Fine tune ice40_dsp.pmg, add support for packing subsets of registers

4 files changed, 47 insertions, 35 deletions

diff --git a/passes/pmgen/ice40_dsp.cc b/passes/pmgen/ice40_dsp.cc
index 963a7d7a1..f6a701540 100644
--- a/passes/pmgen/ice40_dsp.cc
+++ b/passes/pmgen/ice40_dsp.cc
@@ -23,13 +23,16 @@
 USING_YOSYS_NAMESPACE
 PRIVATE_NAMESPACE_BEGIN
 
+template<class T> bool includes(const T &lhs, const T &rhs) {
+	return std::includes(lhs.begin(), lhs.end(), rhs.begin(), rhs.end());
+}
 #include "passes/pmgen/ice40_dsp_pm.h"
 
 void create_ice40_dsp(ice40_dsp_pm &pm)
 {
 	auto &st = pm.st_ice40_dsp;
 
-#if 0
+#if 1
 	log("\n");
 	log("ffA:   %s\n", log_id(st.ffA, "--"));
 	log("ffB:   %s\n", log_id(st.ffB, "--"));
@@ -100,7 +103,7 @@ void create_ice40_dsp(ice40_dsp_pm &pm)
 	cell->setPort("\\IRSTTOP", State::S0);
 	cell->setPort("\\IRSTBOT", State::S0);
 
-	if (st.clock_vld)
+	if (st.clock != SigBit())
 	{
 		cell->setPort("\\CLK", st.clock);
 		cell->setPort("\\CE", State::S1);
diff --git a/passes/pmgen/ice40_dsp.pmg b/passes/pmgen/ice40_dsp.pmg
index f2b7f2169..471b8b519 100644
--- a/passes/pmgen/ice40_dsp.pmg
+++ b/passes/pmgen/ice40_dsp.pmg
@@ -1,8 +1,9 @@
 pattern ice40_dsp
 
 state <SigBit> clock
-state <bool> clock_pol clock_vld
+state <bool> clock_pol
 state <SigSpec> sigA sigB sigY sigS
+state <SigSpec> sigYused
 state <Cell*> addAB muxAB
 
 match mul
@@ -13,68 +14,77 @@ endmatch
 
 match ffA
 	select ffA->type.in($dff)
-	// select nusers(port(ffA, \Q)) == 2
-	index <SigSpec> port(ffA, \Q) === port(mul, \A)
+	filter !port(mul, \A).remove_const().empty()
+	filter includes(port(ffA, \Q).to_sigbit_set(), port(mul, \A).remove_const().to_sigbit_set())
 	optional
 endmatch
 
-code sigA clock clock_pol clock_vld
+code sigA clock clock_pol
 	sigA = port(mul, \A);
 
 	if (ffA) {
-		sigA = port(ffA, \D);
+		sigA.replace(port(ffA, \Q), port(ffA, \D));
 
 		clock = port(ffA, \CLK).as_bit();
 		clock_pol = param(ffA, \CLK_POLARITY).as_bool();
-		clock_vld = true;
 	}
 endcode
 
 match ffB
 	select ffB->type.in($dff)
-	// select nusers(port(ffB, \Q)) == 2
-	index <SigSpec> port(ffB, \Q) === port(mul, \B)
+	filter !port(mul, \B).remove_const().empty()
+	filter includes(port(ffB, \Q).to_sigbit_set(), port(mul, \B).remove_const().to_sigbit_set())
 	optional
 endmatch
 
-code sigB clock clock_pol clock_vld
+code sigB clock clock_pol
 	sigB = port(mul, \B);
 
 	if (ffB) {
-		sigB = port(ffB, \D);
+		sigB.replace(port(ffB, \Q), port(ffB, \D));
+		
 		SigBit c = port(ffB, \CLK).as_bit();
 		bool cp = param(ffB, \CLK_POLARITY).as_bool();
 
-		if (clock_vld && (c != clock || cp != clock_pol))
+		if (clock != SigBit() && (c != clock || cp != clock_pol))
 			reject;
 
 		clock = c;
 		clock_pol = cp;
-		clock_vld = true;
 	}
 endcode
 
+// Extract the bits of Y that actually have a consumer
+// (as opposed to being a sign extension)
+code sigY sigYused
+	sigY = port(mul, \Y);
+	int i;
+	for (i = GetSize(sigY); i > 0; i--)
+		if (nusers(sigY[i-1]) > 1)
+			break;
+	sigYused = sigY.extract(0, i).remove_const();
+endcode
+
 match ffY
 	select ffY->type.in($dff)
 	select nusers(port(ffY, \D)) == 2
-	index <SigSpec> port(ffY, \D) === port(mul, \Y)
+	filter param(ffY, \WIDTH).as_int() >= GetSize(sigYused)
+	filter includes(port(ffY, \D).to_sigbit_set(), sigYused.to_sigbit_set())
 	optional
 endmatch
 
-code sigY clock clock_pol clock_vld
-	sigY = port(mul, \Y);
-
+code clock clock_pol sigY
 	if (ffY) {
-		sigY = port(ffY, \Q);
+		sigY.replace(port(ffY, \D), port(ffY, \Q));
+
 		SigBit c = port(ffY, \CLK).as_bit();
 		bool cp = param(ffY, \CLK_POLARITY).as_bool();
 
-		if (clock_vld && (c != clock || cp != clock_pol))
+		if (clock != SigBit() && (c != clock || cp != clock_pol))
 			reject;
 
 		clock = c;
 		clock_pol = cp;
-		clock_vld = true;
 	}
 endcode
 
@@ -147,16 +157,15 @@ match ffS
 	index <SigSpec> port(ffS, \Q) === sigS
 endmatch
 
-code clock clock_pol clock_vld
+code clock clock_pol
 	if (ffS) {
 		SigBit c = port(ffS, \CLK).as_bit();
 		bool cp = param(ffS, \CLK_POLARITY).as_bool();
 
-		if (clock_vld && (c != clock || cp != clock_pol))
+		if (clock != SigBit() && (c != clock || cp != clock_pol))
 			reject;
 
 		clock = c;
 		clock_pol = cp;
-		clock_vld = true;
 	}
 endcode
diff --git a/passes/pmgen/xilinx_dsp.cc b/passes/pmgen/xilinx_dsp.cc
index c71ac5ef8..d87d63670 100644
--- a/passes/pmgen/xilinx_dsp.cc
+++ b/passes/pmgen/xilinx_dsp.cc
@@ -39,7 +39,7 @@ void pack_xilinx_dsp(xilinx_dsp_pm &pm)
 	log("dsp:   %s\n", log_id(st.dsp, "--"));
 	log("ffP:   %s\n", log_id(st.ffP, "--"));
 	log("muxP:  %s\n", log_id(st.muxP, "--"));
-	log("P_used: %s\n", log_signal(st.P_used));
+	log("sigPused: %s\n", log_signal(st.sigPused));
 	log_module(pm.module);
 #endif
 
diff --git a/passes/pmgen/xilinx_dsp.pmg b/passes/pmgen/xilinx_dsp.pmg
index 7a175123e..a97ab4dd5 100644
--- a/passes/pmgen/xilinx_dsp.pmg
+++ b/passes/pmgen/xilinx_dsp.pmg
@@ -1,7 +1,7 @@
 pattern xilinx_dsp
 
 state <SigBit> clock
-state <SigSpec> P_used
+state <SigSpec> sigPused
 
 match dsp
 	select dsp->type.in(\DSP48E1)
@@ -43,23 +43,23 @@ endcode
 
 // Extract the bits of P that actually have a consumer
 // (as opposed to being a sign extension)
-code P_used
+code sigPused
 	SigSpec P = port(dsp, \P);
 	int i;
 	for (i = GetSize(P); i > 0; i--)
 		if (nusers(P[i-1]) > 1)
 			break;
-	P_used = P.extract(0, i).remove_const();
+	sigPused = P.extract(0, i).remove_const();
 endcode
 
 match ffP
-	if !P_used.empty()
+	if !sigPused.empty()
 	select ffP->type.in($dff, $dffe)
 	select nusers(port(ffP, \D)) == 2
 	// DSP48E1 does not support clock inversion
 	select param(ffP, \CLK_POLARITY).as_bool()
-	filter param(ffP, \WIDTH).as_int() >= GetSize(P_used)
-	filter includes(port(ffP, \D).to_sigbit_set(), P_used.to_sigbit_set())
+	filter param(ffP, \WIDTH).as_int() >= GetSize(sigPused)
+	filter includes(port(ffP, \D).to_sigbit_set(), sigPused.to_sigbit_set())
 	optional
 endmatch
 
@@ -68,12 +68,12 @@ endmatch
 //   since that would lose information helpful for
 //   efficient wide-mux inference
 match muxP
-	if !P_used.empty() && !ffP
+	if !sigPused.empty() && !ffP
 	select muxP->type.in($mux)
 	select nusers(port(muxP, \B)) == 2
 	select port(muxP, \A).is_fully_undef()
-	filter param(muxP, \WIDTH).as_int() >= GetSize(P_used)
-	filter includes(port(muxP, \B).to_sigbit_set(), P_used.to_sigbit_set())
+	filter param(muxP, \WIDTH).as_int() >= GetSize(sigPused)
+	filter includes(port(muxP, \B).to_sigbit_set(), sigPused.to_sigbit_set())
 	optional
 endmatch
 
@@ -83,7 +83,7 @@ match ffY
 	select nusers(port(ffY, \D)) == 2
 	// DSP48E1 does not support clock inversion
 	select param(ffY, \CLK_POLARITY).as_bool()
-	filter param(ffY, \WIDTH).as_int() >= GetSize(P_used)
+	filter param(ffY, \WIDTH).as_int() >= GetSize(sigPused)
 	filter includes(port(ffY, \D).to_sigbit_set(), port(muxP, \Y).to_sigbit_set())
 endmatch