Merge pull request #219 from daveshah1/placer_heap

HeAP-based analytical placer

.cirrus/Dockerfile.ubuntu16.04

@@ -8,7 +8,7 @@ RUN set -e -x ;\
     apt-get -y install \
         build-essential autoconf cmake clang bison wget flex gperf \
         libreadline-dev gawk tcl-dev libffi-dev graphviz xdot python3-dev \
-        libboost-all-dev qt5-default git libftdi-dev pkg-config
+        libboost-all-dev qt5-default git libftdi-dev pkg-config libeigen3-dev
 
 RUN set -e -x ;\
     mkdir -p /usr/local/src ;\

CMakeLists.txt

@@ -5,6 +5,8 @@ project(nextpnr)
 option(BUILD_GUI "Build GUI" ON)
 option(BUILD_PYTHON "Build Python Integration" ON)
 option(BUILD_TESTS "Build GUI" OFF)
+option(BUILD_HEAP "Build HeAP analytic placer" ON)
+option(USE_OPENMP "Use OpenMP to accelerate analytic placer" OFF)
 option(COVERAGE "Add code coverage info" OFF)
 option(STATIC_BUILD "Create static build" OFF)
 option(EXTERNAL_CHIPDB "Create build with pre-built chipdb binaries" OFF)
@@ -53,12 +55,16 @@ endforeach()
 
 set(CMAKE_CXX_STANDARD 11)
 if (MSVC)
-set(CMAKE_CONFIGURATION_TYPES "Debug;Release" CACHE STRING "" FORCE)
+    set(CMAKE_CONFIGURATION_TYPES "Debug;Release" CACHE STRING "" FORCE)
-set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /D_DEBUG /W4 /wd4100 /wd4244 /wd4125 /wd4800 /wd4456 /wd4458 /wd4305 /wd4459 /wd4121 /wd4996")
+    set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /D_DEBUG /W4 /wd4100 /wd4244 /wd4125 /wd4800 /wd4456 /wd4458 /wd4305 /wd4459 /wd4121 /wd4996")
-set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /W4 /wd4100 /wd4244 /wd4125 /wd4800 /wd4456 /wd4458 /wd4305 /wd4459 /wd4121 /wd4996 /wd4127")
+    set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /W4 /wd4100 /wd4244 /wd4125 /wd4800 /wd4456 /wd4458 /wd4305 /wd4459 /wd4121 /wd4996 /wd4127")
 else()
-set(CMAKE_CXX_FLAGS_DEBUG "-Wall -fPIC -ggdb -pipe")
+    set(CMAKE_CXX_FLAGS_DEBUG "-Wall -fPIC -ggdb -pipe")
-set(CMAKE_CXX_FLAGS_RELEASE "-Wall -fPIC -O3 -g -pipe")
+    if (USE_OPENMP)
+        set(CMAKE_CXX_FLAGS_RELEASE "-Wall -fPIC -O3 -g -pipe -fopenmp")
+    else()
+        set(CMAKE_CXX_FLAGS_RELEASE "-Wall -fPIC -O3 -g -pipe")
+    endif()
 endif()
 set(CMAKE_DEFIN)
 
@@ -181,6 +187,14 @@ if (BUILD_PYTHON)
 endif()
 
 include_directories(common/ json/ ${Boost_INCLUDE_DIRS} ${PYTHON_INCLUDE_DIRS})
+
+if(BUILD_HEAP)
+    find_package (Eigen3 REQUIRED NO_MODULE)
+    include_directories(${EIGEN3_INCLUDE_DIRS})
+    add_definitions(${EIGEN3_DEFINITIONS})
+    add_definitions(-DWITH_HEAP)
+endif()
+
 aux_source_directory(common/ COMMON_SRC_FILES)
 aux_source_directory(json/ JSON_PARSER_FILES)
 set(COMMON_FILES ${COMMON_SRC_FILES} ${JSON_PARSER_FILES})

README.md

@@ -36,6 +36,7 @@ of the selected architecture:
 - Python 3.5 or later, including development libraries (`python3-dev` for Ubuntu)
   - on Windows make sure to install same version as supported by [vcpkg](https://github.com/Microsoft/vcpkg/blob/master/ports/python3/CONTROL)
 - Boost libraries (`libboost-dev libboost-filesystem-dev libboost-thread-dev libboost-program-options-dev libboost-python-dev libboost-dev` or `libboost-all-dev` for Ubuntu)
+- Eigen3 (`libeigen3-dev` for Ubuntu) is required to build the analytic placer
 - Latest git Yosys is required to synthesise the demo design
 - For building on Windows with MSVC, usage of vcpkg is advised for dependency installation.
   - For 32 bit builds: `vcpkg install boost-filesystem boost-program-options boost-thread boost-python qt5-base`
@@ -119,11 +120,11 @@ Use cmake `-D` options to specify which version of nextpnr you want to build.
 Use `-DARCH=...` to set the architecture. It is a semicolon separated list.
 Use `cmake . -DARCH=all` to build all supported architectures.
 
-The following runs a debug build of the iCE40 architecture without GUI
+The following runs a debug build of the iCE40 architecture without GUI,
-and without Python support and only HX1K support:
+ without Python support, without the HeAP analytic placer and only HX1K support:
 
 ```
-cmake -DARCH=ice40 -DCMAKE_BUILD_TYPE=Debug -DBUILD_PYTHON=OFF -DBUILD_GUI=OFF -DICE40_HX1K_ONLY=1 .
+cmake -DARCH=ice40 -DCMAKE_BUILD_TYPE=Debug -DBUILD_PYTHON=OFF -DBUILD_GUI=OFF -DBUILD_HEAP=OFF -DICE40_HX1K_ONLY=1 .
 make -j$(nproc)
 ```
 
@@ -134,6 +135,9 @@ cmake -DARCH=ice40 -DBUILD_PYTHON=OFF -DBUILD_GUI=OFF -DSTATIC_BUILD=ON .
 make -j$(nproc)
 ```
 
+The HeAP placer's solver can optionally use OpenMP for a speedup on very large designs. Enable this by passing
+`-DUSE_OPENMP=yes` to cmake (compiler support may vary).
+
 You can change the location where nextpnr will be installed (this will usually default to `/usr/local`) by using
 `-DCMAKE_INSTALL_PREFIX=/install/prefix`.
 

common/command.cc

@@ -27,6 +27,7 @@
 #include "pybindings.h"
 #endif
 
+#include <boost/algorithm/string/join.hpp>
 #include <boost/filesystem/convenience.hpp>
 #include <boost/program_options.hpp>
 #include <fstream>
@@ -120,8 +121,18 @@ po::options_description CommandHandler::getGeneralOptions()
     general.add_options()("json", po::value<std::string>(), "JSON design file to ingest");
     general.add_options()("seed", po::value<int>(), "seed value for random number generator");
     general.add_options()("randomize-seed,r", "randomize seed value for random number generator");
+
+    general.add_options()(
+            "placer", po::value<std::string>(),
+            std::string("placer algorithm to use; available: " + boost::algorithm::join(Arch::availablePlacers, ", ") +
+                        "; default: " + Arch::defaultPlacer)
+                    .c_str());
+
     general.add_options()("slack_redist_iter", po::value<int>(), "number of iterations between slack redistribution");
     general.add_options()("cstrweight", po::value<float>(), "placer weighting for relative constraint satisfaction");
+    general.add_options()("starttemp", po::value<float>(), "placer SA start temperature");
+    general.add_options()("placer-budgets", "use budget rather than criticality in placer timing weights");
+
     general.add_options()("pack-only", "pack design only without placement or routing");
 
     general.add_options()("ignore-loops", "ignore combinational loops in timing analysis");
@@ -183,10 +194,27 @@ void CommandHandler::setupContext(Context *ctx)
         settings->set("timing/allowFail", true);
     }
 
+    if (vm.count("placer")) {
+        std::string placer = vm["placer"].as<std::string>();
+        if (std::find(Arch::availablePlacers.begin(), Arch::availablePlacers.end(), placer) ==
+            Arch::availablePlacers.end())
+            log_error("Placer algorithm '%s' is not supported (available options: %s)\n", placer.c_str(),
+                      boost::algorithm::join(Arch::availablePlacers, ", ").c_str());
+        settings->set("placer", placer);
+    } else {
+        settings->set("placer", Arch::defaultPlacer);
+    }
+
     if (vm.count("cstrweight")) {
         settings->set("placer1/constraintWeight", vm["cstrweight"].as<float>());
     }
+    if (vm.count("starttemp")) {
+        settings->set("placer1/startTemp", vm["starttemp"].as<float>());
+    }
 
+    if (vm.count("placer-budgets")) {
+        settings->set("placer1/budgetBased", true);
+    }
     if (vm.count("freq")) {
         auto freq = vm["freq"].as<double>();
         if (freq > 0)

common/nextpnr.cc

@@ -221,6 +221,9 @@ delay_t Context::getNetinfoRouteDelay(const NetInfo *net_info, const PortRef &us
         return 0;
 #endif
 
+    if (net_info->wires.empty())
+        return predictDelay(net_info, user_info);
+
     WireId src_wire = getNetinfoSourceWire(net_info);
     if (src_wire == WireId())
         return 0;
@@ -421,4 +424,25 @@ void BaseCtx::addClock(IdString net, float freq)
     }
 }
 
+void BaseCtx::createRectangularRegion(IdString name, int x0, int y0, int x1, int y1)
+{
+    std::unique_ptr<Region> new_region(new Region());
+    new_region->name = name;
+    new_region->constr_bels = true;
+    new_region->constr_pips = false;
+    new_region->constr_wires = false;
+    for (int x = x0; x <= x1; x++) {
+        for (int y = y0; y <= y1; y++) {
+            for (auto bel : getCtx()->getBelsByTile(x, y))
+                new_region->bels.insert(bel);
+        }
+    }
+    region[name] = std::move(new_region);
+}
+void BaseCtx::addBelToRegion(IdString name, BelId bel) { region[name]->bels.insert(bel); }
+void BaseCtx::constrainCellToRegion(IdString cell, IdString region_name)
+{
+    cells[cell]->region = region[region_name].get();
+}
+
 NEXTPNR_NAMESPACE_END

common/nextpnr.h

@@ -637,6 +637,9 @@ struct BaseCtx
 
     // Intended to simplify Python API
     void addClock(IdString net, float freq);
+    void createRectangularRegion(IdString name, int x0, int y0, int x1, int y1);
+    void addBelToRegion(IdString name, BelId bel);
+    void constrainCellToRegion(IdString cell, IdString region_name);
 };
 
 NEXTPNR_NAMESPACE_END

common/place_common.cc

@@ -304,7 +304,7 @@ class ConstraintLegaliseWorker
     // Set the strength to locked on all cells in chain
     void lockdown_chain(CellInfo *root)
     {
-        root->belStrength = STRENGTH_LOCKED;
+        root->belStrength = STRENGTH_STRONG;
         for (auto child : root->constr_children)
             lockdown_chain(child);
     }
@@ -380,7 +380,7 @@ class ConstraintLegaliseWorker
                                 rippedCells.insert(confl_cell->name);
                             }
                         }
-                        ctx->bindBel(target, ctx->cells.at(cp.first).get(), STRENGTH_LOCKED);
+                        ctx->bindBel(target, ctx->cells.at(cp.first).get(), STRENGTH_STRONG);
                         rippedCells.erase(cp.first);
                     }
                     for (auto cp : solution) {
@@ -529,4 +529,12 @@ int get_constraints_distance(const Context *ctx, const CellInfo *cell)
     return dist;
 }
 
+bool check_cell_bel_region(const CellInfo *cell, BelId bel)
+{
+    if (cell->region != nullptr && cell->region->constr_bels && !cell->region->bels.count(bel))
+        return false;
+    else
+        return true;
+}
+
 NEXTPNR_NAMESPACE_END

common/place_common.h

@@ -49,6 +49,10 @@ bool legalise_relative_constraints(Context *ctx);
 
 // Get the total distance from satisfied constraints for a cell
 int get_constraints_distance(const Context *ctx, const CellInfo *cell);
+
+// Check that a Bel is within the region for a cell
+bool check_cell_bel_region(const CellInfo *cell, BelId bel);
+
 NEXTPNR_NAMESPACE_END
 
 #endif

common/placer1.cc

@@ -24,6 +24,7 @@
 #include "placer1.h"
 #include <algorithm>
 #include <boost/lexical_cast.hpp>
+#include <boost/range/adaptor/reversed.hpp>
 #include <chrono>
 #include <cmath>
 #include <iostream>
@@ -43,10 +44,32 @@
 #include "timing.h"
 #include "util.h"
 
+namespace std {
+template <> struct hash<std::pair<NEXTPNR_NAMESPACE_PREFIX IdString, std::size_t>>
+{
+    std::size_t operator()(const std::pair<NEXTPNR_NAMESPACE_PREFIX IdString, std::size_t> &idp) const noexcept
+    {
+        std::size_t seed = 0;
+        boost::hash_combine(seed, hash<NEXTPNR_NAMESPACE_PREFIX IdString>()(idp.first));
+        boost::hash_combine(seed, hash<std::size_t>()(idp.second));
+        return seed;
+    }
+};
+} // namespace std
+
 NEXTPNR_NAMESPACE_BEGIN
 
 class SAPlacer
 {
+  private:
+    struct BoundingBox
+    {
+        int x0 = 0, x1 = 0, y0 = 0, y1 = 0;
+        bool is_inside_inc(int x, int y) const { return x >= x0 && x <= x1 && y >= y0 && y <= y1; }
+        bool touches_bounds(int x, int y) const { return x == x0 || x == x1 || y == y0 || y == y1; }
+        wirelen_t hpwl() const { return wirelen_t((x1 - x0) + (y1 - y0)); }
+    };
+
   public:
     SAPlacer(Context *ctx, Placer1Cfg cfg) : ctx(ctx), cfg(cfg)
     {
@@ -78,13 +101,44 @@ class SAPlacer
         }
         diameter = std::max(max_x, max_y) + 1;
 
-        costs.resize(ctx->nets.size());
+        net_bounds.resize(ctx->nets.size());
+        net_arc_tcost.resize(ctx->nets.size());
+        moveChange.already_bounds_changed.resize(ctx->nets.size());
+        moveChange.already_changed_arcs.resize(ctx->nets.size());
         old_udata.reserve(ctx->nets.size());
+        net_by_udata.reserve(ctx->nets.size());
         decltype(NetInfo::udata) n = 0;
         for (auto &net : ctx->nets) {
             old_udata.emplace_back(net.second->udata);
+            net_arc_tcost.at(n).resize(net.second->users.size());
+            moveChange.already_changed_arcs.at(n).resize(net.second->users.size());
             net.second->udata = n++;
+            net_by_udata.push_back(net.second.get());
         }
+        for (auto &region : sorted(ctx->region)) {
+            Region *r = region.second;
+            BoundingBox bb;
+            if (r->constr_bels) {
+                bb.x0 = std::numeric_limits<int>::max();
+                bb.x1 = std::numeric_limits<int>::min();
+                bb.y0 = std::numeric_limits<int>::max();
+                bb.y1 = std::numeric_limits<int>::min();
+                for (auto bel : r->bels) {
+                    Loc loc = ctx->getBelLocation(bel);
+                    bb.x0 = std::min(bb.x0, loc.x);
+                    bb.x1 = std::max(bb.x1, loc.x);
+                    bb.y0 = std::min(bb.y0, loc.y);
+                    bb.y1 = std::max(bb.y1, loc.y);
+                }
+            } else {
+                bb.x0 = 0;
+                bb.y0 = 0;
+                bb.x1 = max_x;
+                bb.y1 = max_y;
+            }
+            region_bounds[r->name] = bb;
+        }
+        build_port_index();
     }
 
     ~SAPlacer()
@@ -93,12 +147,15 @@ class SAPlacer
             net.second->udata = old_udata[net.second->udata];
     }
 
-    bool place()
+    bool place(bool refine = false)
     {
         log_break();
         ctx->lock();
 
         size_t placed_cells = 0;
+        std::vector<CellInfo *> autoplaced;
+        std::vector<CellInfo *> chain_basis;
+        if (!refine) {
             // Initial constraints placer
             for (auto &cell_entry : ctx->cells) {
                 CellInfo *cell = cell_entry.second.get();
@@ -126,7 +183,8 @@ class SAPlacer
 
                     auto bound_cell = ctx->getBoundBelCell(bel);
                     if (bound_cell) {
-                    log_error("Cell \'%s\' cannot be bound to bel \'%s\' since it is already bound to cell \'%s\'\n",
+                        log_error(
+                                "Cell \'%s\' cannot be bound to bel \'%s\' since it is already bound to cell \'%s\'\n",
                                 cell->name.c_str(ctx), loc_name.c_str(), bound_cell->name.c_str(ctx));
                     }
 
@@ -140,7 +198,7 @@ class SAPlacer
             ctx->yield();
 
             // Sort to-place cells for deterministic initial placement
-        std::vector<CellInfo *> autoplaced;
+
             for (auto &cell : ctx->cells) {
                 CellInfo *ci = cell.second.get();
                 if (ci->bel == BelId()) {
@@ -163,27 +221,47 @@ class SAPlacer
             if ((placed_cells - constr_placed_cells) % 500 != 0)
                 log_info("  initial placement placed %d/%d cells\n", int(placed_cells - constr_placed_cells),
                          int(autoplaced.size()));
-        if (ctx->slack_redist_iter > 0)
+            if (cfg.budgetBased && ctx->slack_redist_iter > 0)
                 assign_budget(ctx);
             ctx->yield();
             auto iplace_end = std::chrono::high_resolution_clock::now();
-        log_info("Initial placement time %.02fs\n", std::chrono::duration<float>(iplace_end - iplace_start).count());
+            log_info("Initial placement time %.02fs\n",
-        auto saplace_start = std::chrono::high_resolution_clock::now();
+                     std::chrono::duration<float>(iplace_end - iplace_start).count());
             log_info("Running simulated annealing placer.\n");
-
+        } else {
-        // Calculate metric after initial placement
+            for (auto &cell : ctx->cells) {
-        curr_metric = 0;
+                CellInfo *ci = cell.second.get();
-        curr_tns = 0;
+                if (ci->belStrength > STRENGTH_STRONG)
-        for (auto &net : ctx->nets) {
+                    continue;
-            wirelen_t wl = get_net_metric(ctx, net.second.get(), MetricType::COST, curr_tns);
+                else if (ci->constr_parent != nullptr)
-            costs[net.second->udata] = CostChange{wl, -1};
+                    continue;
-            curr_metric += wl;
+                else if (!ci->constr_children.empty() || ci->constr_z != ci->UNCONSTR)
+                    chain_basis.push_back(ci);
+                else
+                    autoplaced.push_back(ci);
             }
+            require_legal = false;
+            diameter = 3;
+            log_info("Running simulated annealing placer for refinement.\n");
+        }
+        auto saplace_start = std::chrono::high_resolution_clock::now();
+
+        // Invoke timing analysis to obtain criticalities
+        if (!cfg.budgetBased)
+            get_criticalities(ctx, &net_crit);
+
+        // Calculate costs after initial placement
+        setup_costs();
+        curr_wirelen_cost = total_wirelen_cost();
+        curr_timing_cost = total_timing_cost();
+        last_wirelen_cost = curr_wirelen_cost;
+        last_timing_cost = curr_timing_cost;
+
+        wirelen_t avg_wirelen = curr_wirelen_cost;
+        wirelen_t min_wirelen = curr_wirelen_cost;
 
         int n_no_progress = 0;
-        wirelen_t min_metric = curr_metric;
+        temp = refine ? 1e-7 : cfg.startTemp;
-        double avg_metric = curr_metric;
-        temp = 10000;
 
         // Main simulated annealing loop
         for (int iter = 1;; iter++) {
@@ -191,9 +269,9 @@ class SAPlacer
             improved = false;
 
             if (iter % 5 == 0 || iter == 1)
-                log_info("  at iteration #%d: temp = %f, cost = "
+                log_info("  at iteration #%d: temp = %f, timing cost = "
-                         "%.0f, est tns = %.02fns\n",
+                         "%.0f, wirelen = %.0f\n",
-                         iter, temp, double(curr_metric), curr_tns);
+                         iter, temp, double(curr_timing_cost), double(curr_wirelen_cost));
 
             for (int m = 0; m < 15; ++m) {
                 // Loop through all automatically placed cells
@@ -205,10 +283,17 @@ class SAPlacer
                     if (try_bel != BelId() && try_bel != cell->bel)
                         try_swap_position(cell, try_bel);
                 }
+                // Also try swapping chains, if applicable
+                for (auto cb : chain_basis) {
+                    Loc chain_base_loc = ctx->getBelLocation(cb->bel);
+                    BelId try_base = random_bel_for_cell(cb, chain_base_loc.z);
+                    if (try_base != BelId() && try_base != cb->bel)
+                        try_swap_chain(cb, try_base);
+                }
             }
 
-            if (curr_metric < min_metric) {
+            if (curr_wirelen_cost < min_wirelen) {
-                min_metric = curr_metric;
+                min_wirelen = curr_wirelen_cost;
                 improved = true;
             }
 
@@ -218,9 +303,10 @@ class SAPlacer
             else
                 n_no_progress++;
 
-            if (temp <= 1e-3 && n_no_progress >= 5) {
+            if (temp <= 1e-7 && n_no_progress >= (refine ? 1 : 5)) {
-                if (iter % 5 != 0)
+                log_info("  at iteration #%d: temp = %f, timing cost = "
-                    log_info("  at iteration #%d: temp = %f, cost = %f\n", iter, temp, double(curr_metric));
+                         "%.0f, wirelen = %.0f \n",
+                         iter, temp, double(curr_timing_cost), double(curr_wirelen_cost));
                 break;
             }
 
@@ -228,61 +314,64 @@ class SAPlacer
 
             int M = std::max(max_x, max_y) + 1;
 
-            double upper = 0.6, lower = 0.4;
+            if (ctx->verbose)
+                log("iter #%d: temp = %f, timing cost = "
+                    "%.0f, wirelen = %.0f, dia = %d, Ra = %.02f \n",
+                    iter, temp, double(curr_timing_cost), double(curr_wirelen_cost), diameter, Raccept);
 
-            if (curr_metric < 0.95 * avg_metric && curr_metric > 0) {
+            if (curr_wirelen_cost < 0.95 * avg_wirelen && curr_wirelen_cost > 0) {
-                avg_metric = 0.8 * avg_metric + 0.2 * curr_metric;
+                avg_wirelen = 0.8 * avg_wirelen + 0.2 * curr_wirelen_cost;
             } else {
-                if (Raccept >= 0.8) {
+                double diam_next = diameter * (1.0 - 0.44 + Raccept);
-                    temp *= 0.7;
+                diameter = std::max<int>(1, std::min<int>(M, int(diam_next + 0.5)));
-                } else if (Raccept > upper) {
+                if (Raccept > 0.96) {
-                    if (diameter < M)
+                    temp *= 0.5;
-                        diameter++;
+                } else if (Raccept > 0.8) {
-                    else
                     temp *= 0.9;
-                } else if (Raccept > lower) {
+                } else if (Raccept > 0.15 && diameter > 1) {
                     temp *= 0.95;
                 } else {
-                    // Raccept < 0.3
-                    if (diameter > 1)
-                        diameter--;
-                    else
                     temp *= 0.8;
                 }
             }
             // Once cooled below legalise threshold, run legalisation and start requiring
             // legal moves only
-            if (temp < legalise_temp && require_legal) {
+            if (diameter < legalise_dia && require_legal) {
                 if (legalise_relative_constraints(ctx)) {
                     // Only increase temperature if something was moved
                     autoplaced.clear();
+                    chain_basis.clear();
                     for (auto cell : sorted(ctx->cells)) {
-                        if (cell.second->belStrength < STRENGTH_STRONG)
+                        if (cell.second->belStrength <= STRENGTH_STRONG && cell.second->constr_parent == nullptr &&
+                            !cell.second->constr_children.empty())
+                            chain_basis.push_back(cell.second);
+                        else if (cell.second->belStrength < STRENGTH_STRONG)
                             autoplaced.push_back(cell.second);
                     }
-                    temp = post_legalise_temp;
+                    // temp = post_legalise_temp;
-                    diameter *= post_legalise_dia_scale;
+                    // diameter = std::min<int>(M, diameter * post_legalise_dia_scale);
                     ctx->shuffle(autoplaced);
 
                     // Legalisation is a big change so force a slack redistribution here
-                    if (ctx->slack_redist_iter > 0)
+                    if (ctx->slack_redist_iter > 0 && cfg.budgetBased)
                         assign_budget(ctx, true /* quiet */);
                 }
                 require_legal = false;
-            } else if (ctx->slack_redist_iter > 0 && iter % ctx->slack_redist_iter == 0) {
+            } else if (cfg.budgetBased && ctx->slack_redist_iter > 0 && iter % ctx->slack_redist_iter == 0) {
                 assign_budget(ctx, true /* quiet */);
             }
 
+            // Invoke timing analysis to obtain criticalities
+            if (!cfg.budgetBased && ctx->timing_driven)
+                get_criticalities(ctx, &net_crit);
+            // Need to rebuild costs after criticalities change
+            setup_costs();
             // Recalculate total metric entirely to avoid rounding errors
             // accumulating over time
-            curr_metric = 0;
+            curr_wirelen_cost = total_wirelen_cost();
-            curr_tns = 0;
+            curr_timing_cost = total_timing_cost();
-            for (auto &net : ctx->nets) {
+            last_wirelen_cost = curr_wirelen_cost;
-                wirelen_t wl = get_net_metric(ctx, net.second.get(), MetricType::COST, curr_tns);
+            last_timing_cost = curr_timing_cost;
-                costs[net.second->udata] = CostChange{wl, -1};
-                curr_metric += wl;
-            }
-
             // Let the UI show visualization updates.
             ctx->yield();
         }
@@ -334,7 +423,8 @@ class SAPlacer
                 ctx->unbindBel(cell->bel);
             }
             IdString targetType = cell->type;
-            for (auto bel : ctx->getBels()) {
+
+            auto proc_bel = [&](BelId bel) {
                 if (ctx->getBelType(bel) == targetType && ctx->isValidBelForCell(cell, bel)) {
                     if (ctx->checkBelAvail(bel)) {
                         uint64_t score = ctx->rng64();
@@ -352,7 +442,18 @@ class SAPlacer
                         }
                     }
                 }
+            };
+
+            if (cell->region != nullptr && cell->region->constr_bels) {
+                for (auto bel : cell->region->bels) {
+                    proc_bel(bel);
                 }
+            } else {
+                for (auto bel : ctx->getBels()) {
+                    proc_bel(bel);
+                }
+            }
+
             if (best_bel == BelId()) {
                 if (iters == 0 || ripup_bel == BelId())
                     log_error("failed to place cell '%s' of type '%s'\n", cell->name.c_str(ctx), cell->type.c_str(ctx));
@@ -373,49 +474,38 @@ class SAPlacer
     // Attempt a SA position swap, return true on success or false on failure
     bool try_swap_position(CellInfo *cell, BelId newBel)
     {
-        static std::vector<NetInfo *> updates;
+        static const double epsilon = 1e-20;
-        updates.clear();
+        moveChange.reset();
+        if (!require_legal && is_constrained(cell))
+            return false;
         BelId oldBel = cell->bel;
         CellInfo *other_cell = ctx->getBoundBelCell(newBel);
-        if (other_cell != nullptr && other_cell->belStrength > STRENGTH_WEAK) {
+        if (!require_legal && other_cell != nullptr &&
+            (is_constrained(other_cell) || other_cell->belStrength > STRENGTH_WEAK)) {
             return false;
         }
         int old_dist = get_constraints_distance(ctx, cell);
         int new_dist;
         if (other_cell != nullptr)
             old_dist += get_constraints_distance(ctx, other_cell);
-        wirelen_t new_metric = 0, delta;
+        double delta = 0;
         ctx->unbindBel(oldBel);
         if (other_cell != nullptr) {
             ctx->unbindBel(newBel);
         }
 
-        for (const auto &port : cell->ports) {
-            if (port.second.net != nullptr) {
-                auto &cost = costs[port.second.net->udata];
-                if (cost.new_cost == 0)
-                    continue;
-                cost.new_cost = 0;
-                updates.emplace_back(port.second.net);
-            }
-        }
-
-        if (other_cell != nullptr) {
-            for (const auto &port : other_cell->ports)
-                if (port.second.net != nullptr) {
-                    auto &cost = costs[port.second.net->udata];
-                    if (cost.new_cost == 0)
-                        continue;
-                    cost.new_cost = 0;
-                    updates.emplace_back(port.second.net);
-                }
-        }
-
         ctx->bindBel(newBel, cell, STRENGTH_WEAK);
 
         if (other_cell != nullptr) {
             ctx->bindBel(oldBel, other_cell, STRENGTH_WEAK);
         }
+
+        add_move_cell(moveChange, cell, oldBel);
+
+        if (other_cell != nullptr) {
+            add_move_cell(moveChange, other_cell, newBel);
+        }
+
         if (!ctx->isBelLocationValid(newBel) || ((other_cell != nullptr && !ctx->isBelLocationValid(oldBel)))) {
             ctx->unbindBel(newBel);
             if (other_cell != nullptr)
@@ -423,26 +513,18 @@ class SAPlacer
             goto swap_fail;
         }
 
-        new_metric = curr_metric;
-
         // Recalculate metrics for all nets touched by the peturbation
-        for (const auto &net : updates) {
+        compute_cost_changes(moveChange);
-            auto &c = costs[net->udata];
-            new_metric -= c.curr_cost;
-            float temp_tns = 0;
-            wirelen_t net_new_wl = get_net_metric(ctx, net, MetricType::COST, temp_tns);
-            new_metric += net_new_wl;
-            c.new_cost = net_new_wl;
-        }
 
         new_dist = get_constraints_distance(ctx, cell);
         if (other_cell != nullptr)
             new_dist += get_constraints_distance(ctx, other_cell);
-        delta = new_metric - curr_metric;
+        delta = lambda * (moveChange.timing_delta / std::max<double>(last_timing_cost, epsilon)) +
-        delta += (cfg.constraintWeight / temp) * (new_dist - old_dist);
+                (1 - lambda) * (double(moveChange.wirelen_delta) / std::max<double>(last_wirelen_cost, epsilon));
+        delta += (cfg.constraintWeight / temp) * (new_dist - old_dist) / last_wirelen_cost;
         n_move++;
         // SA acceptance criterea
-        if (delta < 0 || (temp > 1e-6 && (ctx->rng() / float(0x3fffffff)) <= std::exp(-delta / temp))) {
+        if (delta < 0 || (temp > 1e-8 && (ctx->rng() / float(0x3fffffff)) <= std::exp(-delta / temp))) {
             n_accept++;
         } else {
             if (other_cell != nullptr)
@@ -450,32 +532,148 @@ class SAPlacer
             ctx->unbindBel(newBel);
             goto swap_fail;
         }
-        curr_metric = new_metric;
+        commit_cost_changes(moveChange);
-        for (const auto &net : updates) {
+#if 0
-            auto &c = costs[net->udata];
+        log_info("swap %s -> %s\n", cell->name.c_str(ctx), ctx->getBelName(newBel).c_str(ctx));
-            c = CostChange{c.new_cost, -1};
+        if (other_cell != nullptr)
-        }
+            log_info("swap %s -> %s\n", other_cell->name.c_str(ctx), ctx->getBelName(oldBel).c_str(ctx));
-
+#endif
         return true;
     swap_fail:
         ctx->bindBel(oldBel, cell, STRENGTH_WEAK);
         if (other_cell != nullptr) {
             ctx->bindBel(newBel, other_cell, STRENGTH_WEAK);
         }
-        for (const auto &net : updates)
+        return false;
-            costs[net->udata].new_cost = -1;
+    }
+
+    inline bool is_constrained(CellInfo *cell)
+    {
+        return cell->constr_parent != nullptr || !cell->constr_children.empty();
+    }
+
+    // Swap the Bel of a cell with another, return the original location
+    BelId swap_cell_bels(CellInfo *cell, BelId newBel)
+    {
+        BelId oldBel = cell->bel;
+#if 0
+        log_info("%s old: %s new: %s\n", cell->name.c_str(ctx), ctx->getBelName(cell->bel).c_str(ctx), ctx->getBelName(newBel).c_str(ctx));
+#endif
+        CellInfo *bound = ctx->getBoundBelCell(newBel);
+        if (bound != nullptr)
+            ctx->unbindBel(newBel);
+        ctx->unbindBel(oldBel);
+        ctx->bindBel(newBel, cell, is_constrained(cell) ? STRENGTH_STRONG : STRENGTH_WEAK);
+        if (bound != nullptr)
+            ctx->bindBel(oldBel, bound, is_constrained(bound) ? STRENGTH_STRONG : STRENGTH_WEAK);
+        return oldBel;
+    }
+
+    // Discover the relative positions of all cells in a chain
+    void discover_chain(Loc baseLoc, CellInfo *cell, std::vector<std::pair<CellInfo *, Loc>> &cell_rel)
+    {
+        Loc cellLoc = ctx->getBelLocation(cell->bel);
+        Loc rel{cellLoc.x - baseLoc.x, cellLoc.y - baseLoc.y, cellLoc.z};
+        cell_rel.emplace_back(std::make_pair(cell, rel));
+        for (auto child : cell->constr_children)
+            discover_chain(baseLoc, child, cell_rel);
+    }
+
+    // Attempt to swap a chain with a non-chain
+    bool try_swap_chain(CellInfo *cell, BelId newBase)
+    {
+        std::vector<std::pair<CellInfo *, Loc>> cell_rel;
+        std::unordered_set<IdString> cells;
+        std::vector<std::pair<CellInfo *, BelId>> moves_made;
+        std::vector<std::pair<CellInfo *, BelId>> dest_bels;
+        double delta = 0;
+        moveChange.reset();
+        if (ctx->debug)
+            log_info("finding cells for chain swap %s\n", cell->name.c_str(ctx));
+
+        Loc baseLoc = ctx->getBelLocation(cell->bel);
+        discover_chain(baseLoc, cell, cell_rel);
+        Loc newBaseLoc = ctx->getBelLocation(newBase);
+        NPNR_ASSERT(newBaseLoc.z == baseLoc.z);
+        for (const auto &cr : cell_rel)
+            cells.insert(cr.first->name);
+
+        for (const auto &cr : cell_rel) {
+            Loc targetLoc = {newBaseLoc.x + cr.second.x, newBaseLoc.y + cr.second.y, cr.second.z};
+            BelId targetBel = ctx->getBelByLocation(targetLoc);
+            if (targetBel == BelId())
+                return false;
+            if (ctx->getBelType(targetBel) != cell->type)
+                return false;
+            CellInfo *bound = ctx->getBoundBelCell(targetBel);
+            // We don't consider swapping chains with other chains, at least for the time being - unless it is
+            // part of this chain
+            if (bound != nullptr && !cells.count(bound->name) &&
+                (bound->belStrength >= STRENGTH_STRONG || is_constrained(bound)))
+                return false;
+            dest_bels.emplace_back(std::make_pair(cr.first, targetBel));
+        }
+        if (ctx->debug)
+            log_info("trying chain swap %s\n", cell->name.c_str(ctx));
+        // <cell, oldBel>
+        for (const auto &db : dest_bels) {
+            BelId oldBel = swap_cell_bels(db.first, db.second);
+            moves_made.emplace_back(std::make_pair(db.first, oldBel));
+        }
+        for (const auto &mm : moves_made) {
+            if (!ctx->isBelLocationValid(mm.first->bel) || !check_cell_bel_region(mm.first, mm.first->bel))
+                goto swap_fail;
+            if (!ctx->isBelLocationValid(mm.second))
+                goto swap_fail;
+            CellInfo *bound = ctx->getBoundBelCell(mm.second);
+            if (bound && !check_cell_bel_region(bound, bound->bel))
+                goto swap_fail;
+            add_move_cell(moveChange, mm.first, mm.second);
+            if (bound != nullptr)
+                add_move_cell(moveChange, bound, mm.first->bel);
+        }
+        compute_cost_changes(moveChange);
+        delta = lambda * (moveChange.timing_delta / last_timing_cost) +
+                (1 - lambda) * (double(moveChange.wirelen_delta) / last_wirelen_cost);
+        n_move++;
+        // SA acceptance criterea
+        if (delta < 0 || (temp > 1e-9 && (ctx->rng() / float(0x3fffffff)) <= std::exp(-delta / temp))) {
+            n_accept++;
+            if (ctx->debug)
+                log_info("accepted chain swap %s\n", cell->name.c_str(ctx));
+        } else {
+            goto swap_fail;
+        }
+        commit_cost_changes(moveChange);
+        return true;
+    swap_fail:
+        for (const auto &entry : boost::adaptors::reverse(moves_made))
+            swap_cell_bels(entry.first, entry.second);
         return false;
     }
 
     // Find a random Bel of the correct type for a cell, within the specified
     // diameter
-    BelId random_bel_for_cell(CellInfo *cell)
+    BelId random_bel_for_cell(CellInfo *cell, int force_z = -1)
     {
         IdString targetType = cell->type;
         Loc curr_loc = ctx->getBelLocation(cell->bel);
+        int count = 0;
+
+        int dx = diameter, dy = diameter;
+        if (cell->region != nullptr && cell->region->constr_bels) {
+            dx = std::min(diameter, (region_bounds[cell->region->name].x1 - region_bounds[cell->region->name].x0) + 1);
+            dy = std::min(diameter, (region_bounds[cell->region->name].y1 - region_bounds[cell->region->name].y0) + 1);
+            // Clamp location to within bounds
+            curr_loc.x = std::max(region_bounds[cell->region->name].x0, curr_loc.x);
+            curr_loc.x = std::min(region_bounds[cell->region->name].x1, curr_loc.x);
+            curr_loc.y = std::max(region_bounds[cell->region->name].y0, curr_loc.y);
+            curr_loc.y = std::min(region_bounds[cell->region->name].y1, curr_loc.y);
+        }
+
         while (true) {
-            int nx = ctx->rng(2 * diameter + 1) + std::max(curr_loc.x - diameter, 0);
+            int nx = ctx->rng(2 * dx + 1) + std::max(curr_loc.x - dx, 0);
-            int ny = ctx->rng(2 * diameter + 1) + std::max(curr_loc.y - diameter, 0);
+            int ny = ctx->rng(2 * dy + 1) + std::max(curr_loc.y - dy, 0);
             int beltype_idx, beltype_cnt;
             std::tie(beltype_idx, beltype_cnt) = bel_types.at(targetType);
             if (beltype_cnt < cfg.minBelsForGridPick)
@@ -488,41 +686,262 @@ class SAPlacer
             if (fb.size() == 0)
                 continue;
             BelId bel = fb.at(ctx->rng(int(fb.size())));
+            if (force_z != -1) {
+                Loc loc = ctx->getBelLocation(bel);
+                if (loc.z != force_z)
+                    continue;
+            }
+            if (!check_cell_bel_region(cell, bel))
+                continue;
             if (locked_bels.find(bel) != locked_bels.end())
                 continue;
+            count++;
             return bel;
         }
     }
 
+    // Return true if a net is to be entirely ignored
+    inline bool ignore_net(NetInfo *net)
+    {
+        return net->driver.cell == nullptr || net->driver.cell->bel == BelId() ||
+               ctx->getBelGlobalBuf(net->driver.cell->bel);
+    }
+
+    // Get the bounding box for a net
+    inline BoundingBox get_net_bounds(NetInfo *net)
+    {
+        BoundingBox bb;
+        NPNR_ASSERT(net->driver.cell != nullptr);
+        Loc dloc = ctx->getBelLocation(net->driver.cell->bel);
+        bb.x0 = dloc.x;
+        bb.x1 = dloc.x;
+        bb.y0 = dloc.y;
+        bb.y1 = dloc.y;
+
+        for (auto user : net->users) {
+            if (user.cell->bel == BelId())
+                continue;
+            Loc uloc = ctx->getBelLocation(user.cell->bel);
+            bb.x0 = std::min(bb.x0, uloc.x);
+            bb.x1 = std::max(bb.x1, uloc.x);
+            bb.y0 = std::min(bb.y0, uloc.y);
+            bb.y1 = std::max(bb.y1, uloc.y);
+        }
+
+        return bb;
+    }
+
+    // Get the timing cost for an arc of a net
+    inline double get_timing_cost(NetInfo *net, size_t user)
+    {
+        int cc;
+        if (net->driver.cell == nullptr)
+            return 0;
+        if (ctx->getPortTimingClass(net->driver.cell, net->driver.port, cc) == TMG_IGNORE)
+            return 0;
+        if (cfg.budgetBased) {
+            double delay = ctx->getDelayNS(ctx->predictDelay(net, net->users.at(user)));
+            return std::min(10.0, std::exp(delay - ctx->getDelayNS(net->users.at(user).budget) / 10));
+        } else {
+            auto crit = net_crit.find(net->name);
+            if (crit == net_crit.end() || crit->second.criticality.empty())
+                return 0;
+            double delay = ctx->getDelayNS(ctx->predictDelay(net, net->users.at(user)));
+            return delay * std::pow(crit->second.criticality.at(user), crit_exp);
+        }
+    }
+
+    // Set up the cost maps
+    void setup_costs()
+    {
+        for (auto net : sorted(ctx->nets)) {
+            NetInfo *ni = net.second;
+            if (ignore_net(ni))
+                continue;
+            net_bounds[ni->udata] = get_net_bounds(ni);
+            if (ctx->timing_driven && int(ni->users.size()) < cfg.timingFanoutThresh)
+                for (size_t i = 0; i < ni->users.size(); i++)
+                    net_arc_tcost[ni->udata][i] = get_timing_cost(ni, i);
+        }
+    }
+
+    // Get the total wiring cost for the design
+    wirelen_t total_wirelen_cost()
+    {
+        wirelen_t cost = 0;
+        for (const auto &net : net_bounds)
+            cost += net.hpwl();
+        return cost;
+    }
+
+    // Get the total timing cost for the design
+    double total_timing_cost()
+    {
+        double cost = 0;
+        for (const auto &net : net_arc_tcost) {
+            for (auto arc_cost : net) {
+                cost += arc_cost;
+            }
+        }
+        return cost;
+    }
+
+    // Cost-change-related data for a move
+    struct MoveChangeData
+    {
+        std::vector<decltype(NetInfo::udata)> bounds_changed_nets;
+        std::vector<std::pair<decltype(NetInfo::udata), size_t>> changed_arcs;
+
+        std::vector<bool> already_bounds_changed;
+        std::vector<std::vector<bool>> already_changed_arcs;
+
+        std::vector<std::pair<decltype(NetInfo::udata), BoundingBox>> new_net_bounds;
+        std::vector<std::pair<std::pair<decltype(NetInfo::udata), size_t>, double>> new_arc_costs;
+
+        wirelen_t wirelen_delta = 0;
+        double timing_delta = 0;
+
+        void reset()
+        {
+            for (auto bc : bounds_changed_nets)
+                already_bounds_changed[bc] = false;
+            for (const auto &tc : changed_arcs)
+                already_changed_arcs[tc.first][tc.second] = false;
+            bounds_changed_nets.clear();
+            changed_arcs.clear();
+            new_net_bounds.clear();
+            new_arc_costs.clear();
+            wirelen_delta = 0;
+            timing_delta = 0;
+        }
+
+    } moveChange;
+
+    void add_move_cell(MoveChangeData &mc, CellInfo *cell, BelId old_bel)
+    {
+        Loc curr_loc = ctx->getBelLocation(cell->bel);
+        Loc old_loc = ctx->getBelLocation(old_bel);
+        // Check net bounds
+        for (const auto &port : cell->ports) {
+            NetInfo *pn = port.second.net;
+            if (pn == nullptr)
+                continue;
+            if (ignore_net(pn))
+                continue;
+            const BoundingBox &curr_bounds = net_bounds[pn->udata];
+            // If the old location was at the edge of the bounds, or the new location exceeds the bounds,
+            // an update is needed
+            if (curr_bounds.touches_bounds(old_loc.x, old_loc.y) || !curr_bounds.is_inside_inc(curr_loc.x, curr_loc.y))
+                if (!mc.already_bounds_changed[pn->udata]) {
+                    mc.bounds_changed_nets.push_back(pn->udata);
+                    mc.already_bounds_changed[pn->udata] = true;
+                }
+            if (ctx->timing_driven && int(pn->users.size()) < cfg.timingFanoutThresh) {
+                // Output ports - all arcs change timing
+                if (port.second.type == PORT_OUT) {
+                    int cc;
+                    TimingPortClass cls = ctx->getPortTimingClass(cell, port.first, cc);
+                    if (cls != TMG_IGNORE)
+                        for (size_t i = 0; i < pn->users.size(); i++)
+                            if (!mc.already_changed_arcs[pn->udata][i]) {
+                                mc.changed_arcs.emplace_back(std::make_pair(pn->udata, i));
+                                mc.already_changed_arcs[pn->udata][i] = true;
+                            }
+                } else if (port.second.type == PORT_IN) {
+                    auto usr = fast_port_to_user.at(&port.second);
+                    if (!mc.already_changed_arcs[pn->udata][usr]) {
+                        mc.changed_arcs.emplace_back(std::make_pair(pn->udata, usr));
+                        mc.already_changed_arcs[pn->udata][usr] = true;
+                    }
+                }
+            }
+        }
+    }
+
+    void compute_cost_changes(MoveChangeData &md)
+    {
+        for (const auto &bc : md.bounds_changed_nets) {
+            wirelen_t old_hpwl = net_bounds.at(bc).hpwl();
+            auto bounds = get_net_bounds(net_by_udata.at(bc));
+            md.new_net_bounds.emplace_back(std::make_pair(bc, bounds));
+            md.wirelen_delta += (bounds.hpwl() - old_hpwl);
+            md.already_bounds_changed[bc] = false;
+        }
+        if (ctx->timing_driven) {
+            for (const auto &tc : md.changed_arcs) {
+                double old_cost = net_arc_tcost.at(tc.first).at(tc.second);
+                double new_cost = get_timing_cost(net_by_udata.at(tc.first), tc.second);
+                md.new_arc_costs.emplace_back(std::make_pair(tc, new_cost));
+                md.timing_delta += (new_cost - old_cost);
+                md.already_changed_arcs[tc.first][tc.second] = false;
+            }
+        }
+    }
+
+    void commit_cost_changes(MoveChangeData &md)
+    {
+        for (const auto &bc : md.new_net_bounds)
+            net_bounds[bc.first] = bc.second;
+        for (const auto &tc : md.new_arc_costs)
+            net_arc_tcost[tc.first.first].at(tc.first.second) = tc.second;
+        curr_wirelen_cost += md.wirelen_delta;
+        curr_timing_cost += md.timing_delta;
+    }
+    // Build the cell port -> user index
+    void build_port_index()
+    {
+        for (auto net : sorted(ctx->nets)) {
+            NetInfo *ni = net.second;
+            for (size_t i = 0; i < ni->users.size(); i++) {
+                auto &usr = ni->users.at(i);
+                fast_port_to_user[&(usr.cell->ports.at(usr.port))] = i;
+            }
+        }
+    }
+
+    // Get the combined wirelen/timing metric
+    inline double curr_metric() { return lambda * curr_timing_cost + (1 - lambda) * curr_wirelen_cost; }
+
+    // Map nets to their bounding box (so we can skip recompute for moves that do not exceed the bounds
+    std::vector<BoundingBox> net_bounds;
+    // Map net arcs to their timing cost (criticality * delay ns)
+    std::vector<std::vector<double>> net_arc_tcost;
+
+    // Fast lookup for cell port to net user index
+    std::unordered_map<const PortInfo *, size_t> fast_port_to_user;
+
+    // Wirelength and timing cost at last and current iteration
+    wirelen_t last_wirelen_cost, curr_wirelen_cost;
+    double last_timing_cost, curr_timing_cost;
+
+    // Criticality data from timing analysis
+    NetCriticalityMap net_crit;
+
     Context *ctx;
-    wirelen_t curr_metric = std::numeric_limits<wirelen_t>::max();
+    float temp = 10;
-    float curr_tns = 0;
+    float crit_exp = 8;
-    float temp = 1000;
+    float lambda = 0.5;
     bool improved = false;
     int n_move, n_accept;
     int diameter = 35, max_x = 1, max_y = 1;
     std::unordered_map<IdString, std::tuple<int, int>> bel_types;
+    std::unordered_map<IdString, BoundingBox> region_bounds;
     std::vector<std::vector<std::vector<std::vector<BelId>>>> fast_bels;
     std::unordered_set<BelId> locked_bels;
-    bool require_legal = true;
+    std::vector<NetInfo *> net_by_udata;
-    const float legalise_temp = 1;
-    const float post_legalise_temp = 10;
-    const float post_legalise_dia_scale = 1.5;
-    Placer1Cfg cfg;
-
-    struct CostChange
-    {
-        wirelen_t curr_cost;
-        wirelen_t new_cost;
-    };
-    std::vector<CostChange> costs;
     std::vector<decltype(NetInfo::udata)> old_udata;
+    bool require_legal = true;
+    const int legalise_dia = 4;
+    Placer1Cfg cfg;
 };
 
 Placer1Cfg::Placer1Cfg(Context *ctx) : Settings(ctx)
 {
     constraintWeight = get<float>("placer1/constraintWeight", 10);
     minBelsForGridPick = get<int>("placer1/minBelsForGridPick", 64);
+    budgetBased = get<bool>("placer1/budgetBased", false);
+    startTemp = get<float>("placer1/startTemp", 1);
+    timingFanoutThresh = std::numeric_limits<int>::max();
 }
 
 bool placer1(Context *ctx, Placer1Cfg cfg)
@@ -545,4 +964,24 @@ bool placer1(Context *ctx, Placer1Cfg cfg)
     }
 }
 
+bool placer1_refine(Context *ctx, Placer1Cfg cfg)
+{
+    try {
+        SAPlacer placer(ctx, cfg);
+        placer.place(true);
+        log_info("Checksum: 0x%08x\n", ctx->checksum());
+#ifndef NDEBUG
+        ctx->lock();
+        ctx->check();
+        ctx->unlock();
+#endif
+        return true;
+    } catch (log_execution_error_exception) {
+#ifndef NDEBUG
+        ctx->check();
+#endif
+        return false;
+    }
+}
+
 NEXTPNR_NAMESPACE_END

common/placer1.h

@@ -29,9 +29,13 @@ struct Placer1Cfg : public Settings
     Placer1Cfg(Context *ctx);
     float constraintWeight;
     int minBelsForGridPick;
+    bool budgetBased;
+    float startTemp;
+    int timingFanoutThresh;
 };
 
 extern bool placer1(Context *ctx, Placer1Cfg cfg);
+extern bool placer1_refine(Context *ctx, Placer1Cfg cfg);
 
 NEXTPNR_NAMESPACE_END
 

common/placer_heap.h

@@ -0,0 +1,47 @@
+/*
+ *  nextpnr -- Next Generation Place and Route
+ *
+ *  Copyright (C) 2019  David Shah <david@symbioticeda.com>
+ *
+ *  Permission to use, copy, modify, and/or distribute this software for any
+ *  purpose with or without fee is hereby granted, provided that the above
+ *  copyright notice and this permission notice appear in all copies.
+ *
+ *  THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+ *  WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+ *  MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+ *  ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+ *  WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+ *  ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+ *  OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+ *
+ *  [[cite]] HeAP
+ *  Analytical Placement for Heterogeneous FPGAs, Marcel Gort and Jason H. Anderson
+ *  https://janders.eecg.utoronto.ca/pdfs/marcelfpl12.pdf
+ *
+ *  [[cite]] SimPL
+ *  SimPL: An Effective Placement Algorithm, Myung-Chul Kim, Dong-Jin Lee and Igor L. Markov
+ *  http://www.ece.umich.edu/cse/awards/pdfs/iccad10-simpl.pdf
+ */
+
+#ifndef PLACER_HEAP_H
+#define PLACER_HEAP
+#include "nextpnr.h"
+#include "settings.h"
+
+NEXTPNR_NAMESPACE_BEGIN
+
+struct PlacerHeapCfg : public Settings
+{
+    PlacerHeapCfg(Context *ctx);
+
+    float alpha;
+    float criticalityExponent;
+    float timingWeight;
+
+    std::unordered_set<IdString> ioBufTypes;
+};
+
+extern bool placer_heap(Context *ctx, PlacerHeapCfg cfg);
+NEXTPNR_NAMESPACE_END
+#endif

common/pybindings.cc

@@ -104,6 +104,7 @@ BOOST_PYTHON_MODULE(MODULE_NAME)
     typedef std::unordered_map<IdString, std::string> AttrMap;
     typedef std::unordered_map<IdString, PortInfo> PortMap;
     typedef std::unordered_map<IdString, IdString> PinMap;
+    typedef std::unordered_map<IdString, std::unique_ptr<Region>> RegionMap;
 
     class_<BaseCtx, BaseCtx *, boost::noncopyable>("BaseCtx", no_init);
 
@@ -135,6 +136,8 @@ BOOST_PYTHON_MODULE(MODULE_NAME)
 
     typedef std::vector<PortRef> PortRefVector;
     typedef std::unordered_map<WireId, PipMap> WireMap;
+    typedef std::unordered_set<BelId> BelSet;
+    typedef std::unordered_set<WireId> WireSet;
 
     auto ni_cls = class_<ContextualWrapper<NetInfo &>>("NetInfo", no_init);
     readwrite_wrapper<NetInfo &, decltype(&NetInfo::name), &NetInfo::name, conv_to_str<IdString>,
@@ -163,10 +166,25 @@ BOOST_PYTHON_MODULE(MODULE_NAME)
     def("parse_json", parse_json_shim);
     def("load_design", load_design_shim, return_value_policy<manage_new_object>());
 
+    auto region_cls = class_<ContextualWrapper<Region &>>("Region", no_init);
+    readwrite_wrapper<Region &, decltype(&Region::name), &Region::name, conv_to_str<IdString>,
+                      conv_from_str<IdString>>::def_wrap(region_cls, "name");
+    readwrite_wrapper<Region &, decltype(&Region::constr_bels), &Region::constr_bels, pass_through<bool>,
+                      pass_through<bool>>::def_wrap(region_cls, "constr_bels");
+    readwrite_wrapper<Region &, decltype(&Region::constr_wires), &Region::constr_wires, pass_through<bool>,
+                      pass_through<bool>>::def_wrap(region_cls, "constr_bels");
+    readwrite_wrapper<Region &, decltype(&Region::constr_pips), &Region::constr_pips, pass_through<bool>,
+                      pass_through<bool>>::def_wrap(region_cls, "constr_pips");
+    readonly_wrapper<Region &, decltype(&Region::bels), &Region::bels, wrap_context<BelSet &>>::def_wrap(region_cls,
+                                                                                                         "bels");
+    readonly_wrapper<Region &, decltype(&Region::wires), &Region::wires, wrap_context<WireSet &>>::def_wrap(region_cls,
+                                                                                                            "wires");
+
     WRAP_MAP(AttrMap, pass_through<std::string>, "AttrMap");
     WRAP_MAP(PortMap, wrap_context<PortInfo &>, "PortMap");
     WRAP_MAP(PinMap, conv_to_str<IdString>, "PinMap");
     WRAP_MAP(WireMap, wrap_context<PipMap &>, "WireMap");
+    WRAP_MAP_UPTR(RegionMap, "RegionMap");
 
     WRAP_VECTOR(PortRefVector, wrap_context<PortRef &>);
 

common/pycontainers.h

@@ -345,6 +345,12 @@ template <typename T, typename value_conv> struct map_wrapper
         std::terminate();
     }
 
+    static bool contains(wrapped_map &x, std::string const &i)
+    {
+        K k = PythonConversion::string_converter<K>().from_str(x.ctx, i);
+        return x.base.count(k);
+    }
+
     static void wrap(const char *map_name, const char *kv_name, const char *kv_iter_name, const char *iter_name)
     {
         map_pair_wrapper<typename KV::first_type, typename KV::second_type, value_conv>::wrap(kv_name, kv_iter_name);
@@ -353,6 +359,7 @@ template <typename T, typename value_conv> struct map_wrapper
         class_<wrapped_map>(map_name, no_init)
                 .def("__iter__", rw::iter)
                 .def("__len__", len)
+                .def("__contains__", contains)
                 .def("__getitem__", get)
                 .def("__setitem__", set, with_custodian_and_ward<1, 2>());
     }
@@ -465,6 +472,12 @@ template <typename T> struct map_wrapper_uptr
         std::terminate();
     }
 
+    static bool contains(wrapped_map &x, std::string const &i)
+    {
+        K k = PythonConversion::string_converter<K>().from_str(x.ctx, i);
+        return x.base.count(k);
+    }
+
     static void wrap(const char *map_name, const char *kv_name, const char *kv_iter_name, const char *iter_name)
     {
         map_pair_wrapper_uptr<typename KV::first_type, typename KV::second_type>::wrap(kv_name, kv_iter_name);
@@ -473,6 +486,7 @@ template <typename T> struct map_wrapper_uptr
         class_<wrapped_map>(map_name, no_init)
                 .def("__iter__", rw::iter)
                 .def("__len__", len)
+                .def("__contains__", contains)
                 .def("__getitem__", get)
                 .def("__setitem__", set, with_custodian_and_ward<1, 2>());
     }

common/pywrappers.h

@@ -269,7 +269,7 @@ template <typename Class, typename FuncT, FuncT fn, typename arg1_conv, typename
     template <typename WrapCls> static void def_wrap(WrapCls cls_, const char *name) { cls_.def(name, wrapped_fn); }
 };
 
-// Three parameters, one return
+// Three parameters, no return
 template <typename Class, typename FuncT, FuncT fn, typename arg1_conv, typename arg2_conv, typename arg3_conv>
 struct fn_wrapper_3a_v
 {
@@ -288,6 +288,30 @@ struct fn_wrapper_3a_v
     template <typename WrapCls> static void def_wrap(WrapCls cls_, const char *name) { cls_.def(name, wrapped_fn); }
 };
 
+// Five parameters, no return
+template <typename Class, typename FuncT, FuncT fn, typename arg1_conv, typename arg2_conv, typename arg3_conv,
+          typename arg4_conv, typename arg5_conv>
+struct fn_wrapper_5a_v
+{
+    using class_type = typename WrapIfNotContext<Class>::maybe_wrapped_t;
+    using conv_arg1_type = typename arg1_conv::arg_type;
+    using conv_arg2_type = typename arg2_conv::arg_type;
+    using conv_arg3_type = typename arg3_conv::arg_type;
+    using conv_arg4_type = typename arg4_conv::arg_type;
+    using conv_arg5_type = typename arg5_conv::arg_type;
+
+    static void wrapped_fn(class_type &cls, conv_arg1_type arg1, conv_arg2_type arg2, conv_arg3_type arg3,
+                           conv_arg4_type arg4, conv_arg5_type arg5)
+    {
+        Context *ctx = get_ctx<Class>(cls);
+        Class &base = get_base<Class>(cls);
+        return (base.*fn)(arg1_conv()(ctx, arg1), arg2_conv()(ctx, arg2), arg3_conv()(ctx, arg3),
+                          arg4_conv()(ctx, arg4), arg5_conv()(ctx, arg5));
+    }
+
+    template <typename WrapCls> static void def_wrap(WrapCls cls_, const char *name) { cls_.def(name, wrapped_fn); }
+};
+
 // Wrapped getter
 template <typename Class, typename MemT, MemT mem, typename v_conv> struct readonly_wrapper
 {

common/settings.h

@@ -45,19 +45,30 @@ class Settings
         return defaultValue;
     }
 
-    template <typename T> void set(const char *name, T value)
+    template <typename T> void set(const char *name, T value);
-    {
-        IdString id = ctx->id(name);
-        auto pair = ctx->settings.emplace(id, std::to_string(value));
-        if (!pair.second) {
-            ctx->settings[pair.first->first] = value;
-        }
-    }
 
   private:
     Context *ctx;
 };
 
+template <typename T> inline void Settings::set(const char *name, T value)
+{
+    IdString id = ctx->id(name);
+    auto pair = ctx->settings.emplace(id, std::to_string(value));
+    if (!pair.second) {
+        ctx->settings[pair.first->first] = value;
+    }
+}
+
+template <> inline void Settings::set<std::string>(const char *name, std::string value)
+{
+    IdString id = ctx->id(name);
+    auto pair = ctx->settings.emplace(id, value);
+    if (!pair.second) {
+        ctx->settings[pair.first->first] = value;
+    }
+}
+
 NEXTPNR_NAMESPACE_END
 
 #endif // SETTINGS_H

common/timing.cc

@@ -904,8 +904,7 @@ void timing_analysis(Context *ctx, bool print_histogram, bool print_fmax, bool p
             if (!warn_on_failure || passed)
                 log_info("Max frequency for clock %*s'%s': %.02f MHz (%s at %.02f MHz)\n", width, "",
                          clock_name.c_str(), clock_fmax[clock.first], passed ? "PASS" : "FAIL", target);
-            else
+            else if (bool_or_default(ctx->settings, ctx->id("timing/allowFail"), false))
-            if (bool_or_default(ctx->settings, ctx->id("timing/allowFail"), false))
                 log_warning("Max frequency for clock %*s'%s': %.02f MHz (%s at %.02f MHz)\n", width, "",
                             clock_name.c_str(), clock_fmax[clock.first], passed ? "PASS" : "FAIL", target);
             else

docs/archapi.md

@@ -490,3 +490,14 @@ a certain number of different clock signals allowed for a group of bels.
 
 Returns true if a bell in the current configuration is valid, i.e. if
 `isValidBelForCell()` would return true for the current mapping.
+
+
+### static const std::string defaultPlacer
+
+Name of the default placement algorithm for the architecture, if
+`--placer` isn't specified on the command line.
+
+### static const std::vector\<std::string\> availablePlacers
+
+Name of available placer algorithms for the architecture, used
+to provide help for and validate `--placer`.

ecp5/arch.cc

@@ -28,6 +28,7 @@
 #include "log.h"
 #include "nextpnr.h"
 #include "placer1.h"
+#include "placer_heap.h"
 #include "router1.h"
 #include "timing.h"
 #include "util.h"
@@ -456,6 +457,7 @@ delay_t Arch::estimateDelay(WireId src, WireId dst) const
     auto src_loc = est_location(src), dst_loc = est_location(dst);
 
     int dx = abs(src_loc.first - dst_loc.first), dy = abs(src_loc.second - dst_loc.second);
+
     return (130 - 25 * args.speed) *
            (6 + std::max(dx - 5, 0) + std::max(dy - 5, 0) + 2 * (std::min(dx, 5) + std::min(dy, 5)));
 }
@@ -467,7 +469,6 @@ delay_t Arch::predictDelay(const NetInfo *net_info, const PortRef &sink) const
         return 0;
     auto driver_loc = getBelLocation(driver.cell->bel);
     auto sink_loc = getBelLocation(sink.cell->bel);
-
     // Encourage use of direct interconnect
     if (driver_loc.x == sink_loc.x && driver_loc.y == sink_loc.y) {
         if ((sink.port == id_A0 || sink.port == id_A1) && (driver.port == id_F1) &&
@@ -485,6 +486,7 @@ delay_t Arch::predictDelay(const NetInfo *net_info, const PortRef &sink) const
     }
 
     int dx = abs(driver_loc.x - sink_loc.x), dy = abs(driver_loc.y - sink_loc.y);
+
     return (130 - 25 * args.speed) *
            (6 + std::max(dx - 5, 0) + std::max(dy - 5, 0) + 2 * (std::min(dx, 5) + std::min(dy, 5)));
 }
@@ -506,10 +508,23 @@ bool Arch::getBudgetOverride(const NetInfo *net_info, const PortRef &sink, delay
 
 bool Arch::place()
 {
-    bool result = placer1(getCtx(), Placer1Cfg(getCtx()));
+    std::string placer = str_or_default(settings, id("placer"), defaultPlacer);
-    if (result)
+
+    if (placer == "heap") {
+        PlacerHeapCfg cfg(getCtx());
+        cfg.criticalityExponent = 7;
+        cfg.ioBufTypes.insert(id_TRELLIS_IO);
+        if (!placer_heap(getCtx(), cfg))
+            return false;
+    } else if (placer == "sa") {
+        if (!placer1(getCtx(), Placer1Cfg(getCtx())))
+            return false;
+    } else {
+        log_error("ECP5 architecture does not support placer '%s'\n", placer.c_str());
+    }
+
     permute_luts();
-    return result;
+    return true;
 }
 
 bool Arch::route()
@@ -605,6 +620,11 @@ DecalXY Arch::getGroupDecal(GroupId pip) const { return {}; };
 
 bool Arch::getDelayFromTimingDatabase(IdString tctype, IdString from, IdString to, DelayInfo &delay) const
 {
+    auto fnd_dk = celldelay_cache.find({tctype, from, to});
+    if (fnd_dk != celldelay_cache.end()) {
+        delay = fnd_dk->second.second;
+        return fnd_dk->second.first;
+    }
     for (int i = 0; i < speed_grade->num_cell_timings; i++) {
         const auto &tc = speed_grade->cell_timings[i];
         if (tc.cell_type == tctype.index) {
@@ -613,9 +633,11 @@ bool Arch::getDelayFromTimingDatabase(IdString tctype, IdString from, IdString t
                 if (dly.from_port == from.index && dly.to_port == to.index) {
                     delay.max_delay = dly.max_delay;
                     delay.min_delay = dly.min_delay;
+                    celldelay_cache[{tctype, from, to}] = std::make_pair(true, delay);
                     return true;
                 }
             }
+            celldelay_cache[{tctype, from, to}] = std::make_pair(false, DelayInfo());
             return false;
         }
     }
@@ -645,7 +667,6 @@ void Arch::getSetupHoldFromTimingDatabase(IdString tctype, IdString clock, IdStr
 
 bool Arch::getCellDelay(const CellInfo *cell, IdString fromPort, IdString toPort, DelayInfo &delay) const
 {
-
     // Data for -8 grade
     if (cell->type == id_TRELLIS_SLICE) {
         bool has_carry = cell->sliceInfo.is_carry;
@@ -965,4 +986,16 @@ WireId Arch::getBankECLK(int bank, int eclk)
     return getWireByLocAndBasename(Location(0, 0), "G_BANK" + std::to_string(bank) + "ECLK" + std::to_string(eclk));
 }
 
+#ifdef WITH_HEAP
+const std::string Arch::defaultPlacer = "heap";
+#else
+const std::string Arch::defaultPlacer = "sa";
+#endif
+
+const std::vector<std::string> Arch::availablePlacers = {"sa",
+#ifdef WITH_HEAP
+                                                         "heap"
+#endif
+};
+
 NEXTPNR_NAMESPACE_END

ecp5/arch.h

@@ -448,6 +448,30 @@ struct ArchArgs
     } speed = SPEED_6;
 };
 
+struct DelayKey
+{
+    IdString celltype, from, to;
+    inline bool operator==(const DelayKey &other) const
+    {
+        return celltype == other.celltype && from == other.from && to == other.to;
+    }
+};
+
+NEXTPNR_NAMESPACE_END
+namespace std {
+template <> struct hash<NEXTPNR_NAMESPACE_PREFIX DelayKey>
+{
+    std::size_t operator()(const NEXTPNR_NAMESPACE_PREFIX DelayKey &dk) const noexcept
+    {
+        std::size_t seed = std::hash<NEXTPNR_NAMESPACE_PREFIX IdString>()(dk.celltype);
+        seed ^= std::hash<NEXTPNR_NAMESPACE_PREFIX IdString>()(dk.from) + 0x9e3779b9 + (seed << 6) + (seed >> 2);
+        seed ^= std::hash<NEXTPNR_NAMESPACE_PREFIX IdString>()(dk.to) + 0x9e3779b9 + (seed << 6) + (seed >> 2);
+        return seed;
+    }
+};
+} // namespace std
+NEXTPNR_NAMESPACE_BEGIN
+
 struct Arch : BaseCtx
 {
     const ChipInfoPOD *chip_info;
@@ -1019,6 +1043,11 @@ struct Arch : BaseCtx
     IdString id_clk, id_lsr;
     IdString id_clkmux, id_lsrmux;
     IdString id_srmode, id_mode;
+
+    mutable std::unordered_map<DelayKey, std::pair<bool, DelayInfo>> celldelay_cache;
+
+    static const std::string defaultPlacer;
+    static const std::vector<std::string> availablePlacers;
 };
 
 NEXTPNR_NAMESPACE_END

ecp5/arch_pybindings.cc

@@ -133,6 +133,13 @@ void arch_wrap_python()
 
     fn_wrapper_2a_v<Context, decltype(&Context::addClock), &Context::addClock, conv_from_str<IdString>,
                     pass_through<float>>::def_wrap(ctx_cls, "addClock");
+    fn_wrapper_5a_v<Context, decltype(&Context::createRectangularRegion), &Context::createRectangularRegion,
+                    conv_from_str<IdString>, pass_through<int>, pass_through<int>, pass_through<int>,
+                    pass_through<int>>::def_wrap(ctx_cls, "createRectangularRegion");
+    fn_wrapper_2a_v<Context, decltype(&Context::addBelToRegion), &Context::addBelToRegion, conv_from_str<IdString>,
+                    conv_from_str<BelId>>::def_wrap(ctx_cls, "addBelToRegion");
+    fn_wrapper_2a_v<Context, decltype(&Context::constrainCellToRegion), &Context::constrainCellToRegion,
+                    conv_from_str<IdString>, conv_from_str<IdString>>::def_wrap(ctx_cls, "constrainCellToRegion");
 
     WRAP_RANGE(Bel, conv_to_str<BelId>);
     WRAP_RANGE(Wire, conv_to_str<WireId>);

ecp5/main.cc

@@ -149,8 +149,8 @@ std::unique_ptr<Context> ECP5CommandHandler::createContext()
             chipArgs.speed = ArchArgs::SPEED_6;
         }
     }
-
+    auto ctx = std::unique_ptr<Context>(new Context(chipArgs));
-    return std::unique_ptr<Context>(new Context(chipArgs));
+    return ctx;
 }
 
 void ECP5CommandHandler::customAfterLoad(Context *ctx)

generic/arch.cc

@@ -21,6 +21,7 @@
 #include "nextpnr.h"
 #include "placer1.h"
 #include "router1.h"
+#include "util.h"
 
 NEXTPNR_NAMESPACE_BEGIN
 
@@ -439,7 +440,16 @@ bool Arch::getBudgetOverride(const NetInfo *net_info, const PortRef &sink, delay
 
 // ---------------------------------------------------------------
 
-bool Arch::place() { return placer1(getCtx(), Placer1Cfg(getCtx())); }
+bool Arch::place()
+{
+    std::string placer = str_or_default(settings, id("placer"), defaultPlacer);
+    // FIXME: No HeAP because it needs a list of IO buffers
+    if (placer == "sa") {
+        return placer1(getCtx(), Placer1Cfg(getCtx()));
+    } else {
+        log_error("Generic architecture does not support placer '%s'\n", placer.c_str());
+    }
+}
 
 bool Arch::route() { return router1(getCtx(), Router1Cfg(getCtx())); }
 
@@ -476,4 +486,7 @@ TimingClockingInfo Arch::getPortClockingInfo(const CellInfo *cell, IdString port
 bool Arch::isValidBelForCell(CellInfo *cell, BelId bel) const { return true; }
 bool Arch::isBelLocationValid(BelId bel) const { return true; }
 
+const std::string Arch::defaultPlacer = "sa";
+const std::vector<std::string> Arch::availablePlacers = {"sa"};
+
 NEXTPNR_NAMESPACE_END

generic/arch.h

@@ -240,6 +240,9 @@ struct Arch : BaseCtx
 
     bool isValidBelForCell(CellInfo *cell, BelId bel) const;
     bool isBelLocationValid(BelId bel) const;
+
+    static const std::string defaultPlacer;
+    static const std::vector<std::string> availablePlacers;
 };
 
 NEXTPNR_NAMESPACE_END

ice40/arch.cc

@@ -26,10 +26,10 @@
 #include "log.h"
 #include "nextpnr.h"
 #include "placer1.h"
+#include "placer_heap.h"
 #include "router1.h"
 #include "timing_opt.h"
 #include "util.h"
-
 NEXTPNR_NAMESPACE_BEGIN
 
 // -----------------------------------------------------------------------
@@ -671,8 +671,18 @@ bool Arch::getBudgetOverride(const NetInfo *net_info, const PortRef &sink, delay
 
 bool Arch::place()
 {
+    std::string placer = str_or_default(settings, id("placer"), defaultPlacer);
+    if (placer == "heap") {
+        PlacerHeapCfg cfg(getCtx());
+        cfg.ioBufTypes.insert(id_SB_IO);
+        if (!placer_heap(getCtx(), cfg))
+            return false;
+    } else if (placer == "sa") {
         if (!placer1(getCtx(), Placer1Cfg(getCtx())))
             return false;
+    } else {
+        log_error("iCE40 architecture does not support placer '%s'\n", placer.c_str());
+    }
     if (bool_or_default(settings, id("opt_timing"), false)) {
         TimingOptCfg tocfg(getCtx());
         tocfg.cellTypes.insert(id_ICESTORM_LC);
@@ -1198,4 +1208,12 @@ void Arch::assignCellInfo(CellInfo *cell)
     }
 }
 
+const std::string Arch::defaultPlacer = "sa";
+
+const std::vector<std::string> Arch::availablePlacers = {"sa",
+#ifdef WITH_HEAP
+                                                         "heap"
+#endif
+};
+
 NEXTPNR_NAMESPACE_END

ice40/arch.h

@@ -897,6 +897,9 @@ struct Arch : BaseCtx
         IdString glb_net = getWireName(getBelPinWire(bel, id_GLOBAL_BUFFER_OUTPUT));
         return std::stoi(std::string("") + glb_net.str(this).back());
     }
+
+    static const std::string defaultPlacer;
+    static const std::vector<std::string> availablePlacers;
 };
 
 void ice40DelayFuzzerMain(Context *ctx);

ice40/arch_pybindings.cc

@@ -144,6 +144,13 @@ void arch_wrap_python()
 
     fn_wrapper_2a_v<Context, decltype(&Context::addClock), &Context::addClock, conv_from_str<IdString>,
                     pass_through<float>>::def_wrap(ctx_cls, "addClock");
+    fn_wrapper_5a_v<Context, decltype(&Context::createRectangularRegion), &Context::createRectangularRegion,
+                    conv_from_str<IdString>, pass_through<int>, pass_through<int>, pass_through<int>,
+                    pass_through<int>>::def_wrap(ctx_cls, "createRectangularRegion");
+    fn_wrapper_2a_v<Context, decltype(&Context::addBelToRegion), &Context::addBelToRegion, conv_from_str<IdString>,
+                    conv_from_str<BelId>>::def_wrap(ctx_cls, "addBelToRegion");
+    fn_wrapper_2a_v<Context, decltype(&Context::constrainCellToRegion), &Context::constrainCellToRegion,
+                    conv_from_str<IdString>, conv_from_str<IdString>>::def_wrap(ctx_cls, "constrainCellToRegion");
 
     WRAP_RANGE(Bel, conv_to_str<BelId>);
     WRAP_RANGE(Wire, conv_to_str<WireId>);

ice40/examples/floorplan/.gitignore

@@ -0,0 +1,4 @@
+*.json
+*.asc
+*.bin
+__pycache__

ice40/examples/floorplan/floorplan.py

@@ -0,0 +1,5 @@
+ctx.createRectangularRegion("osc", 1, 1, 1, 4)
+for cell, cellinfo in ctx.cells:
+    if "ringosc" in cellinfo.attrs:
+        print("Floorplanned cell %s" % cell)
+        ctx.constrainCellToRegion(cell, "osc")

ice40/examples/floorplan/floorplan.sh

@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+set -ex
+yosys -p "synth_ice40 -top top -json floorplan.json" floorplan.v
+../../../nextpnr-ice40 --up5k --json floorplan.json --pcf icebreaker.pcf --asc floorplan.asc --ignore-loops --pre-place floorplan.py
+icepack floorplan.asc floorplan.bin
+iceprog floorplan.bin

ice40/examples/floorplan/floorplan.v

@@ -0,0 +1,22 @@
+module top(output LED1, LED2, LED3, LED4, LED5);
+    localparam N = 31;
+    wire [N:0] x;
+    assign x[0] = x[N];
+
+    genvar ii;
+    generate
+
+    for (ii = 0; ii < N; ii = ii + 1) begin
+        (* ringosc *)
+        SB_LUT4 #(.LUT_INIT(1)) lut_i(.I0(x[ii]), .I1(), .I2(), .I3(), .O(x[ii+1]));
+    end
+    endgenerate
+
+    assign clk = x[N];
+
+
+    reg [19:0] ctr;
+    always @(posedge clk)
+        ctr <= ctr + 1'b1;
+    assign {LED5, LED4, LED3, LED2, LED1} = ctr[19:15];
+endmodule

ice40/examples/floorplan/icebreaker.pcf

@@ -0,0 +1,5 @@
+set_io -nowarn LED1       26
+set_io -nowarn LED2       27
+set_io -nowarn LED3       25
+set_io -nowarn LED4       23
+set_io -nowarn LED5       21

ice40/main.cc

@@ -176,7 +176,6 @@ std::unique_ptr<Context> Ice40CommandHandler::createContext()
         ctx->settings[ctx->id("opt_timing")] = "1";
     if (vm.count("pcf-allow-unconstrained"))
         ctx->settings[ctx->id("pcf_allow_unconstrained")] = "1";
-
     return ctx;
 }
 

tests

@@ -1 +1 @@
-Subproject commit f29dcbe187b517d01964b1074eb7ff0b90849eed
+Subproject commit 32a683071758ee59d47e2c5cb29c87882993facd