Merge pull request #44 from YosysHQ/improve_timing_spec

Speed up budget allocator using topographical ordering and update cell timing API

common/nextpnr.h

@@ -291,6 +291,19 @@ struct CellInfo : ArchCellInfo
     // parent.[xyz] := 0 when (constr_parent == nullptr)
 };
 
+enum TimingPortClass
+{
+    TMG_CLOCK_INPUT,     // Clock input to a sequential cell
+    TMG_GEN_CLOCK,       // Generated clock output (PLL, DCC, etc)
+    TMG_REGISTER_INPUT,  // Input to a register, with an associated clock (may also have comb. fanout too)
+    TMG_REGISTER_OUTPUT, // Output from a register
+    TMG_COMB_INPUT,      // Combinational input, no paths end here
+    TMG_COMB_OUTPUT,     // Combinational output, no paths start here
+    TMG_STARTPOINT,      // Unclocked primary startpoint, such as an IO cell output
+    TMG_ENDPOINT,        // Unclocked primary endpoint, such as an IO cell input
+    TMG_IGNORE,          // Asynchronous to all clocks, "don't care", and should be ignored (false path) for analysis
+};
+
 struct DeterministicRNG
 {
     uint64_t rngstate;
@@ -440,7 +453,7 @@ struct BaseCtx
 
     const Context *getCtx() const { return reinterpret_cast<const Context *>(this); }
 
-    template<typename T> const char *nameOf(const T *obj)
+    template <typename T> const char *nameOf(const T *obj)
     {
         if (obj == nullptr)
             return "";

common/placer1.cc

@@ -236,7 +236,10 @@ class SAPlacer
                 temp = post_legalise_temp;
                 diameter *= post_legalise_dia_scale;
                 ctx->shuffle(autoplaced);
-                assign_budget(ctx);
+
+                // Legalisation is a big change so force a slack redistribution here
+                if (ctx->slack_redist_iter > 0)
+                    assign_budget(ctx, true /* quiet */);
             } else if (ctx->slack_redist_iter > 0 && iter % ctx->slack_redist_iter == 0) {
                 assign_budget(ctx, true /* quiet */);
             }

common/timing.cc

@@ -20,6 +20,7 @@
 
 #include "timing.h"
 #include <algorithm>
+#include <boost/range/adaptor/reversed.hpp>
 #include <unordered_map>
 #include <utility>
 #include "log.h"
@@ -36,10 +37,18 @@ struct Timing
     bool net_delays;
     bool update;
     delay_t min_slack;
-    PortRefVector current_path;
     PortRefVector *crit_path;
     DelayFrequency *slack_histogram;
 
+    struct TimingData
+    {
+        TimingData() : max_arrival(), max_path_length(), min_remaining_budget() {}
+        TimingData(delay_t max_arrival) : max_arrival(max_arrival), max_path_length(), min_remaining_budget() {}
+        delay_t max_arrival;
+        unsigned max_path_length = 0;
+        delay_t min_remaining_budget;
+    };
+
     Timing(Context *ctx, bool net_delays, bool update, PortRefVector *crit_path = nullptr,
            DelayFrequency *slack_histogram = nullptr)
             : ctx(ctx), net_delays(net_delays), update(update), min_slack(1.0e12 / ctx->target_freq),
@@ -47,104 +56,256 @@ struct Timing
     {
     }
 
-    delay_t follow_net(NetInfo *net, int path_length, delay_t slack)
+    delay_t walk_paths()
     {
-        const delay_t default_budget = slack / (path_length + 1);
+        const auto clk_period = delay_t(1.0e12 / ctx->target_freq);
-        delay_t net_budget = default_budget;
-        for (auto &usr : net->users) {
-            auto delay = net_delays ? ctx->getNetinfoRouteDelay(net, usr) : delay_t();
-            if (crit_path)
-                current_path.push_back(&usr);
-            // If budget override exists, use that value and do not increment path_length
-            auto budget = default_budget;
-            if (ctx->getBudgetOverride(net, usr, budget)) {
-                if (update)
-                    usr.budget = std::min(usr.budget, budget);
-                budget = follow_user_port(usr, path_length, slack - budget);
-                net_budget = std::min(net_budget, budget);
-            }
-            else {
-                budget = follow_user_port(usr, path_length + 1, slack - delay);
-                net_budget = std::min(net_budget, budget);
-                if (update)
-                    usr.budget = std::min(usr.budget, delay + budget);
-            }
-            if (crit_path)
-                current_path.pop_back();
-        }
-        return net_budget;
-    }
 
-    // Follow a path, returning budget to annotate
+        // First, compute the topographical order of nets to walk through the circuit, assuming it is a _acyclic_ graph
-    delay_t follow_user_port(PortRef &user, int path_length, delay_t slack)
+        // TODO(eddieh): Handle the case where it is cyclic, e.g. combinatorial loops
-    {
+        std::vector<NetInfo *> topographical_order;
-        delay_t value;
+        std::unordered_map<const NetInfo *, TimingData> net_data;
-        if (ctx->getPortClock(user.cell, user.port) != IdString()) {
+        // In lieu of deleting edges from the graph, simply count the number of fanins to each output port
-            // At the end of a timing path (arguably, should check setup time
+        std::unordered_map<const PortInfo *, unsigned> port_fanin;
-            // here too)
+
-            value = slack / path_length;
+        std::vector<IdString> input_ports;
-            if (slack < min_slack) {
+        std::vector<const PortInfo *> output_ports;
-                min_slack = slack;
+        for (auto &cell : ctx->cells) {
-                if (crit_path)
+            input_ports.clear();
-                    *crit_path = current_path;
+            output_ports.clear();
+            for (auto &port : cell.second->ports) {
+                if (!port.second.net)
+                    continue;
+                if (port.second.type == PORT_OUT)
+                    output_ports.push_back(&port.second);
+                else
+                    input_ports.push_back(port.first);
             }
-            if (slack_histogram) {
+
-                int slack_ps = ctx->getDelayNS(slack) * 1000;
+            for (auto o : output_ports) {
-                (*slack_histogram)[slack_ps]++;
+                IdString clockPort;
+                TimingPortClass portClass = ctx->getPortTimingClass(cell.second.get(), o->name, clockPort);
+                // If output port is influenced by a clock (e.g. FF output) then add it to the ordering as a timing
+                // start-point
+                if (portClass == TMG_REGISTER_OUTPUT) {
+                    DelayInfo clkToQ;
+                    ctx->getCellDelay(cell.second.get(), clockPort, o->name, clkToQ);
+                    topographical_order.emplace_back(o->net);
+                    net_data.emplace(o->net, TimingData{clkToQ.maxDelay()});
+                } else {
+                    // TODO(eddieh): Generated clocks and ignored ports are currently added into the ordering as if it
+                    // was a regular timing start point in order to enable the full topographical order to be computed,
+                    // however these false nets (and their downstream paths) should not be in the final ordering
+                    if (portClass == TMG_STARTPOINT || portClass == TMG_GEN_CLOCK || portClass == TMG_IGNORE) {
+                        topographical_order.emplace_back(o->net);
+                        net_data.emplace(o->net, TimingData{});
+                    }
+                    // Otherwise, for all driven input ports on this cell, if a timing arc exists between the input and
+                    // the current output port, increment fanin counter
+                    for (auto i : input_ports) {
+                        DelayInfo comb_delay;
+                        bool is_path = ctx->getCellDelay(cell.second.get(), i, o->name, comb_delay);
+                        if (is_path)
+                            port_fanin[o]++;
+                    }
+                }
             }
-        } else {
+        }
-            // Default to the path ending here, if no further paths found
+
-            value = slack / path_length;
+        // If these constant nets exist, add them to the topographical ordering too
-            // Follow outputs of the user
+        // TODO(eddieh): Also false paths and should be removed from ordering
-            for (auto port : user.cell->ports) {
+        auto it = ctx->nets.find(ctx->id("$PACKER_VCC_NET"));
-                if (port.second.type == PORT_OUT) {
+        if (it != ctx->nets.end()) {
+            topographical_order.emplace_back(it->second.get());
+            net_data.emplace(it->second.get(), TimingData{});
+        }
+        it = ctx->nets.find(ctx->id("$PACKER_GND_NET"));
+        if (it != ctx->nets.end()) {
+            topographical_order.emplace_back(it->second.get());
+            net_data.emplace(it->second.get(), TimingData{});
+        }
+
+        std::deque<NetInfo *> queue(topographical_order.begin(), topographical_order.end());
+
+        // Now walk the design, from the start points identified previously, building up a topographical order
+        while (!queue.empty()) {
+            const auto net = queue.front();
+            queue.pop_front();
+
+            for (auto &usr : net->users) {
+                IdString clockPort;
+                TimingPortClass usrClass = ctx->getPortTimingClass(usr.cell, usr.port, clockPort);
+                if (usrClass == TMG_IGNORE || usrClass == TMG_CLOCK_INPUT)
+                    continue;
+                for (auto &port : usr.cell->ports) {
+                    if (port.second.type != PORT_OUT || !port.second.net)
+                        continue;
+                    TimingPortClass portClass = ctx->getPortTimingClass(usr.cell, port.first, clockPort);
+
+                    // Skip if this is a clocked output (but allow non-clocked ones)
+                    if (portClass == TMG_REGISTER_OUTPUT || portClass == TMG_STARTPOINT || portClass == TMG_IGNORE ||
+                        portClass == TMG_GEN_CLOCK)
+                        continue;
                     DelayInfo comb_delay;
-                    // Look up delay through this path
+                    bool is_path = ctx->getCellDelay(usr.cell, usr.port, port.first, comb_delay);
-                    bool is_path = ctx->getCellDelay(user.cell, user.port, port.first, comb_delay);
+                    if (!is_path)
-                    if (is_path) {
+                        continue;
-                        NetInfo *net = port.second.net;
+                    // Decrement the fanin count, and only add to topographical order if all its fanins have already
-                        if (net) {
+                    // been visited
-                            delay_t path_budget = follow_net(net, path_length, slack - comb_delay.maxDelay());
+                    auto it = port_fanin.find(&port.second);
-                            value = std::min(value, path_budget);
+                    NPNR_ASSERT(it != port_fanin.end());
+                    if (--it->second == 0) {
+                        topographical_order.emplace_back(port.second.net);
+                        queue.emplace_back(port.second.net);
+                        port_fanin.erase(it);
+                    }
+                }
+            }
+        }
+
+        // Sanity check to ensure that all ports where fanins were recorded were indeed visited
+        NPNR_ASSERT(port_fanin.empty());
+
+        // Go forwards topographically to find the maximum arrival time and max path length for each net
+        for (auto net : topographical_order) {
+            auto &nd = net_data.at(net);
+            const auto net_arrival = nd.max_arrival;
+            const auto net_length_plus_one = nd.max_path_length + 1;
+            nd.min_remaining_budget = clk_period;
+            for (auto &usr : net->users) {
+                IdString clockPort;
+                TimingPortClass portClass = ctx->getPortTimingClass(usr.cell, usr.port, clockPort);
+                if (portClass == TMG_REGISTER_INPUT || portClass == TMG_ENDPOINT || portClass == TMG_IGNORE) {
+                } else {
+                    auto net_delay = net_delays ? ctx->getNetinfoRouteDelay(net, usr) : delay_t();
+                    auto budget_override = ctx->getBudgetOverride(net, usr, net_delay);
+                    auto usr_arrival = net_arrival + net_delay;
+                    // Iterate over all output ports on the same cell as the sink
+                    for (auto port : usr.cell->ports) {
+                        if (port.second.type != PORT_OUT || !port.second.net)
+                            continue;
+                        DelayInfo comb_delay;
+                        // Look up delay through this path
+                        bool is_path = ctx->getCellDelay(usr.cell, usr.port, port.first, comb_delay);
+                        if (!is_path)
+                            continue;
+                        auto &data = net_data[port.second.net];
+                        auto &arrival = data.max_arrival;
+                        arrival = std::max(arrival, usr_arrival + comb_delay.maxDelay());
+                        if (!budget_override) { // Do not increment path length if budget overriden since it doesn't
+                                                // require a share of the slack
+                            auto &path_length = data.max_path_length;
+                            path_length = std::max(path_length, net_length_plus_one);
                         }
                     }
                 }
             }
         }
-        return value;
-    }
 
-    delay_t walk_paths()
+        const NetInfo *crit_net = nullptr;
-    {
-        delay_t default_slack = delay_t(1.0e12 / ctx->target_freq);
 
-        // Go through all clocked drivers and distribute the available path
+        // Now go backwards topographically to determine the minimum path slack, and to distribute all path slack evenly
-        //   slack evenly into the budget of every sink on the path
+        // between all nets on the path
-        for (auto &cell : ctx->cells) {
+        for (auto net : boost::adaptors::reverse(topographical_order)) {
-            for (auto port : cell.second->ports) {
+            auto &nd = net_data.at(net);
-                if (port.second.type == PORT_OUT) {
+            const delay_t net_length_plus_one = nd.max_path_length + 1;
-                    IdString clock_domain = ctx->getPortClock(cell.second.get(), port.first);
+            auto &net_min_remaining_budget = nd.min_remaining_budget;
-                    if (clock_domain != IdString()) {
+            for (auto &usr : net->users) {
-                        delay_t slack = default_slack; // TODO: clock constraints
+                auto net_delay = net_delays ? ctx->getNetinfoRouteDelay(net, usr) : delay_t();
-                        DelayInfo clkToQ;
+                auto budget_override = ctx->getBudgetOverride(net, usr, net_delay);
-                        if (ctx->getCellDelay(cell.second.get(), clock_domain, port.first, clkToQ))
+                IdString associatedClock;
-                            slack -= clkToQ.maxDelay();
+                TimingPortClass portClass = ctx->getPortTimingClass(usr.cell, usr.port, associatedClock);
-                        if (port.second.net)
+                if (portClass == TMG_REGISTER_INPUT || portClass == TMG_ENDPOINT) {
-                            follow_net(port.second.net, 0, slack);
+                    const auto net_arrival = nd.max_arrival;
+                    auto path_budget = clk_period - (net_arrival + net_delay);
+                    if (update) {
+                        auto budget_share = budget_override ? 0 : path_budget / net_length_plus_one;
+                        usr.budget = std::min(usr.budget, net_delay + budget_share);
+                        net_min_remaining_budget = std::min(net_min_remaining_budget, path_budget - budget_share);
+                    }
+
+                    if (path_budget < min_slack) {
+                        min_slack = path_budget;
+                        if (crit_path) {
+                            crit_path->clear();
+                            crit_path->push_back(&usr);
+                            crit_net = net;
+                        }
+                    }
+                    if (slack_histogram) {
+                        int slack_ps = ctx->getDelayNS(path_budget) * 1000;
+                        (*slack_histogram)[slack_ps]++;
+                    }
+                } else if (update) {
+                    // Iterate over all output ports on the same cell as the sink
+                    for (const auto &port : usr.cell->ports) {
+                        if (port.second.type != PORT_OUT || !port.second.net)
+                            continue;
+                        DelayInfo comb_delay;
+                        bool is_path = ctx->getCellDelay(usr.cell, usr.port, port.first, comb_delay);
+                        if (!is_path)
+                            continue;
+                        auto path_budget = net_data.at(port.second.net).min_remaining_budget;
+                        auto budget_share = budget_override ? 0 : path_budget / net_length_plus_one;
+                        usr.budget = std::min(usr.budget, net_delay + budget_share);
+                        net_min_remaining_budget = std::min(net_min_remaining_budget, path_budget - budget_share);
                     }
                 }
             }
         }
+
+        if (crit_path) {
+            // Walk backwards from the most critical net
+            while (crit_net) {
+                const PortInfo *crit_ipin = nullptr;
+                delay_t max_arrival = std::numeric_limits<delay_t>::min();
+
+                // Look at all input ports on its driving cell
+                for (const auto &port : crit_net->driver.cell->ports) {
+                    if (port.second.type != PORT_IN || !port.second.net)
+                        continue;
+                    DelayInfo comb_delay;
+                    bool is_path =
+                            ctx->getCellDelay(crit_net->driver.cell, port.first, crit_net->driver.port, comb_delay);
+                    if (!is_path)
+                        continue;
+                    // If input port is influenced by a clock, skip
+                    IdString portClock;
+                    TimingPortClass portClass = ctx->getPortTimingClass(crit_net->driver.cell, port.first, portClock);
+                    if (portClass == TMG_REGISTER_INPUT || portClass == TMG_CLOCK_INPUT || portClass == TMG_ENDPOINT ||
+                        portClass == TMG_IGNORE)
+                        continue;
+
+                    // And find the fanin net with the latest arrival time
+                    const auto net_arrival = net_data.at(port.second.net).max_arrival;
+                    if (net_arrival > max_arrival) {
+                        max_arrival = net_arrival;
+                        crit_ipin = &port.second;
+                    }
+                }
+
+                if (!crit_ipin)
+                    break;
+
+                // Now convert PortInfo* into a PortRef*
+                for (auto &usr : crit_ipin->net->users) {
+                    if (usr.cell->name == crit_net->driver.cell->name && usr.port == crit_ipin->name) {
+                        crit_path->push_back(&usr);
+                        break;
+                    }
+                }
+                crit_net = crit_ipin->net;
+            }
+            std::reverse(crit_path->begin(), crit_path->end());
+        }
         return min_slack;
     }
 
     void assign_budget()
     {
         // Clear delays to a very high value first
-        delay_t default_slack = delay_t(1.0e12 / ctx->target_freq);
         for (auto &net : ctx->nets) {
             for (auto &usr : net.second->users) {
-                usr.budget = default_slack;
+                usr.budget = std::numeric_limits<delay_t>::max();
             }
         }
 
@@ -180,16 +341,15 @@ void assign_budget(Context *ctx, bool quiet)
         }
     }
 
-    // For slack redistribution, if user has not specified a frequency
+    // For slack redistribution, if user has not specified a frequency dynamically adjust the target frequency to be the
-    //   dynamically adjust the target frequency to be the currently
+    // currently achieved maximum
-    //   achieved maximum
     if (ctx->auto_freq && ctx->slack_redist_iter > 0) {
-        delay_t default_slack = delay_t(1.0e12 / ctx->target_freq);
+        delay_t default_slack = delay_t((1.0e9 / ctx->getDelayNS(1)) / ctx->target_freq);
-        ctx->target_freq = 1e12 / (default_slack - timing.min_slack);
+        ctx->target_freq = 1.0e9 / ctx->getDelayNS(default_slack - timing.min_slack);
         if (ctx->verbose)
-            log_info("minimum slack for this assign = %d, target Fmax for next "
+            log_info("minimum slack for this assign = %.2f ns, target Fmax for next "
                      "update = %.2f MHz\n",
-                     timing.min_slack, ctx->target_freq / 1e6);
+                     ctx->getDelayNS(timing.min_slack), ctx->target_freq / 1e6);
     }
 
     if (!quiet)
@@ -217,7 +377,9 @@ void timing_analysis(Context *ctx, bool print_histogram, bool print_path)
             auto &front = crit_path.front();
             auto &front_port = front->cell->ports.at(front->port);
             auto &front_driver = front_port.net->driver;
-            auto last_port = ctx->getPortClock(front_driver.cell, front_driver.port);
+
+            IdString last_port;
+            ctx->getPortTimingClass(front_driver.cell, front_driver.port, last_port);
             for (auto sink : crit_path) {
                 auto sink_cell = sink->cell;
                 auto &port = sink_cell->ports.at(sink->port);
@@ -227,14 +389,15 @@ void timing_analysis(Context *ctx, bool print_histogram, bool print_path)
                 DelayInfo comb_delay;
                 ctx->getCellDelay(sink_cell, last_port, driver.port, comb_delay);
                 total += comb_delay.maxDelay();
-                log_info("%4d %4d  Source %s.%s\n", comb_delay.maxDelay(), total, driver_cell->name.c_str(ctx),
+                log_info("%4.1f %4.1f  Source %s.%s\n", ctx->getDelayNS(comb_delay.maxDelay()), ctx->getDelayNS(total),
-                         driver.port.c_str(ctx));
+                         driver_cell->name.c_str(ctx), driver.port.c_str(ctx));
                 auto net_delay = ctx->getNetinfoRouteDelay(net, *sink);
                 total += net_delay;
                 auto driver_loc = ctx->getBelLocation(driver_cell->bel);
                 auto sink_loc = ctx->getBelLocation(sink_cell->bel);
-                log_info("%4d %4d    Net %s budget %d (%d,%d) -> (%d,%d)\n", net_delay, total, net->name.c_str(ctx),
+                log_info("%4.1f %4.1f    Net %s budget %f ns (%d,%d) -> (%d,%d)\n", ctx->getDelayNS(net_delay),
-                         sink->budget, driver_loc.x, driver_loc.y, sink_loc.x, sink_loc.y);
+                         ctx->getDelayNS(total), net->name.c_str(ctx), ctx->getDelayNS(sink->budget), driver_loc.x,
+                         driver_loc.y, sink_loc.x, sink_loc.y);
                 log_info("                Sink %s.%s\n", sink_cell->name.c_str(ctx), sink->port.c_str(ctx));
                 last_port = sink->port;
             }
@@ -242,8 +405,8 @@ void timing_analysis(Context *ctx, bool print_histogram, bool print_path)
         }
     }
 
-    delay_t default_slack = delay_t(1.0e12 / ctx->target_freq);
+    delay_t default_slack = delay_t((1.0e9 / ctx->getDelayNS(1)) / ctx->target_freq);
-    log_info("estimated Fmax = %.2f MHz\n", 1e6 / (default_slack - min_slack));
+    log_info("estimated Fmax = %.2f MHz\n", 1e3 / ctx->getDelayNS(default_slack - min_slack));
 
     if (print_histogram && slack_histogram.size() > 0) {
         constexpr unsigned num_bins = 20;

docs/archapi.md

@@ -455,13 +455,11 @@ Cell Delay Methods
 Returns the delay for the specified path through a cell in the `&delay` argument. The method returns
 false if there is no timing relationship from `fromPort` to `toPort`.
 
-### IdString getPortClock(const CellInfo \*cell, IdString port) const
+### TimingPortClass getPortTimingClass(const CellInfo *cell, IdString port, IdString &clockPort) const
 
-Returns the clock input port for the specified output port.
+Return the _timing port class_ of a port. This can be a register or combinational input or output; clock input or
-
+output; general startpoint or endpoint; or a port ignored for timing purposes. For register ports, clockPort is set
-### bool isClockPort(const CellInfo \*cell, IdString port) const
+to the associated clock port.
-
-Returns true if the specified port is a clock input.
 
 Placer Methods
 --------------

ecp5/arch.cc

@@ -375,7 +375,6 @@ BelId Arch::getPioByFunctionName(const std::string &name) const
 }
 
 std::vector<PortPin> Arch::getBelPins(BelId bel) const
-
 {
     std::vector<PortPin> ret;
     NPNR_ASSERT(bel != BelId());
@@ -496,9 +495,10 @@ bool Arch::getCellDelay(const CellInfo *cell, IdString fromPort, IdString toPort
     return false;
 }
 
-IdString Arch::getPortClock(const CellInfo *cell, IdString port) const { return IdString(); }
+TimingPortClass Arch::getPortTimingClass(const CellInfo *cell, IdString port, IdString &clockPort) const
-
+{
-bool Arch::isClockPort(const CellInfo *cell, IdString port) const { return false; }
+    return TMG_IGNORE;
+}
 
 std::vector<std::pair<std::string, std::string>> Arch::getTilesAtLocation(int row, int col)
 {

ecp5/arch.h

@@ -828,10 +828,8 @@ struct Arch : BaseCtx
     // Get the delay through a cell from one port to another, returning false
     // if no path exists
     bool getCellDelay(const CellInfo *cell, IdString fromPort, IdString toPort, DelayInfo &delay) const;
-    // Get the associated clock to a port, or empty if the port is combinational
+    // Get the port class, also setting clockPort if applicable
-    IdString getPortClock(const CellInfo *cell, IdString port) const;
+    TimingPortClass getPortTimingClass(const CellInfo *cell, IdString port, IdString &clockPort) const;
-    // Return true if a port is a clock
-    bool isClockPort(const CellInfo *cell, IdString port) const;
     // Return true if a port is a net
     bool isGlobalNet(const NetInfo *net) const;
 

generic/arch.cc

@@ -435,9 +435,11 @@ bool Arch::getCellDelay(const CellInfo *cell, IdString fromPort, IdString toPort
     return false;
 }
 
-IdString Arch::getPortClock(const CellInfo *cell, IdString port) const { return IdString(); }
+// Get the port class, also setting clockPort if applicable
-
+TimingPortClass Arch::getPortTimingClass(const CellInfo *cell, IdString port, IdString &clockPort) const
-bool Arch::isClockPort(const CellInfo *cell, IdString port) const { return false; }
+{
+    return TMG_IGNORE;
+}
 
 bool Arch::isValidBelForCell(CellInfo *cell, BelId bel) const { return true; }
 bool Arch::isBelLocationValid(BelId bel) const { return true; }

generic/arch.h

@@ -215,8 +215,8 @@ struct Arch : BaseCtx
     DecalXY getGroupDecal(GroupId group) const;
 
     bool getCellDelay(const CellInfo *cell, IdString fromPort, IdString toPort, DelayInfo &delay) const;
-    IdString getPortClock(const CellInfo *cell, IdString port) const;
+    // Get the port class, also setting clockPort if applicable
-    bool isClockPort(const CellInfo *cell, IdString port) const;
+    TimingPortClass getPortTimingClass(const CellInfo *cell, IdString port, IdString &clockPort) const;
 
     bool isValidBelForCell(CellInfo *cell, BelId bel) const;
     bool isBelLocationValid(BelId bel) const;

ice40/arch.cc

@@ -27,6 +27,7 @@
 #include "placer1.h"
 #include "router1.h"
 #include "util.h"
+
 NEXTPNR_NAMESPACE_BEGIN
 
 // -----------------------------------------------------------------------
@@ -106,7 +107,9 @@ BelType Arch::belTypeFromId(IdString type) const
 void IdString::initialize_arch(const BaseCtx *ctx)
 {
 #define X(t) initialize_add(ctx, #t, PIN_##t);
+
 #include "portpins.inc"
+
 #undef X
 }
 
@@ -291,7 +294,8 @@ BelId Arch::getBelByLocation(Loc loc) const
 
 BelRange Arch::getBelsByTile(int x, int y) const
 {
-    // In iCE40 chipdb bels at the same tile are consecutive and dense z ordinates are used
+    // In iCE40 chipdb bels at the same tile are consecutive and dense z ordinates
+    // are used
     BelRange br;
 
     br.b.cursor = Arch::getBelByLocation(Loc(x, y, 0)).index;
@@ -645,23 +649,27 @@ bool Arch::getBudgetOverride(const NetInfo *net_info, const PortRef &sink, delay
         auto sink_loc = getBelLocation(sink.cell->bel);
         if (driver_loc.y == sink_loc.y)
             budget = 0;
-        else switch (args.type) {
+        else
+            switch (args.type) {
 #ifndef ICE40_HX1K_ONLY
             case ArchArgs::HX8K:
 #endif
             case ArchArgs::HX1K:
-                budget = 190; break;
+                budget = 190;
+                break;
 #ifndef ICE40_HX1K_ONLY
             case ArchArgs::LP384:
             case ArchArgs::LP1K:
             case ArchArgs::LP8K:
-                budget = 290; break;
+                budget = 290;
+                break;
             case ArchArgs::UP5K:
-                budget = 560; break;
+                budget = 560;
+                break;
 #endif
             default:
                 log_error("Unsupported iCE40 chip type.\n");
-        }
+            }
         return true;
     }
     return false;
@@ -883,27 +891,76 @@ bool Arch::getCellDelay(const CellInfo *cell, IdString fromPort, IdString toPort
     return false;
 }
 
-IdString Arch::getPortClock(const CellInfo *cell, IdString port) const
+// Get the port class, also setting clockPort to associated clock if applicable
+TimingPortClass Arch::getPortTimingClass(const CellInfo *cell, IdString port, IdString &clockPort) const
 {
-    if (cell->type == id_icestorm_lc && cell->lcInfo.dffEnable) {
+    if (cell->type == id_icestorm_lc) {
-        if (port != id_lo && port != id_cin && port != id_cout)
+        if (port == id_clk)
-            return id_clk;
+            return TMG_CLOCK_INPUT;
+        if (port == id_cin)
+            return TMG_COMB_INPUT;
+        if (port == id_cout || port == id_lo)
+            return TMG_COMB_OUTPUT;
+        if (cell->lcInfo.dffEnable) {
+            clockPort = id_clk;
+            if (port == id_o)
+                return TMG_REGISTER_OUTPUT;
+            else
+                return TMG_REGISTER_INPUT;
+        } else {
+            if (port == id_o)
+                return TMG_COMB_OUTPUT;
+            else
+                return TMG_COMB_INPUT;
+        }
     } else if (cell->type == id_icestorm_ram) {
-        if (port.str(this)[0] == 'R')
-            return id_rclk;
-        else
-            return id_wclk;
-    }
-    return IdString();
-}
 
-bool Arch::isClockPort(const CellInfo *cell, IdString port) const
+        if (port == id_rclk || port == id_wclk)
-{
+            return TMG_CLOCK_INPUT;
-    if (cell->type == id("ICESTORM_LC") && port == id("CLK"))
+
-        return true;
+        if (port.str(this)[0] == 'R')
-    if (cell->type == id("ICESTORM_RAM") && (port == id("RCLK") || (port == id("WCLK"))))
+            clockPort = id_rclk;
-        return true;
+        else
-    return false;
+            clockPort = id_wclk;
+
+        if (cell->ports.at(port).type == PORT_OUT)
+            return TMG_REGISTER_OUTPUT;
+        else
+            return TMG_REGISTER_INPUT;
+    } else if (cell->type == id("ICESTORM_DSP") || cell->type == id("ICESTORM_SPRAM")) {
+        clockPort = id_clk;
+        if (port == id_clk)
+            return TMG_CLOCK_INPUT;
+        else if (cell->ports.at(port).type == PORT_OUT)
+            return TMG_REGISTER_OUTPUT;
+        else
+            return TMG_REGISTER_INPUT;
+    } else if (cell->type == id_sb_io) {
+        if (port == id("D_IN_0") || port == id("D_IN_1"))
+            return TMG_STARTPOINT;
+        if (port == id("D_OUT_0") || port == id("D_OUT_1") || port == id("OUTPUT_ENABLE"))
+            return TMG_ENDPOINT;
+        return TMG_IGNORE;
+    } else if (cell->type == id("ICESTORM_PLL")) {
+        if (port == id("PLLOUT_A") || port == id("PLLOUT_B"))
+            return TMG_GEN_CLOCK;
+        return TMG_IGNORE;
+    } else if (cell->type == id("ICESTORM_LFOSC")) {
+        if (port == id("CLKLF"))
+            return TMG_GEN_CLOCK;
+        return TMG_IGNORE;
+    } else if (cell->type == id("ICESTORM_HFOSC")) {
+        if (port == id("CLKHF"))
+            return TMG_GEN_CLOCK;
+        return TMG_IGNORE;
+    } else if (cell->type == id_sb_gb) {
+        if (port == id_glb_buf_out)
+            return TMG_COMB_OUTPUT;
+        return TMG_COMB_INPUT;
+    } else if (cell->type == id("SB_WARMBOOT")) {
+        return TMG_ENDPOINT;
+    }
+    log_error("no timing info for port '%s' of cell type '%s'\n", port.c_str(this), cell->type.c_str(this));
 }
 
 bool Arch::isGlobalNet(const NetInfo *net) const

ice40/arch.h

@@ -400,10 +400,10 @@ struct Arch : BaseCtx
     mutable std::unordered_map<Loc, int> bel_by_loc;
 
     std::vector<bool> bel_carry;
-    std::vector<CellInfo*> bel_to_cell;
+    std::vector<CellInfo *> bel_to_cell;
-    std::vector<NetInfo*> wire_to_net;
+    std::vector<NetInfo *> wire_to_net;
-    std::vector<NetInfo*> pip_to_net;
+    std::vector<NetInfo *> pip_to_net;
-    std::vector<NetInfo*> switches_locked;
+    std::vector<NetInfo *> switches_locked;
 
     ArchArgs args;
     Arch(ArchArgs args);
@@ -789,16 +789,15 @@ struct Arch : BaseCtx
     // Get the delay through a cell from one port to another, returning false
     // if no path exists
     bool getCellDelay(const CellInfo *cell, IdString fromPort, IdString toPort, DelayInfo &delay) const;
-    // Get the associated clock to a port, or empty if the port is combinational
+    // Get the port class, also setting clockDomain if applicable
-    IdString getPortClock(const CellInfo *cell, IdString port) const;
+    TimingPortClass getPortTimingClass(const CellInfo *cell, IdString port, IdString &clockDomain) const;
-    // Return true if a port is a clock
-    bool isClockPort(const CellInfo *cell, IdString port) const;
     // Return true if a port is a net
     bool isGlobalNet(const NetInfo *net) const;
 
     // -------------------------------------------------
 
-    // Perform placement validity checks, returning false on failure (all implemented in arch_place.cc)
+    // Perform placement validity checks, returning false on failure (all
+    // implemented in arch_place.cc)
 
     // Whether or not a given cell can be placed at a given Bel
     // This is not intended for Bel type checks, but finer-grained constraints
@@ -812,7 +811,8 @@ struct Arch : BaseCtx
     bool logicCellsCompatible(const std::vector<const CellInfo *> &cells) const;
 
     // -------------------------------------------------
-    // Assign architecure-specific arguments to nets and cells, which must be called between packing or further
+    // Assign architecure-specific arguments to nets and cells, which must be
+    // called between packing or further
     // netlist modifications, and validity checks
     void assignArchInfo();
     void assignCellInfo(CellInfo *cell);