From 303a386ecc89bbff7cfbc0d4a64535087826a372 Mon Sep 17 00:00:00 2001
From: "N. Engelhardt" <nak@yosyshq.com>
Date: Fri, 31 Jan 2025 11:28:57 +0100
Subject: [PATCH] create duplicate IOFFs if multiple output ports are connected
 to the same register

---
 techlibs/quicklogic/ql_ioff.cc           | 66 +++++++++++++++++++-----
 techlibs/quicklogic/synth_quicklogic.cc  |  5 +-
 tests/arch/quicklogic/qlf_k6n10f/ioff.ys | 37 +++++++++++++
 3 files changed, 93 insertions(+), 15 deletions(-)

diff --git a/techlibs/quicklogic/ql_ioff.cc b/techlibs/quicklogic/ql_ioff.cc
index 4cc9785d5..87b62e855 100644
--- a/techlibs/quicklogic/ql_ioff.cc
+++ b/techlibs/quicklogic/ql_ioff.cc
@@ -29,7 +29,17 @@ struct QlIoffPass : public Pass {
 		if (!module)
 			return;
 		modwalker.setup(module);
-		pool<RTLIL::Cell *> cells_to_replace;
+		pool<RTLIL::Cell *> input_ffs;
+		dict<RTLIL::Wire *, std::vector<Cell*>> output_ffs;
+		dict<SigBit, pool<SigBit>> output_bit_aliases;
+
+		for (Wire* wire : module->wires())
+			if (wire->port_output) {
+				output_ffs[wire].resize(wire->width, nullptr);
+				for (SigBit bit : SigSpec(wire))
+					output_bit_aliases[modwalker.sigmap(bit)].insert(bit);
+			}
+
 		for (auto cell : module->selected_cells()) {
 			if (cell->type.in(ID(dffsre), ID(sdffsre))) {
 				log_debug("Checking cell %s.\n", cell->name.c_str());
@@ -53,32 +63,62 @@ struct QlIoffPass : public Pass {
 						log_debug("not promoting: D has other consumers\n");
 						continue;
 					}
-					cells_to_replace.insert(cell);
-					continue; // no need to check Q if we already put it on the list
+					input_ffs.insert(cell);
+					continue; // prefer input FFs over output FFs
 				}
+
 				SigSpec q = cell->getPort(ID::Q);
 				log_assert(GetSize(q) == 1);
-				if (modwalker.has_outputs(q)) {
+				if (modwalker.has_outputs(q) && !modwalker.has_consumers(q)) {
 					log_debug("Cell %s is potentially eligible for promotion to output IOFF.\n", cell->name.c_str());
-					// check that q_sig has no other consumers
-					pool<ModWalker::PortBit> portbits;
-					modwalker.get_consumers(portbits, q);
-					if (GetSize(portbits) > 0) {
-						log_debug("not promoting: Q has other consumers\n");
-						continue;
+					for (SigBit bit : output_bit_aliases[modwalker.sigmap(q)]) {
+						log_assert(bit.is_wire());
+						output_ffs[bit.wire][bit.offset] = cell;
 					}
-					cells_to_replace.insert(cell);
+
 				}
 			}
 		}
 
-		for (auto cell : cells_to_replace) {
-			log("Promoting register %s to IOFF.\n", log_signal(cell->getPort(ID::Q)));
+		for (auto cell : input_ffs) {
+			log("Promoting register %s to input IOFF.\n", log_signal(cell->getPort(ID::Q)));
 			cell->type = ID(dff);
 			cell->unsetPort(ID::E);
 			cell->unsetPort(ID::R);
 			cell->unsetPort(ID::S);
 		}
+		for (auto & [old_port_output, ioff_cells] : output_ffs) {
+			if (std::any_of(ioff_cells.begin(), ioff_cells.end(), [](Cell * c) { return c != nullptr; }))
+			{
+				// create replacement output wire
+				RTLIL::Wire* new_port_output = module->addWire(NEW_ID, old_port_output->width);
+				new_port_output->start_offset = old_port_output->start_offset;
+				module->swap_names(old_port_output, new_port_output);
+				std::swap(old_port_output->port_id, new_port_output->port_id);
+				std::swap(old_port_output->port_input, new_port_output->port_input);
+				std::swap(old_port_output->port_output, new_port_output->port_output);
+				std::swap(old_port_output->upto, new_port_output->upto);
+				std::swap(old_port_output->is_signed, new_port_output->is_signed);
+				std::swap(old_port_output->attributes, new_port_output->attributes);
+
+				// create new output FFs
+				SigSpec sig_o(old_port_output);
+				SigSpec sig_n(new_port_output);
+				for (int i = 0; i < new_port_output->width; i++) {
+					if (ioff_cells[i]) {
+						log("Promoting %s to output IOFF.\n", log_signal(sig_n[i]));
+
+						RTLIL::Cell *new_cell = module->addCell(NEW_ID, ID(dff));
+						new_cell->setPort(ID::C, ioff_cells[i]->getPort(ID::C));
+						new_cell->setPort(ID::D, ioff_cells[i]->getPort(ID::D));
+						new_cell->setPort(ID::Q, sig_n[i]);
+						new_cell->set_bool_attribute(ID::keep);
+					} else {
+						module->connect(sig_n[i], sig_o[i]);
+					}
+				}
+			}
+		}
 	}
 } QlIoffPass;
 
diff --git a/techlibs/quicklogic/synth_quicklogic.cc b/techlibs/quicklogic/synth_quicklogic.cc
index 16f660943..07ec769b5 100644
--- a/techlibs/quicklogic/synth_quicklogic.cc
+++ b/techlibs/quicklogic/synth_quicklogic.cc
@@ -334,9 +334,10 @@ struct SynthQuickLogicPass : public ScriptPass {
 			run("opt_lut");
 		}
 		
-		if (check_label("iomap", "(for qlf_k6n10f)") && (family == "qlf_k6n10f" || help_mode)) {
+		if (check_label("iomap", "(for qlf_k6n10f, skip if -noioff)") && (family == "qlf_k6n10f" || help_mode)) {
 			if (ioff || help_mode) {
-				run("ql_ioff", "(unless -noioff)");
+				run("ql_ioff");
+				run("opt_clean");
 			}
 		}
 
diff --git a/tests/arch/quicklogic/qlf_k6n10f/ioff.ys b/tests/arch/quicklogic/qlf_k6n10f/ioff.ys
index 3144eda13..da1fb2946 100644
--- a/tests/arch/quicklogic/qlf_k6n10f/ioff.ys
+++ b/tests/arch/quicklogic/qlf_k6n10f/ioff.ys
@@ -21,6 +21,21 @@ EOF
 synth_quicklogic -family qlf_k6n10f -top top
 select -assert-count 4 t:dff
 
+design -reset
+# test: acceptable for output IOFF promotion; duplicate output FF
+read_verilog <<EOF
+module top (input clk, input [3:0] a, output [3:0] o, output [3:0] p);
+    reg [3:0] r;
+    always @(posedge clk) begin
+        r <= ~a;
+    end
+    assign o = r;
+    assign p = r;
+endmodule
+EOF
+synth_quicklogic -family qlf_k6n10f -top top
+select -assert-count 8 t:dff
+
 design -reset
 # test: acceptable for input IOFF promotion
 read_verilog <<EOF
@@ -170,3 +185,25 @@ endmodule
 EOF
 synth_quicklogic -family qlf_k6n10f -top top
 select -assert-count 0 t:dff
+
+design -reset
+# test: duplicate registers driving multiple output ports
+read_verilog <<EOF
+module top (
+    input clk,
+    input en,
+    input [3:0] a,
+    output reg [3:0] o_1,
+	output wire [3:0] o_2
+);
+always @(posedge clk) begin
+    o_1[1:0] <= ~a[1:0];
+    if (en)
+        o_1[2] <= a[2];
+end
+always @(*) o_1[3] = a[3];
+assign o_2 = o_1;
+endmodule
+EOF
+synth_quicklogic -family qlf_k6n10f -top top
+select -assert-count 4 t:dff