygons(bottom)), stTop);
                         }
                     }
 
@@ -1897,7 +1911,7 @@ void PrintObject::detect_surfaces_type()
                     {
                         Polygons topbottom = to_polygons(top);
                         polygons_append(topbottom, to_polygons(bottom));
-                        surfaces_append(surfaces_out, diff_ex(surfaces_prev_expolys, topbottom), stInternal);
+                        surfaces_append(surfaces_out, diff_ex_by_piece(surfaces_prev_expolys, topbottom), stInternal);
                     }
 
                     surfaces_append(surfaces_out, std::move(top));
@@ -2074,29 +2088,31 @@ void PrintObject::detect_surfaces_type()
                 }
             }
             );
-            // ==============================================================================================================
-            // === ORCA: Interim workaround - for now the new stInternalAfterExternalBridge surfaace is re-classified  ==============
-            // === back to a bottom bridge. As a starting point, this improves bridging reliability as it extrudes ==========
-            // === two external bridge layers. However, TODO: Implement a new surface type throughout the codebase ==========
-            // ==============================================================================================================
-            for (size_t region_id = 0; region_id < this->num_printing_regions(); ++region_id) {
-                tbb::parallel_for( tbb::blocked_range<size_t>(0, m_layers.size()), [this, region_id](const tbb::blocked_range<size_t> &range) {
-                    for (size_t idx_layer = range.begin(); idx_layer < range.end(); ++idx_layer) {
-                        Surfaces &surfs = m_layers[idx_layer]->m_regions[region_id]->slices.surfaces;
-                        for (Surface &s : surfs) {
-                            if (s.surface_type == stInternalAfterExternalBridge) {
-                                s.surface_type = stBottomBridge;
-                            }
-                        }
-                    }
-                }
-              );
-            }
         }
         // ==============================================================================================================
         // === ORCA: End of second external bridge layer changes  =======================================================
         // ==============================================================================================================
         
+    }); // for each this->print->region_count
+
+    // ==============================================================================================================
+    // === ORCA: Interim workaround - for now the new stInternalAfterExternalBridge surfaace is re-classified  ==============
+    // === back to a bottom bridge. As a starting point, this improves bridging reliability as it extrudes ==========
+    // === two external bridge layers. However, TODO: Implement a new surface type throughout the codebase ==========
+    // ==============================================================================================================
+    // Once all the regions have their second bridge layer, and before their slices are trimmed into fill surfaces below.
+    if ((this->config().enable_extra_bridge_layer.value == eblApplyToAll) || (this->config().enable_extra_bridge_layer.value == eblExternalBridgeOnly)) {
+        tbb::parallel_for(tbb::blocked_range<size_t>(0, m_layers.size()), [this](const tbb::blocked_range<size_t> &range) {
+            for (size_t idx_layer = range.begin(); idx_layer < range.end(); ++idx_layer)
+                for (LayerRegion *layerm : m_layers[idx_layer]->regions())
+                    for (Surface &s : layerm->slices.surfaces)
+                        if (s.surface_type == stInternalAfterExternalBridge)
+                            s.surface_type = stBottomBridge;
+        });
+        m_print->throw_if_canceled();
+    }
+
+    tbb::parallel_for(size_t(0), this->num_printing_regions(), [&](size_t region_id) {
         BOOST_LOG_TRIVIAL(debug) << "Detecting solid surfaces for region " << region_id << " - clipping in parallel - start";
         // Fill in layerm->fill_surfaces by trimming the layerm->slices by the cummulative layerm->fill_surfaces.
         tbb::parallel_for(
@@ -2113,7 +2129,7 @@ void PrintObject::detect_surfaces_type()
             });
         m_print->throw_if_canceled();
         BOOST_LOG_TRIVIAL(debug) << "Detecting solid surfaces for region " << region_id << " - clipping in parallel - end";
-    } // for each this->print->region_count
+    });
 
     // Mark the object to have the region slices classified (typed, which also means they are split based on whether they are supported, bridging, top layers etc.)
     m_typed_slices = true;
@@ -2180,8 +2196,10 @@ void PrintObject::process_external_surfaces()
 	    BOOST_LOG_TRIVIAL(debug) << "Collecting surfaces covered with extrusions in parallel - end";
 	}
 
-	for (size_t region_id = 0; region_id < this->num_printing_regions(); ++region_id) {
-        BOOST_LOG_TRIVIAL(debug) << "Processing external surfaces for region " << region_id << " in parallel - start";
+    BOOST_LOG_TRIVIAL(debug) << "Processing external surfaces in parallel - start";
+    // The regions of a layer do not see each other here, and a layer cut through a fine relief takes far longer than the
+    // others, so the regions run next to each other instead of one after another, each still over all layers.
+    tbb::parallel_for(size_t(0), this->num_printing_regions(), [this, &surfaces_covered](size_t region_id) {
         tbb::parallel_for(
             tbb::blocked_range<size_t>(0, m_layers.size()),
             [this, &surfaces_covered, region_id](const tbb::blocked_range<size_t>& range) {
@@ -2196,9 +2214,9 @@ void PrintObject::process_external_surfaces()
                 }
             }
         );
+    });
     m_print->throw_if_canceled();
-        BOOST_LOG_TRIVIAL(debug) << "Processing external surfaces for region " << region_id << " in parallel - end";
-    }
+    BOOST_LOG_TRIVIAL(debug) << "Processing external surfaces in parallel - end";
 }
 
 void PrintObject::discover_vertical_shells()
@@ -2237,10 +2255,10 @@ void PrintObject::discover_vertical_shells()
             // The "ensure vertical wall thickness" feature is not applicable to any of the regions. Quit.
             return;
         BOOST_LOG_TRIVIAL(debug) << "Discovering vertical shells in parallel - start : cache top / bottom";
-        //FIXME Improve the heuristics for a grain size.
-        size_t grain_size = std::max(num_layers / 16, size_t(1));
+        // One layer per task: on a layer cut through a fine relief the unions below take far longer than elsewhere, and a
+        // few such layers next to each other must not end up in one task.
         tbb::parallel_for(
-            tbb::blocked_range<size_t>(0, num_layers, grain_size),
+            tbb::blocked_range<size_t>(0, num_layers, 1),
             [this, &cache_top_botom_regions](const tbb::blocked_range<size_t>& range) {
                 const std::initializer_list<SurfaceType> surfaces_bottom { stBottom, stBottomBridge };
                 const size_t num_regions = this->num_printing_regions();
@@ -2248,26 +2266,38 @@ void PrintObject::discover_vertical_shells()
                     m_print->throw_if_canceled();
                     const Layer                      &layer = *m_layers[idx_layer];
                     DiscoverVerticalShellsCacheEntry &cache = cache_top_botom_regions[idx_layer];
-                    // Simulate single set of perimeters over all merged regions.
-                    float                             perimeter_offset = 0.f;
-                    float                             perimeter_min_spacing = FLT_MAX;
+                    const auto top_bottom_expansion = [&layer](size_t region_id) {
+                        return float(layer.m_regions[region_id]->flow(frSolidInfill).scaled_spacing()) * top_bottom_expansion_coeff;
+                    };
 #ifdef SLIC3R_DEBUG_SLICE_PROCESSING
                     static size_t debug_idx = 0;
                     ++ debug_idx;
 #endif /* SLIC3R_DEBUG_SLICE_PROCESSING */
-                    for (size_t region_id = 0; region_id < num_regions; ++ region_id) {
-                        LayerRegion &layerm               = *layer.m_regions[region_id];
-                        float        top_bottom_expansion = float(layerm.flow(frSolidInfill).scaled_spacing()) * top_bottom_expansion_coeff;
-                        // Top surfaces.
-                        append(cache.top_surfaces, offset(layerm.slices.filter_by_type(stTop), top_bottom_expansion));
+                    // The top surfaces, the bottom surfaces and the holes are independent of each other.
+                    tbb::parallel_invoke(
+                        [&]() {
+                            for (size_t region_id = 0; region_id < num_regions; ++ region_id)
+                                append(cache.top_surfaces, offset(layer.m_regions[region_id]->slices.filter_by_type(stTop), top_bottom_expansion(region_id)));
 //                              append(cache.top_surfaces, offset(layerm.fill_surfaces.filter_by_type(stTop), top_bottom_expansion));
-                        // Bottom surfaces.
-                        append(cache.bottom_surfaces, offset(layerm.slices.filter_by_types(surfaces_bottom), top_bottom_expansion));
+                            // Save some computing time by reducing the number of polygons.
+                            cache.top_surfaces = union_(cache.top_surfaces);
+                        },
+                        [&]() {
+                            for (size_t region_id = 0; region_id < num_regions; ++ region_id)
+                                append(cache.bottom_surfaces, offset(layer.m_regions[region_id]->slices.filter_by_types(surfaces_bottom), top_bottom_expansion(region_id)));
 //                              append(cache.bottom_surfaces, offset(layerm.fill_surfaces.filter_by_types(surfaces_bottom), top_bottom_expansion));
+                            cache.bottom_surfaces = union_(cache.bottom_surfaces);
+                        },
+                        [&]() {
+                            // Simulate single set of perimeters over all merged regions.
+                            float perimeter_offset      = 0.f;
+                            float perimeter_min_spacing = FLT_MAX;
+                            for (size_t region_id = 0; region_id < num_regions; ++ region_id) {
+                                const LayerRegion &layerm = *layer.m_regions[region_id];
                                 // Calculate the maximum perimeter offset as if the slice was extruded with a single extruder only.
                                 // First find the maxium number of perimeters per region slice.
                                 unsigned int perimeters = 0;
-                        for (Surface &s : layerm.slices.surfaces)
+                                for (const Surface &s : layerm.slices.surfaces)
                                     perimeters = std::max<unsigned int>(perimeters, s.extra_perimeters);
                                 perimeters += layerm.region().config().wall_loops.value;
                                 // Then calculate the infill offset.
@@ -2280,9 +2310,6 @@ void PrintObject::discover_vertical_shells()
                                 }
                                 polygons_append(cache.holes, to_polygons(layerm.fill_expolygons));
                             }
-                    // Save some computing time by reducing the number of polygons.
-                    cache.top_surfaces    = union_(cache.top_surfaces);
-                    cache.bottom_surfaces = union_(cache.bottom_surfaces);
                             // For a multi-material print, simulate perimeter / infill split as if only a single extruder has been used for the whole print.
                             if (perimeter_offset > 0.) {
                                 // The layer.lslices are forced to merge by expanding them first.
@@ -2298,106 +2325,32 @@ void PrintObject::discover_vertical_shells()
 #endif /* SLIC3R_DEBUG_SLICE_PROCESSING */
                             }
                             cache.holes = union_(cache.holes);
+                        });
                 }
             });
         m_print->throw_if_canceled();
         BOOST_LOG_TRIVIAL(debug) << "Discovering vertical shells in parallel - end : cache top / bottom";
     }
 
-    for (size_t region_id = 0; region_id < this->num_printing_regions(); ++ region_id) {
-        const PrintRegion &region = this->printing_region(region_id);
-        if (region.config().ensure_vertical_shell_thickness.value != evstAll )
-            // This region will be handled by discover_horizontal_shells().
-            continue;
-
-        //FIXME Improve the heuristics for a grain size.
-        size_t grain_size = std::max(num_layers / 16, size_t(1));
-
-        if (! top_bottom_surfaces_all_regions) {
-            // This is either a single material print, or a multi-material print and interface_shells are enabled, meaning that the vertical shell thickness
-            // is calculated over a single material.
-            BOOST_LOG_TRIVIAL(debug) << "Discovering vertical shells for region " << region_id << " in parallel - start : cache top / bottom";
-            tbb::parallel_for(
-                tbb::blocked_range<size_t>(0, num_layers, grain_size),
-                [this, region_id, &cache_top_botom_regions](const tbb::blocked_range<size_t>& range) {
-                    const std::initializer_list<SurfaceType> surfaces_bottom { stBottom, stBottomBridge };
-                    for (size_t idx_layer = range.begin(); idx_layer < range.end(); ++ idx_layer) {
-                        m_print->throw_if_canceled();
-                        Layer       &layer                = *m_layers[idx_layer];
-                        LayerRegion &layerm               = *layer.m_regions[region_id];
-                        float        top_bottom_expansion = float(layerm.flow(frSolidInfill).scaled_spacing()) * top_bottom_expansion_coeff;
-                        // Top surfaces.
-                        auto &cache = cache_top_botom_regions[idx_layer];
-                        cache.top_surfaces = offset(layerm.slices.filter_by_type(stTop), top_bottom_expansion);
-//                        append(cache.top_surfaces, offset(layerm.fill_surfaces.filter_by_type(stTop), top_bottom_expansion));
-                        // Bottom surfaces.
-                        cache.bottom_surfaces = offset(layerm.slices.filter_by_types(surfaces_bottom), top_bottom_expansion);
-//                        append(cache.bottom_surfaces, offset(layerm.fill_surfaces.filter_by_types(surfaces_bottom), top_bottom_expansion));
-                        // Holes over all regions. Only collect them once, they are valid for all region_id iterations.
-                        if (cache.holes.empty()) {
-                            for (size_t region_id = 0; region_id < layer.regions().size(); ++ region_id)
-                                polygons_append(cache.holes, to_polygons(layer.regions()[region_id]->fill_expolygons));
-                        }
-                    }
-                });
-            m_print->throw_if_canceled();
-            BOOST_LOG_TRIVIAL(debug) << "Discovering vertical shells for region " << region_id << " in parallel - end : cache top / bottom";
-        }
-
-        BOOST_LOG_TRIVIAL(debug) << "Discovering vertical shells for region " << region_id << " in parallel - start : ensure vertical wall thickness";
-        grain_size = 1;
-        tbb::parallel_for(
-            tbb::blocked_range<size_t>(0, num_layers, grain_size),
-            [this, region_id, &cache_top_botom_regions]
-            (const tbb::blocked_range<size_t>& range) {
-                // printf("discover_vertical_shells from %d to %d\n", range.begin(), range.end());
-                for (size_t idx_layer = range.begin(); idx_layer < range.end(); ++ idx_layer) {
-                    m_print->throw_if_canceled();
-#ifdef SLIC3R_DEBUG_SLICE_PROCESSING
-        			static size_t debug_idx = 0;
-        			++ debug_idx;
-#endif /* SLIC3R_DEBUG_SLICE_PROCESSING */
-
-                    Layer       	        *layer          = m_layers[idx_layer];
-                    LayerRegion 	        *layerm         = layer->m_regions[region_id];
-                    const PrintRegionConfig &region_config  = layerm->region().config();
-
-#ifdef SLIC3R_DEBUG_SLICE_PROCESSING
-                    layerm->export_region_slices_to_svg_debug("3_discover_vertical_shells-initial");
-                    layerm->export_region_fill_surfaces_to_svg_debug("3_discover_vertical_shells-initial");
-#endif /* SLIC3R_DEBUG_SLICE_PROCESSING */
-
-                    Flow         solid_infill_flow   = layerm->flow(frSolidInfill);
-                    coord_t      infill_line_spacing = solid_infill_flow.scaled_spacing();
-                    // Find a union of perimeters below / above this surface to guarantee a minimum shell thickness.
+    // With one top/bottom cache for all regions, the shell and hole accumulation in the loop below depends on nothing
+    // region-specific but the shell settings and the external perimeter spacing, so a region sharing them with an earlier
+    // one reuses its result instead of repeating it: that accumulation is a union over several layers of top/bottom
+    // surfaces, and a multi-material print has a region per filament.
+    using AccumulationKey = std::array<double, 5>;
+    struct ShellAccumulation
+    {
+        AccumulationKey key;
         Polygons        shell;
         Polygons        holes;
-#ifdef SLIC3R_DEBUG_SLICE_PROCESSING
-                    ExPolygons shell_ex;
-#endif /* SLIC3R_DEBUG_SLICE_PROCESSING */
-                    float min_perimeter_infill_spacing = float(infill_line_spacing) * 1.05f;
-#if 0
-// #ifdef SLIC3R_DEBUG_SLICE_PROCESSING
-                    {
-        				Slic3r::SVG svg_cummulative(debug_out_path("discover_vertical_shells-perimeters-before-union-run%d.svg", debug_idx), this->bounding_box());
-                        for (int n = (int)idx_layer - n_extra_bottom_layers; n <= (int)idx_layer + n_extra_top_layers; ++ n) {
-                            if (n < 0 || n >= (int)m_layers.size())
-                                continue;
-                            ExPolygons &expolys = m_layers[n]->perimeter_expolygons;
-                            for (size_t i = 0; i < expolys.size(); ++ i) {
-        						Slic3r::SVG svg(debug_out_path("discover_vertical_shells-perimeters-before-union-run%d-layer%d-expoly%d.svg", debug_idx, n, i), get_extents(expolys[i]));
-                                svg.draw(expolys[i]);
-                                svg.draw_outline(expolys[i].contour, "black", scale_(0.05));
-                                svg.draw_outline(expolys[i].holes, "blue", scale_(0.05));
-                                svg.Close();
-
-                                svg_cummulative.draw(expolys[i]);
-                                svg_cummulative.draw_outline(expolys[i].contour, "black", scale_(0.05));
-                                svg_cummulative.draw_outline(expolys[i].holes, "blue", scale_(0.05));
-                            }
-                        }
-                    }
-#endif /* SLIC3R_DEBUG_SLICE_PROCESSING */
+    };
+    const auto accumulation_key = [](const PrintRegionConfig &region_config, const LayerRegion *layerm) {
+        return AccumulationKey{ double(region_config.top_shell_layers.value), region_config.top_shell_thickness.value,
+                                double(region_config.bottom_shell_layers.value), region_config.bottom_shell_thickness.value,
+                                double(layerm->flow(frExternalPerimeter).scaled_spacing()) };
+    };
+    const auto accumulate_shell = [this, &cache_top_botom_regions](size_t idx_layer, const PrintRegionConfig &region_config,
+                                                                   const LayerRegion *layerm, Polygons &shell, Polygons &holes) {
+        const Layer *layer = m_layers[idx_layer];
     			        polygons_append(holes, cache_top_botom_regions[idx_layer].holes);
                         auto combine_holes = [&holes](const Polygons &holes2) {
                             if (holes.empty() || holes2.empty())
@@ -2472,6 +2425,141 @@ void PrintObject::discover_vertical_shells()
                                     (i > ibottom || bottom_z - m_layers[i]->print_z < region_config.bottom_shell_thickness - EPSILON))
                                     combine_holes(cache_top_botom_regions[i].holes);
     	                }
+    };
+    std::vector<std::vector<ShellAccumulation>> shell_accumulations(top_bottom_surfaces_all_regions ? num_layers : 0);
+    if (! shell_accumulations.empty()) {
+        // Every (layer, key) pair is accumulated once, before the regions, so that nothing in the loop below is shared
+        // between them and they can run next to each other.
+        std::vector<std::array<size_t, 3>> todo; // layer, its slot, a region holding the key
+        for (size_t idx_layer = 0; idx_layer < num_layers; ++ idx_layer) {
+            std::vector<ShellAccumulation> &accumulations = shell_accumulations[idx_layer];
+            for (size_t region_id = 0; region_id < this->num_printing_regions(); ++ region_id) {
+                if (this->printing_region(region_id).config().ensure_vertical_shell_thickness.value != evstAll)
+                    continue;
+                const LayerRegion    *layerm = m_layers[idx_layer]->m_regions[region_id];
+                const AccumulationKey key    = accumulation_key(layerm->region().config(), layerm);
+                if (std::none_of(accumulations.begin(), accumulations.end(), [&key](const ShellAccumulation &a) { return a.key == key; })) {
+                    todo.push_back({ idx_layer, accumulations.size(), region_id });
+                    accumulations.push_back({ key, {}, {} });
+                }
+            }
+        }
+        tbb::parallel_for(size_t(0), todo.size(), [this, &todo, &shell_accumulations, &accumulate_shell](size_t i) {
+            m_print->throw_if_canceled();
+            const LayerRegion *layerm = m_layers[todo[i][0]]->m_regions[todo[i][2]];
+            ShellAccumulation &out    = shell_accumulations[todo[i][0]][todo[i][1]];
+            accumulate_shell(todo[i][0], layerm->region().config(), layerm, out.shell, out.holes);
+        });
+        m_print->throw_if_canceled();
+    }
+
+    const auto process_region = [&](size_t region_id) {
+        const PrintRegion &region = this->printing_region(region_id);
+        if (region.config().ensure_vertical_shell_thickness.value != evstAll )
+            // This region will be handled by discover_horizontal_shells().
+            return;
+
+        //FIXME Improve the heuristics for a grain size.
+        size_t grain_size = std::max(num_layers / 16, size_t(1));
+
+        if (! top_bottom_surfaces_all_regions) {
+            // This is either a single material print, or a multi-material print and interface_shells are enabled, meaning that the vertical shell thickness
+            // is calculated over a single material.
+            BOOST_LOG_TRIVIAL(debug) << "Discovering vertical shells for region " << region_id << " in parallel - start : cache top / bottom";
+            tbb::parallel_for(
+                tbb::blocked_range<size_t>(0, num_layers, grain_size),
+                [this, region_id, &cache_top_botom_regions](const tbb::blocked_range<size_t>& range) {
+                    const std::initializer_list<SurfaceType> surfaces_bottom { stBottom, stBottomBridge };
+                    for (size_t idx_layer = range.begin(); idx_layer < range.end(); ++ idx_layer) {
+                        m_print->throw_if_canceled();
+                        Layer       &layer                = *m_layers[idx_layer];
+                        LayerRegion &layerm               = *layer.m_regions[region_id];
+                        float        top_bottom_expansion = float(layerm.flow(frSolidInfill).scaled_spacing()) * top_bottom_expansion_coeff;
+                        // Top surfaces.
+                        auto &cache = cache_top_botom_regions[idx_layer];
+                        cache.top_surfaces = offset(layerm.slices.filter_by_type(stTop), top_bottom_expansion);
+//                        append(cache.top_surfaces, offset(layerm.fill_surfaces.filter_by_type(stTop), top_bottom_expansion));
+                        // Bottom surfaces.
+                        cache.bottom_surfaces = offset(layerm.slices.filter_by_types(surfaces_bottom), top_bottom_expansion);
+//                        append(cache.bottom_surfaces, offset(layerm.fill_surfaces.filter_by_types(surfaces_bottom), top_bottom_expansion));
+                        // Holes over all regions. Only collect them once, they are valid for all region_id iterations.
+                        if (cache.holes.empty()) {
+                            for (size_t region_id = 0; region_id < layer.regions().size(); ++ region_id)
+                                polygons_append(cache.holes, to_polygons(layer.regions()[region_id]->fill_expolygons));
+                        }
+                    }
+                });
+            m_print->throw_if_canceled();
+            BOOST_LOG_TRIVIAL(debug) << "Discovering vertical shells for region " << region_id << " in parallel - end : cache top / bottom";
+        }
+
+        BOOST_LOG_TRIVIAL(debug) << "Discovering vertical shells for region " << region_id << " in parallel - start : ensure vertical wall thickness";
+        grain_size = 1;
+        tbb::parallel_for(
+            tbb::blocked_range<size_t>(0, num_layers, grain_size),
+            [this, region_id, &shell_accumulations, &accumulation_key, &accumulate_shell]
+            (const tbb::blocked_range<size_t>& range) {
+                // printf("discover_vertical_shells from %d to %d\n", range.begin(), range.end());
+                for (size_t idx_layer = range.begin(); idx_layer < range.end(); ++ idx_layer) {
+                    m_print->throw_if_canceled();
+#ifdef SLIC3R_DEBUG_SLICE_PROCESSING
+        			static size_t debug_idx = 0;
+        			++ debug_idx;
+#endif /* SLIC3R_DEBUG_SLICE_PROCESSING */
+
+                    Layer       	        *layer          = m_layers[idx_layer];
+                    LayerRegion 	        *layerm         = layer->m_regions[region_id];
+                    const PrintRegionConfig &region_config  = layerm->region().config();
+
+#ifdef SLIC3R_DEBUG_SLICE_PROCESSING
+                    layerm->export_region_slices_to_svg_debug("3_discover_vertical_shells-initial");
+                    layerm->export_region_fill_surfaces_to_svg_debug("3_discover_vertical_shells-initial");
+#endif /* SLIC3R_DEBUG_SLICE_PROCESSING */
+
+                    Flow         solid_infill_flow   = layerm->flow(frSolidInfill);
+                    coord_t      infill_line_spacing = solid_infill_flow.scaled_spacing();
+                    // Find a union of perimeters below / above this surface to guarantee a minimum shell thickness.
+                    Polygons shell;
+                    Polygons holes;
+#ifdef SLIC3R_DEBUG_SLICE_PROCESSING
+                    ExPolygons shell_ex;
+#endif /* SLIC3R_DEBUG_SLICE_PROCESSING */
+                    float min_perimeter_infill_spacing = float(infill_line_spacing) * 1.05f;
+#if 0
+// #ifdef SLIC3R_DEBUG_SLICE_PROCESSING
+                    {
+        				Slic3r::SVG svg_cummulative(debug_out_path("discover_vertical_shells-perimeters-before-union-run%d.svg", debug_idx), this->bounding_box());
+                        for (int n = (int)idx_layer - n_extra_bottom_layers; n <= (int)idx_layer + n_extra_top_layers; ++ n) {
+                            if (n < 0 || n >= (int)m_layers.size())
+                                continue;
+                            ExPolygons &expolys = m_layers[n]->perimeter_expolygons;
+                            for (size_t i = 0; i < expolys.size(); ++ i) {
+        						Slic3r::SVG svg(debug_out_path("discover_vertical_shells-perimeters-before-union-run%d-layer%d-expoly%d.svg", debug_idx, n, i), get_extents(expolys[i]));
+                                svg.draw(expolys[i]);
+                                svg.draw_outline(expolys[i].contour, "black", scale_(0.05));
+                                svg.draw_outline(expolys[i].holes, "blue", scale_(0.05));
+                                svg.Close();
+
+                                svg_cummulative.draw(expolys[i]);
+                                svg_cummulative.draw_outline(expolys[i].contour, "black", scale_(0.05));
+                                svg_cummulative.draw_outline(expolys[i].holes, "blue", scale_(0.05));
+                            }
+                        }
+                    }
+#endif /* SLIC3R_DEBUG_SLICE_PROCESSING */
+                    const AccumulationKey key = accumulation_key(region_config, layerm);
+                    const ShellAccumulation *reused = shell_accumulations.empty() ? nullptr :
+                        [&]() -> const ShellAccumulation * {
+                            for (const ShellAccumulation &a : shell_accumulations[idx_layer])
+                                if (a.key == key)
+                                    return &a;
+                            return nullptr;
+                        }();
+                    if (reused != nullptr) {
+                        shell = reused->shell;
+                        holes = reused->holes;
+                    } else
+                        accumulate_shell(idx_layer, region_config, layerm, shell, holes);
 #ifdef SLIC3R_DEBUG_SLICE_PROCESSING
                     {
         				Slic3r::SVG svg(debug_out_path("discover_vertical_shells-perimeters-before-union-%d.svg", debug_idx), get_extents(shell));
@@ -2565,11 +2653,8 @@ void PrintObject::discover_vertical_shells()
                         Polygons object_volume;
                         Polygons internal_volume;
                         {
-                            Polygons shrinked_bottom_slice = idx_layer > 0 ? to_polygons(m_layers[idx_layer - 1]->lslices) : Polygons{};
-                            Polygons shrinked_upper_slice  = (idx_layer + 1) < m_layers.size() ?
-                                                                 to_polygons(m_layers[idx_layer + 1]->lslices) :
-                                                                 Polygons{};
-                            object_volume = intersection(shrinked_bottom_slice, shrinked_upper_slice);
+                            if (idx_layer > 0 && idx_layer + 1 < m_layers.size())
+                                object_volume = to_polygons(intersection_ex_by_piece(m_layers[idx_layer - 1]->lslices, to_polygons(m_layers[idx_layer + 1]->lslices)));
                             internal_volume = closing(polygonsInternal, SCALED_EPSILON);
                         }
 
@@ -2580,15 +2665,34 @@ void PrintObject::discover_vertical_shells()
                         //      the in-model condition is there due to small sloping surfaces, e.g. top of the hull of the benchy
                         //   2. the area does not fully cover an internal polygon
                         //         This is there mainly for a very thin parts, where the solid layers would be missing if the part area is quite small
+                        // Both tests below compare a small piece against the whole layer. Done literally, that is
+                        // quadratic in the number of pieces, which is what a layer split up by colour painting has,
+                        // so each is restricted to the part of the layer near the piece with an identical result:
+                        // object_volume is clipped to the piece's box, and only the internal polygons whose box meets
+                        // the expanded piece take part in the count, since the others pass through the difference
+                        // unchanged and add the same number to both sides of it.
+                        std::vector<BoundingBox> internal_bboxes;
+                        internal_bboxes.reserve(internal_volume.size());
+                        for (const Polygon &poly : internal_volume)
+                            internal_bboxes.emplace_back(get_extents(poly));
                         regularized_shell.erase(std::remove_if(regularized_shell.begin(), regularized_shell.end(),
-                                                               [&internal_volume, &min_perimeter_infill_spacing,
+                                                               [&internal_volume, &internal_bboxes, &min_perimeter_infill_spacing,
                                                                 &object_volume](const ExPolygon &p) {
-                                                                   return (p.area() < min_perimeter_infill_spacing * scaled(1.5) ||
+                                                                   const bool small = p.area() < min_perimeter_infill_spacing * scaled(1.5) ||
                                                                                       (p.area() < min_perimeter_infill_spacing * scaled(8.0) &&
-                                                                            diff(to_polygons(p), object_volume).empty())) &&
-                                                                          diff(internal_volume,
-                                                                               expand(to_polygons(p), min_perimeter_infill_spacing))
-                                                                                  .size() >= internal_volume.size();
+                                                                                       diff(to_polygons(p),
+                                                                                            ClipperUtils::clip_clipper_polygons_with_subject_bbox(
+                                                                                                object_volume, get_extents(p).inflated(SCALED_EPSILON)))
+                                                                                           .empty());
+                                                                   if (!small)
+                                                                       return false;
+                                                                   const Polygons    expanded = expand(to_polygons(p), min_perimeter_infill_spacing);
+                                                                   const BoundingBox bbox     = get_extents(expanded);
+                                                                   Polygons          nearby;
+                                                                   for (size_t i = 0; i < internal_volume.size(); ++i)
+                                                                       if (internal_bboxes[i].overlap(bbox))
+                                                                           nearby.emplace_back(internal_volume[i]);
+                                                                   return diff(nearby, expanded).size() >= nearby.size();
                                                                }),
                                                 regularized_shell.end());
                     }
@@ -2610,8 +2714,9 @@ void PrintObject::discover_vertical_shells()
 #endif /* SLIC3R_DEBUG_SLICE_PROCESSING */
 
                     // Trim the internal & internalvoid by the shell.
-                    Slic3r::ExPolygons new_internal = diff_ex(layerm->fill_surfaces.filter_by_type(stInternal), regularized_shell);
-                    Slic3r::ExPolygons new_internal_void = diff_ex(layerm->fill_surfaces.filter_by_type(stInternalVoid), regularized_shell);
+                    const Polygons     regularized_shell_polygons = to_polygons(regularized_shell);
+                    Slic3r::ExPolygons new_internal = diff_ex_by_piece(to_expolygons(layerm->fill_surfaces.filter_by_type(stInternal)), regularized_shell_polygons);
+                    Slic3r::ExPolygons new_internal_void = diff_ex_by_piece(to_expolygons(layerm->fill_surfaces.filter_by_type(stInternalVoid)), regularized_shell_polygons);
 
 #ifdef SLIC3R_DEBUG_SLICE_PROCESSING
                     {
@@ -2638,7 +2743,15 @@ void PrintObject::discover_vertical_shells()
 			layerm->export_region_fill_surfaces_to_svg_debug("3_discover_vertical_shells-final");
 		}
 #endif /* SLIC3R_DEBUG_SLICE_PROCESSING */
-    } // for each region
+    }; // for each region
+    if (top_bottom_surfaces_all_regions)
+        // Nothing is shared between the regions, and a layer cut through a fine relief takes far longer than the others,
+        // so they run next to each other instead of one after another.
+        tbb::parallel_for(size_t(0), this->num_printing_regions(), process_region);
+    else
+        // Here every region fills the one top/bottom cache with its own surfaces first.
+        for (size_t region_id = 0; region_id < this->num_printing_regions(); ++ region_id)
+            process_region(region_id);
 } // void PrintObject::discover_vertical_shells()
 
 // #define DEBUG_BRIDGE_OVER_INFILL
@@ -3159,6 +3272,16 @@ void PrintObject::bridge_over_infill()
                 vertical_lines[i].b = Point{x, y_max};
             }
 
+            // The vertical lines only span the bridged area's x range, so anchors entirely outside it can never be
+            // hit. Leaving them out gives the same intersections without building a tree over the whole layer's
+            // boundary for every bridge.
+            const coord_t scan_x_min = bb_x.min.x();
+            const coord_t scan_x_max = bb_x.min.x() + coord_t(n_vlines) * scan_spacing;
+            anchors.erase(std::remove_if(anchors.begin(), anchors.end(),
+                                         [scan_x_min, scan_x_max](const Line &l) {
+                                             return std::max(l.a.x(), l.b.x()) < scan_x_min || std::min(l.a.x(), l.b.x()) > scan_x_max;
+                                         }),
+                          anchors.end());
             auto anchors_and_walls_tree = AABBTreeLines::LinesDistancer<Line>{std::move(anchors)};
             auto bridged_area_tree      = AABBTreeLines::LinesDistancer<Line>{to_lines(bridged_area)};
 
@@ -3403,28 +3526,63 @@ void PrintObject::bridge_over_infill()
 
                 std::vector<CandidateSurface> expanded_surfaces;
                 expanded_surfaces.reserve(surfaces_by_layer[lidx].size());
+                // The expanded fill boundary depends only on the bridging flow, and total_fill_area is not
+                // modified below, so build it once per spacing rather than once per candidate. A layer split
+                // into many candidates (e.g. by colour painting) otherwise repeats a layer-wide offset for each.
+                std::map<coord_t, Polylines> boundary_by_spacing;
+                // expansion_area is a clean, non-overlapping set, so uniting it with a bridge or cutting a bridge
+                // out of it only changes the polygons near that bridge. The rest are passed through untouched
+                // instead of being fed to ClipperLib with the whole layer again for every candidate.
+                // Not `near`/`far`: the Windows headers still define those as macros, and they expand to
+                // nothing, which turns the declaration below into an empty one.
+                const auto split_near = [](const Polygons &polys, const BoundingBox &bbox, Polygons &rest) {
+                    Polygons nearby;
+                    for (const Polygon &p : polys)
+                        (get_extents(p).overlap(bbox) ? nearby : rest).emplace_back(p);
+                    return nearby;
+                };
                 for (const CandidateSurface &candidate : surfaces_by_layer[lidx]) {
                     const auto &region_config = candidate.region->region().config();
                     const bool turning_pattern = region_config.sparse_infill_pattern == ipHilbertCurve ||
                                                  region_config.sparse_infill_pattern == ipOctagramSpiral;
                     const Flow &flow              = candidate.region->bridging_flow(frSolidInfill, true);
                     Polygons    area_to_be_bridge = expand(candidate.new_polys, flow.scaled_spacing());
-                    area_to_be_bridge             = intersection(area_to_be_bridge, deep_infill_area);
+                    // deep_infill_area and internal_unsupported_area cover the whole layer; only their part under
+                    // this candidate can change the results, so they are clipped to its box first.
+                    if (!area_to_be_bridge.empty())
+                        area_to_be_bridge = intersection(area_to_be_bridge,
+                                                         ClipperUtils::clip_clipper_polygons_with_subject_bbox(
+                                                             deep_infill_area, get_extents(area_to_be_bridge).inflated(SCALED_EPSILON)));
 
                     area_to_be_bridge.erase(std::remove_if(area_to_be_bridge.begin(), area_to_be_bridge.end(),
-                                                           [internal_unsupported_area](const Polygon &p) {
-                                                               return intersection({p}, internal_unsupported_area).empty();
+                                                           [&internal_unsupported_area](const Polygon &p) {
+                                                               return intersection({p}, ClipperUtils::clip_clipper_polygons_with_subject_bbox(
+                                                                                            internal_unsupported_area,
+                                                                                            get_extents(p).inflated(SCALED_EPSILON)))
+                                                                   .empty();
                                                            }),
                                             area_to_be_bridge.end());
 
-                    Polygons limiting_area = union_(area_to_be_bridge, expansion_area);
-
                     if (area_to_be_bridge.empty())
                         continue;
 
-                    Polylines boundary_plines = to_polylines(expand(total_fill_area, 1.3 * flow.scaled_spacing()));
+                    Polygons       limiting_area;
+                    const Polygons near_expansion = split_near(expansion_area, get_extents(area_to_be_bridge).inflated(SCALED_EPSILON),
+                                                               limiting_area);
+                    append(limiting_area, union_(area_to_be_bridge, near_expansion));
+
+                    auto boundary_it = boundary_by_spacing.find(flow.scaled_spacing());
+                    if (boundary_it == boundary_by_spacing.end())
+                        boundary_it = boundary_by_spacing
+                                          .emplace(flow.scaled_spacing(), to_polylines(expand(total_fill_area, 1.3 * flow.scaled_spacing())))
+                                          .first;
+                    Polylines boundary_plines = boundary_it->second;
                     {
-                        Polylines limiting_plines = to_polylines(expand(limiting_area, 0.3*flow.spacing()));
+                        // No offset here: flow.spacing() is in mm, so the expand(limiting_area, 0.3 * flow.spacing())
+                        // this used to be moved the outline by 0.135 scaled units - nothing beyond rounding - while
+                        // costing a whole-layer ClipperLib pass for every candidate. limiting_area is already a clean
+                        // union, so its own outline is the same boundary.
+                        Polylines limiting_plines = to_polylines(limiting_area);
                         boundary_plines.insert(boundary_plines.end(), limiting_plines.begin(), limiting_plines.end());
                     }
 
@@ -3499,8 +3657,11 @@ void PrintObject::bridge_over_infill()
                     {
                         bool     reconstruct       = false;
                         Polygons          tmp_expanded_area = expand(bridging_area, 3.0 * flow.scaled_spacing());
+                        const BoundingBox tmp_expanded_bbox = get_extents(tmp_expanded_area);
                         for (const CandidateSurface &s : expanded_surfaces) {
-                            if (!intersection(s.new_polys, tmp_expanded_area).empty()) {
+                            // Surfaces whose boxes miss each other cannot intersect, which is most pairs on a busy layer.
+                            if (get_extents(s.new_polys).overlap(tmp_expanded_bbox) &&
+                                !intersection(s.new_polys, tmp_expanded_area).empty()) {
                                 bridging_angle = s.bridge_angle;
                                 reconstruct    = true;
                                 break;
@@ -3524,10 +3685,20 @@ void PrintObject::bridge_over_infill()
                         bridging_area = union_(bridging_area, construct_anchored_polygon(bridging_area, to_lines(boundary_plines), flow,
                                                                                        bridging_angle, scan_spacing, true));
                     }
-                    bridging_area          = intersection(bridging_area, limiting_area);
-                    bridging_area          = intersection(bridging_area, total_fill_area);
-                    bridging_area          = diff(bridging_area, total_top_area);
-                    expansion_area         = diff(expansion_area, bridging_area);
+                    // Each of these meets one bridge with the whole layer, so the layer side is first cut down to the
+                    // bridge's box (and expansion_area split as above); the result is the same.
+                    if (!bridging_area.empty()) {
+                        const BoundingBox bridging_bbox = get_extents(bridging_area).inflated(SCALED_EPSILON);
+                        bridging_area = intersection(bridging_area, ClipperUtils::clip_clipper_polygons_with_subject_bbox(limiting_area, bridging_bbox));
+                        bridging_area = intersection(bridging_area, ClipperUtils::clip_clipper_polygons_with_subject_bbox(total_fill_area, bridging_bbox));
+                        bridging_area = diff(bridging_area, ClipperUtils::clip_clipper_polygons_with_subject_bbox(total_top_area, bridging_bbox));
+                    }
+                    if (!bridging_area.empty()) {
+                        Polygons       kept;
+                        const Polygons cut = split_near(expansion_area, get_extents(bridging_area).inflated(SCALED_EPSILON), kept);
+                        append(kept, diff(cut, bridging_area));
+                        expansion_area = std::move(kept);
+                    }
 
 #ifdef DEBUG_BRIDGE_OVER_INFILL
                     debug_draw(std::to_string(lidx) + "_" + std::to_string(cluster_idx) + "_" + std::to_string(job_idx) + "_" + "_expanded_bridging" +  std::to_string(r),

src/libslic3r/Support/SupportMaterial.cpp

@@ -929,9 +929,9 @@ public:
                 ::fread(&y, sizeof(coord_t), 1, file);
                 poly.points.emplace_back(Point(x * scale, y * scale));
             }
+            printf("Polygon %d, area: %lf\n", i, area(poly.points));
             if (which == -1 || which == i)
                 m_support_polygons_deserialized.emplace_back(std::move(poly));
-            printf("Polygon %d, area: %lf\n", i, area(poly.points));
         }
         ::fread(&n_polygons, 4, 1, file);
         m_trimming_polygons_deserialized.reserve(n_polygons);

src/libslic3r/Support/TreeSupport.cpp

@@ -854,11 +854,41 @@ void TreeSupport::detect_overhangs(bool check_support_necessity/* = false*/)
                 if (is_auto(stype) && config_detect_sharp_tails)
                 {
                     // BBS detect sharp tail
+                    // Each island is tested only against the lower islands whose box meets its own: overlaps() tries every
+                    // pair, which on a layer cut through a fine relief (thousands of islands above thousands) never ends.
+                    std::vector<BoundingBox> lower_bboxes;
+                    lower_bboxes.reserve(lower_polys.size());
+                    for (const ExPolygon &lower : lower_polys)
+                        lower_bboxes.emplace_back(get_extents(lower));
                     for (const ExPolygon& expoly : curr_polys) {
                         bool  is_sharp_tail = false;
                         // 1. nothing below
                         // this is a sharp tail region if it's floating and non-ignorable
-                        if (!overlaps(offset_ex(expoly, 0.1 * extrusion_width_scaled), lower_polys)) {
+                        const ExPolygons  expanded = offset_ex(expoly, 0.1 * extrusion_width_scaled);
+                        const BoundingBox bbox     = get_extents(expanded);
+                        ExPolygons        lower_nearby;
+                        for (size_t i = 0; i < lower_polys.size(); ++i)
+                            if (lower_bboxes[i].overlap(bbox))
+                                lower_nearby.emplace_back(lower_polys[i]);
+                        // As overlaps(expanded, lower_nearby), with each lower island cut to the island's box first: below
+                        // a fine relief the lower layer is a few islands with thousands of holes, whose whole boundary
+                        // was otherwise intersected again for every island above.
+                        const auto overlaps_nearby = [&]() {
+                            for (const ExPolygon &a : expanded) {
+                                if (a.empty())
+                                    continue;
+                                const BoundingBox a_bbox = get_extents(a);
+                                for (const ExPolygon &b : lower_nearby) {
+                                    if (b.empty() || !get_extents(b).overlap(a_bbox))
+                                        continue;
+                                    const Polygons b_near = ClipperUtils::clip_clipper_polygons_with_subject_bbox(b, a_bbox.inflated(SCALED_EPSILON));
+                                    if (!intersection_pl(to_polylines(b_near), a).empty() || b.contains(a.contour.points.front()))
+                                        return true;
+                                }
+                            }
+                            return false;
+                        };
+                        if (!overlaps_nearby()) {
                             is_sharp_tail = !offset_ex(expoly, -0.1 * extrusion_width_scaled).empty();
                         }
 

src/libslic3r/Surface.hpp

@@ -61,7 +61,7 @@ public:
             thickness(other.thickness), thickness_layers(other.thickness_layers),
             bridge_angle(other.bridge_angle), extra_perimeters(other.extra_perimeters)
         {};
-    Surface(Surface &&rhs)
+    Surface(Surface &&rhs) noexcept
         : surface_type(rhs.surface_type), expolygon(std::move(rhs.expolygon)),
             thickness(rhs.thickness), thickness_layers(rhs.thickness_layers),
             bridge_angle(rhs.bridge_angle), extra_perimeters(rhs.extra_perimeters)
@@ -87,7 +87,7 @@ public:
         return *this;
     }
 
-    Surface& operator=(Surface &&rhs)
+    Surface& operator=(Surface &&rhs) noexcept
     {
         surface_type     = rhs.surface_type;
         expolygon        = std::move(rhs.expolygon);

src/libslic3r/libslic3r.h

@@ -162,10 +162,10 @@ inline void append(std::vector<T, Alloc> &dest, std::vector<T, Alloc> &&src)
 {
     if (dest.empty())
         dest = std::move(src);
-    else {
-        dest.reserve(dest.size() + src.size());
-        std::move(std::begin(src), std::end(src), std::back_inserter(dest));
-    }
+    else
+        // insert() grows the capacity geometrically; reserving exactly the new size reallocated on every call, which
+        // made appending piece by piece quadratic.
+        dest.insert(dest.end(), std::make_move_iterator(src.begin()), std::make_move_iterator(src.end()));
     src.clear();
     src.shrink_to_fit();
 }

tests/libslic3r/CMakeLists.txt

@@ -28,6 +28,7 @@ add_executable(${_TEST_NAME}_tests
     test_filament_mixer.cpp
     test_fill_plane_path.cpp
     test_geometry.cpp
+    test_kdtree.cpp
     test_multimaterial_segmentation.cpp
     test_placeholder_parser.cpp
     test_polygon.cpp

tests/libslic3r/test_clipper_utils.cpp

@@ -299,3 +299,47 @@ TEST_CASE("Traversing Clipper PolyTree", "[ClipperUtils]") {
         REQUIRE(count_polys(output) == reference.size());
     }
 }
+
+TEST_CASE("Tiled diff and intersection cover the same area as the plain calls", "[ClipperUtils]") {
+    // A grid of disjoint framed squares, enough of them to be split into several tiles.
+    const int  n    = 40;
+    const coord_t cell = scaled<coord_t>(2.), side = scaled<coord_t>(1.5), frame = scaled<coord_t>(0.3);
+    ExPolygons subject;
+    for (int y = 0; y < n; ++ y)
+        for (int x = 0; x < n; ++ x) {
+            const Point o(x * cell, y * cell);
+            ExPolygon square(Polygon({ o, o + Point(side, 0), o + Point(side, side), o + Point(0, side) }));
+            Polygon hole({ o + Point(frame, frame), o + Point(frame, side - frame), o + Point(side - frame, side - frame), o + Point(side - frame, frame) });
+            square.holes.emplace_back(std::move(hole));
+            subject.emplace_back(std::move(square));
+        }
+    // Clip polygons crossing many squares, one of them large with holes of its own.
+    Polygons clip;
+    const coord_t span = n * cell;
+    for (int i = 0; i < 8; ++ i) {
+        const coord_t y0 = coord_t(i) * span / 8, y1 = y0 + scaled<coord_t>(0.9);
+        clip.emplace_back(Polygon({ Point(- cell, y0), Point(span, y0 + cell * 3), Point(span, y1 + cell * 3), Point(- cell, y1) }));
+    }
+    ExPolygon big(Polygon({ Point(span / 4, span / 4), Point(3 * span / 4, span / 4), Point(3 * span / 4, 3 * span / 4), Point(span / 4, 3 * span / 4) }));
+    for (int i = 0; i < 4; ++ i) {
+        const Point o(span / 4 + scaled<coord_t>(3.1) + i * scaled<coord_t>(9.7), span / 4 + scaled<coord_t>(5.3));
+        big.holes.emplace_back(Polygon({ o, o + Point(0, scaled<coord_t>(20.)), o + Point(scaled<coord_t>(5.), scaled<coord_t>(20.)), o + Point(scaled<coord_t>(5.), 0) }));
+    }
+    polygons_append(clip, to_polygons(big));
+
+    const auto xor_area = [](const ExPolygons &a, const ExPolygons &b) { return area(diff_ex(a, b)) + area(diff_ex(b, a)); };
+    const ApplySafetyOffset safety = GENERATE(ApplySafetyOffset::No, ApplySafetyOffset::Yes);
+    const double tolerance = double(scaled<coord_t>(0.001)) * double(span);
+
+    const ExPolygons diff_plain = diff_ex(subject, clip, safety);
+    const ExPolygons diff_tiled = diff_ex_by_piece(subject, clip, safety);
+    REQUIRE(area(diff_plain) > 0.);
+    CHECK_THAT(area(diff_tiled), Catch::Matchers::WithinRel(area(diff_plain), 1e-9));
+    CHECK(xor_area(diff_tiled, diff_plain) < tolerance);
+
+    const ExPolygons intersection_plain = intersection_ex(subject, clip, safety);
+    const ExPolygons intersection_tiled = intersection_ex_by_piece(subject, clip, safety);
+    REQUIRE(area(intersection_plain) > 0.);
+    CHECK_THAT(area(intersection_tiled), Catch::Matchers::WithinRel(area(intersection_plain), 1e-9));
+    CHECK(xor_area(intersection_tiled, intersection_plain) < tolerance);
+}

tests/libslic3r/test_kdtree.cpp

@@ -0,0 +1,67 @@
+#include <catch2/catch_all.hpp>
+
+#include <numeric>
+#include <random>
+#include <vector>
+
+#include "libslic3r/KDTreeIndirect.hpp"
+#include "libslic3r/Point.hpp"
+
+using namespace Slic3r;
+
+TEST_CASE("Visiting the nearby points gives what collecting them gives", "[KDTree]") {
+    std::mt19937                          rng(19937);
+    std::uniform_real_distribution<float> coord(-50.f, 50.f);
+    // Points in a box, so that a radius search returns anything from none of them to all of them.
+    std::vector<Vec3f> points(2000);
+    for (Vec3f &p : points)
+        p = Vec3f(coord(rng), coord(rng), coord(rng));
+
+    auto coordinate = [&points](size_t idx, size_t dimension) { return points[idx](int(dimension)); };
+    KDTreeIndirect<3, float, decltype(coordinate)> tree(coordinate);
+    std::vector<size_t> indices(points.size());
+    std::iota(indices.begin(), indices.end(), 0);
+    tree.build(indices);
+
+    const float radius = GENERATE(0.5f, 5.f, 25.f, 200.f);
+    for (int i = 0; i < 20; ++ i) {
+        const Vec3f center(coord(rng), coord(rng), coord(rng));
+
+        const std::vector<size_t> collected = find_nearby_points(tree, center, radius);
+        std::vector<size_t>       visited;
+        visit_nearby_points(tree, center, radius, [&visited](size_t idx) { visited.emplace_back(idx); });
+
+        // Same points, and in the same order: a caller that keeps the first of several equally good ones
+        // must get the same answer either way.
+        REQUIRE(visited == collected);
+    }
+}
+
+TEST_CASE("A radius search returns every point within the radius and no other", "[KDTree]") {
+    std::mt19937                          rng(2024);
+    std::uniform_real_distribution<float> coord(-20.f, 20.f);
+    std::vector<Vec3f> points(500);
+    for (Vec3f &p : points)
+        p = Vec3f(coord(rng), coord(rng), coord(rng));
+
+    auto coordinate = [&points](size_t idx, size_t dimension) { return points[idx](int(dimension)); };
+    KDTreeIndirect<3, float, decltype(coordinate)> tree(coordinate);
+    std::vector<size_t> indices(points.size());
+    std::iota(indices.begin(), indices.end(), 0);
+    tree.build(indices);
+
+    const Vec3f center(1.f, -2.f, 3.f);
+    const float radius = 7.f;
+
+    std::vector<size_t> expected;
+    for (size_t i = 0; i < points.size(); ++ i)
+        if ((points[i] - center).squaredNorm() < radius * radius)
+            expected.emplace_back(i);
+
+    std::vector<size_t> visited;
+    visit_nearby_points(tree, center, radius, [&visited](size_t idx) { visited.emplace_back(idx); });
+    std::sort(visited.begin(), visited.end());
+
+    REQUIRE(! expected.empty());
+    REQUIRE(visited == expected);
+}