From d678b6533dfac431a36abd22c01e9df184b723a4 Mon Sep 17 00:00:00 2001
From: Clifford Wolf <clifford@clifford.at>
Date: Wed, 16 Jul 2014 13:37:41 +0200
Subject: [PATCH] improved opt_reduce for $mem/$memwr WR_EN multiplexers
---
passes/opt/opt_reduce.cc | 80 ++++++++++++++++++++++++++++++++++++++++
1 file changed, 80 insertions(+)
diff --git a/passes/opt/opt_reduce.cc b/passes/opt/opt_reduce.cc
index dfe214416..d5bcb7f00 100644
--- a/passes/opt/opt_reduce.cc
+++ b/passes/opt/opt_reduce.cc
@@ -171,6 +171,66 @@ struct OptReduceWorker
}
}
+ void opt_mux_bits(RTLIL::Cell *cell)
+ {
+ std::vector<RTLIL::SigBit> sig_a = assign_map(cell->connections["\\A"]).to_sigbit_vector();
+ std::vector<RTLIL::SigBit> sig_b = assign_map(cell->connections["\\B"]).to_sigbit_vector();
+ std::vector<RTLIL::SigBit> sig_y = assign_map(cell->connections["\\Y"]).to_sigbit_vector();
+
+ std::vector<RTLIL::SigBit> new_sig_y;
+ RTLIL::SigSig old_sig_conn;
+
+ std::vector<std::vector<RTLIL::SigBit>> consolidated_in_tuples;
+ std::map<std::vector<RTLIL::SigBit>, RTLIL::SigBit> consolidated_in_tuples_map;
+
+ for (int i = 0; i < int(sig_y.size()); i++)
+ {
+ std::vector<RTLIL::SigBit> in_tuple;
+ in_tuple.push_back(sig_a.at(i));
+ for (int j = i; j < int(sig_b.size()); j += int(sig_a.size()))
+ in_tuple.push_back(sig_b.at(j));
+
+ if (consolidated_in_tuples_map.count(in_tuple))
+ {
+ old_sig_conn.first.append_bit(sig_y.at(i));
+ old_sig_conn.second.append_bit(consolidated_in_tuples_map.at(in_tuple));
+ }
+ else
+ {
+ consolidated_in_tuples_map[in_tuple] = sig_y.at(i);
+ consolidated_in_tuples.push_back(in_tuple);
+ new_sig_y.push_back(sig_y.at(i));
+ }
+ }
+
+ if (new_sig_y.size() != sig_y.size())
+ {
+ log(" Consolidated identical input bits for %s cell %s:\n", cell->type.c_str(), cell->name.c_str());
+ log(" Old inputs: A=%s, B=%s\n", log_signal(cell->connections["\\A"]), log_signal(cell->connections["\\B"]));
+
+ cell->connections["\\A"] = RTLIL::SigSpec();
+ for (auto &in_tuple : consolidated_in_tuples)
+ cell->connections["\\A"].append(in_tuple.at(0));
+
+ cell->connections["\\B"] = RTLIL::SigSpec();
+ for (int i = 1; i <= cell->connections["\\S"].width; i++)
+ for (auto &in_tuple : consolidated_in_tuples)
+ cell->connections["\\B"].append(in_tuple.at(i));
+
+ log(" New inputs: A=%s, B=%s\n", log_signal(cell->connections["\\A"]), log_signal(cell->connections["\\B"]));
+
+ cell->parameters["\\WIDTH"] = RTLIL::Const(new_sig_y.size());
+ cell->connections["\\Y"] = new_sig_y;
+
+ module->connections.push_back(old_sig_conn);
+ module->check();
+
+ did_something = true;
+ OPT_DID_SOMETHING = true;
+ total_count++;
+ }
+ }
+
OptReduceWorker(RTLIL::Design *design, RTLIL::Module *module) :
design(design), module(module), assign_map(module)
{
@@ -179,6 +239,20 @@ struct OptReduceWorker
total_count = 0;
did_something = true;
+ SigPool mem_wren_sigs;
+ for (auto &cell_it : module->cells) {
+ RTLIL::Cell *cell = cell_it.second;
+ if (cell->type == "$mem")
+ mem_wren_sigs.add(assign_map(cell->connections["\\WR_EN"]));
+ if (cell->type == "$memwr")
+ mem_wren_sigs.add(assign_map(cell->connections["\\EN"]));
+ }
+ for (auto &cell_it : module->cells) {
+ RTLIL::Cell *cell = cell_it.second;
+ if (cell->type == "$dff" && mem_wren_sigs.check_any(assign_map(cell->connections["\\Q"])))
+ mem_wren_sigs.add(assign_map(cell->connections["\\D"]));
+ }
+
while (did_something)
{
did_something = false;
@@ -213,6 +287,12 @@ struct OptReduceWorker
RTLIL::Cell *cell = cell_it.second;
if ((cell->type != "$mux" && cell->type != "$pmux" && cell->type != "$safe_pmux") || !design->selected(module, cell))
continue;
+
+ // this optimization is to aggressive for most coarse-grain applications.
+ // but we always want it for multiplexers driving write enable ports.
+ if (mem_wren_sigs.check_any(assign_map(cell->connections.at("\\Y"))))
+ opt_mux_bits(cell);
+
opt_mux(cell);
}
}