Slice fine texture relief without stalling

A colour texture baked at 0.1 mm / 2000k made the top layers thousands
of islands and slicing never finished. Colour segmentation runs per
island, the merge subtracts piece by piece, the support check tests only
nearby islands, and the travel ordering finds crossings through a grid.

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;
 }