pattern ice40_dsp

state <SigBit> clock
state <bool> clock_pol cd_signed o_lo
state <SigSpec> sigA sigB sigCD sigH sigO
state <Cell*> add mux
state <IdString> addAB muxAB

state <Cell*> ffA ffB ffCD
state <Cell*> ffFJKG ffH ffO

// subpattern
state <bool> argSdff
state <SigSpec> argQ argD
udata <SigSpec> dffD dffQ
udata <SigBit> dffclock
udata <Cell*> dff
udata <bool> dffclock_pol

match mul
	select mul->type.in($mul, \SB_MAC16)
	select GetSize(mul->getPort(\A)) + GetSize(mul->getPort(\B)) > 10
endmatch

code sigA sigB sigH
	auto unextend = [](const SigSpec &sig) {
		int i;
		for (i = GetSize(sig)-1; i > 0; i--)
			if (sig[i] != sig[i-1])
				break;
		// Do not remove sign bit
		++i;
		return sig.extract(0, i);
	};
	sigA = unextend(port(mul, \A));
	sigB = unextend(port(mul, \B));

	SigSpec O;
	if (mul->type == $mul)
		O = mul->getPort(\Y);
	else if (mul->type == \SB_MAC16)
		O = mul->getPort(\O);
	else log_abort();
	if (GetSize(O) <= 10)
		reject;

	// Only care about those bits that are used
	int i;
	for (i = GetSize(O) - 1; i > 0 && nusers(O[i]) <= 1; i--)
		;
	// This sigM could have no users if downstream sinks (e.g. $add) is
	//   narrower than $mul result, for example
	if (i == 0)
		reject;

	for (int j = 0, wire_width = 0; j <= i; j++) 
		if (nusers(O[j]) == 0)
			wire_width++;
		else {
			if (wire_width) { // add empty wires for bit offset if needed
				sigH.append(module->addWire(NEW_ID, wire_width));
				wire_width = 0;
			}
			sigH.append(O[j]);
		}

endcode

code argQ ffA sigA clock clock_pol
	if (mul->type != \SB_MAC16 || !param(mul, \A_REG).as_bool()) {
		argQ = sigA;
		subpattern(in_dffe);
		if (dff) {
			ffA = dff;
			clock = dffclock;
			clock_pol = dffclock_pol;
			sigA = dffD;
		}
	}
endcode

code argQ ffB sigB clock clock_pol
	if (mul->type != \SB_MAC16 || !param(mul, \B_REG).as_bool()) {
		argQ = sigB;
		subpattern(in_dffe);
		if (dff) {
			ffB = dff;
			clock = dffclock;
			clock_pol = dffclock_pol;
			sigB = dffD;
		}
	}
endcode

code argD argSdff ffFJKG sigH clock clock_pol
	if (nusers(sigH) == 2 &&
			(mul->type != \SB_MAC16 ||
			 (!param(mul, \TOP_8x8_MULT_REG).as_bool() && !param(mul, \BOT_8x8_MULT_REG).as_bool() && !param(mul, \PIPELINE_16x16_MULT_REG1).as_bool() && !param(mul, \PIPELINE_16x16_MULT_REG1).as_bool()))) {
		argD = sigH;
		argSdff = false;
		subpattern(out_dffe);
		if (dff) {
			// F/J/K/G do not have a CE-like (hold) input
			if (dff->hasPort(\EN))
				goto reject_ffFJKG;

			// Reset signal of F/J (IRSTTOP) and K/G (IRSTBOT)
			//   shared with A and B
			if (ffA) {
				if (ffA->hasPort(\ARST) != dff->hasPort(\ARST))
					goto reject_ffFJKG;
				if (ffA->hasPort(\ARST)) {
					if (port(ffA, \ARST) != port(dff, \ARST))
						goto reject_ffFJKG;
					if (param(ffA, \ARST_POLARITY) != param(dff, \ARST_POLARITY))
						goto reject_ffFJKG;
				}
			}
			if (ffB) {
				if (ffB->hasPort(\ARST) != dff->hasPort(\ARST))
					goto reject_ffFJKG;
				if (ffB->hasPort(\ARST)) {
					if (port(ffB, \ARST) != port(dff, \ARST))
						goto reject_ffFJKG;
					if (param(ffB, \ARST_POLARITY) != param(dff, \ARST_POLARITY))
						goto reject_ffFJKG;
				}
			}

			ffFJKG = dff;
			clock = dffclock;
			clock_pol = dffclock_pol;
			sigH = dffQ;

reject_ffFJKG: 		;
		}
	}
endcode

code argD argSdff ffH sigH sigO clock clock_pol
	if (ffFJKG && nusers(sigH) == 2 &&
			(mul->type != \SB_MAC16 || !param(mul, \PIPELINE_16x16_MULT_REG2).as_bool())) {
		argD = sigH;
		argSdff = false;
		subpattern(out_dffe);
		if (dff) {
			// H does not have a CE-like (hold) input
			if (dff->hasPort(\EN))
				goto reject_ffH;

			// Reset signal of H (IRSTBOT) shared with B
			if (ffB->hasPort(\ARST) != dff->hasPort(\ARST))
				goto reject_ffH;
			if (ffB->hasPort(\ARST)) {
				if (port(ffB, \ARST) != port(dff, \ARST))
					goto reject_ffH;
				if (param(ffB, \ARST_POLARITY) != param(dff, \ARST_POLARITY))
					goto reject_ffH;
			}

			ffH = dff;
			clock = dffclock;
			clock_pol = dffclock_pol;
			sigH = dffQ;

reject_ffH:		;
		}
	}

	sigO = sigH;
endcode

match add
	if mul->type != \SB_MAC16 || (param(mul, \TOPOUTPUT_SELECT).as_int() == 3 && param(mul, \BOTOUTPUT_SELECT).as_int() == 3)

	select add->type.in($add)
	choice <IdString> AB {\A, \B}
	select nusers(port(add, AB)) == 2

	index <SigBit> port(add, AB)[0] === sigH[0]
	filter GetSize(port(add, AB)) <= GetSize(sigH)
	filter port(add, AB) == sigH.extract(0, GetSize(port(add, AB)))
	filter nusers(sigH.extract_end(GetSize(port(add, AB)))) <= 1
	set addAB AB
	optional
endmatch

code sigCD sigO cd_signed
	if (add) {
		sigCD = port(add, addAB == \A ? \B : \A);
		cd_signed = param(add, addAB == \A ? \B_SIGNED : \A_SIGNED).as_bool();

		int natural_mul_width = GetSize(sigA) + GetSize(sigB);
		int actual_mul_width = GetSize(sigH);
		int actual_acc_width = GetSize(sigCD);

		if ((actual_acc_width > actual_mul_width) && (natural_mul_width > actual_mul_width))
			reject;
		// If accumulator, check adder width and signedness
		if (sigCD == sigH && (actual_acc_width != actual_mul_width) && (param(mul, \A_SIGNED).as_bool() != param(add, \A_SIGNED).as_bool()))
			reject;

		sigO = port(add, \Y);
	}
endcode

match mux
	select mux->type == $mux
	choice <IdString> AB {\A, \B}
	select nusers(port(mux, AB)) == 2
	index <SigSpec> port(mux, AB) === sigO
	set muxAB AB
	optional
endmatch

code sigO
	if (mux)
		sigO = port(mux, \Y);
endcode

code argD argSdff ffO sigO sigCD clock clock_pol cd_signed o_lo
	if (mul->type != \SB_MAC16 ||
			// Ensure that register is not already used
			((param(mul, \TOPOUTPUT_SELECT).as_int() != 1 && param(mul, \BOTOUTPUT_SELECT).as_int() != 1) &&
			 // Ensure that OLOADTOP/OLOADBOT is unused or zero
			 (port(mul, \OLOADTOP, State::S0).is_fully_zero() && port(mul, \OLOADBOT, State::S0).is_fully_zero()))) {

		dff = nullptr;

		// First try entire sigO
		if (nusers(sigO) == 2) {
			argD = sigO;
			argSdff = !mux;
			subpattern(out_dffe);
		}

		// Otherwise try just its least significant 16 bits
		if (!dff && GetSize(sigO) > 16) {
			argD = sigO.extract(0, 16);
			if (nusers(argD) == 2) {
				argSdff = !mux;
				subpattern(out_dffe);
				o_lo = dff;
			}
		}

		if (dff) {
			ffO = dff;
			clock = dffclock;
			clock_pol = dffclock_pol;

			sigO.replace(sigO.extract(0, GetSize(dffQ)), dffQ);
		}

		// Loading value into output register is not
		//   supported unless using accumulator
		if (mux) {
			if (sigCD != sigO)
				reject;
			sigCD = port(mux, muxAB == \B ? \A : \B);

			cd_signed = add && param(add, \A_SIGNED).as_bool() && param(add, \B_SIGNED).as_bool();
		} else if (dff && dff->hasPort(\SRST)) {
			if (sigCD != sigO)
				reject;
			sigCD = param(dff, \SRST_VALUE);

			cd_signed = add && param(add, \A_SIGNED).as_bool() && param(add, \B_SIGNED).as_bool();
		}
	}
endcode

code argQ ffCD sigCD clock clock_pol
	if (!sigCD.empty() && sigCD != sigO &&
			(mul->type != \SB_MAC16 || (!param(mul, \C_REG).as_bool() && !param(mul, \D_REG).as_bool()))) {
		argQ = sigCD;
		subpattern(in_dffe);
		if (dff) {
			// Reset signal of C (IRSTTOP) and D (IRSTBOT)
			//   shared with A and B
			if (ffA) {
				if (ffA->hasPort(\ARST) != dff->hasPort(\ARST))
					goto reject_ffCD;
				if (ffA->hasPort(\ARST)) {
					if (port(ffA, \ARST) != port(dff, \ARST))
						goto reject_ffCD;
					if (param(ffA, \ARST_POLARITY) != param(dff, \ARST_POLARITY))
						goto reject_ffCD;
				}
			}
			if (ffB) {
				if (ffB->hasPort(\ARST) != dff->hasPort(\ARST))
					goto reject_ffCD;
				if (ffB->hasPort(\ARST)) {
					if (port(ffB, \ARST) != port(dff, \ARST))
						goto reject_ffCD;
					if (param(ffB, \ARST_POLARITY) != param(dff, \ARST_POLARITY))
						goto reject_ffCD;
				}
			}

			ffCD = dff;
			clock = dffclock;
			clock_pol = dffclock_pol;
			sigCD = dffD;

reject_ffCD: 		;
		}
	}
endcode

code sigCD
	sigCD.extend_u0(32, cd_signed);
endcode

code
	accept;
endcode

// #######################

subpattern in_dffe
arg argD argQ clock clock_pol

code
	dff = nullptr;
	if (argQ.empty())
		reject;
	for (auto c : argQ.chunks()) {
		if (!c.wire)
			reject;
		if (c.wire->get_bool_attribute(\keep))
			reject;
		Const init = c.wire->attributes.at(\init, State::Sx);
		if (!init.is_fully_undef() && !init.is_fully_zero())
			reject;
	}
endcode

match ff
	select ff->type.in($dff, $dffe)
	// DSP48E1 does not support clock inversion
	select param(ff, \CLK_POLARITY).as_bool()

	slice offset GetSize(port(ff, \D))
	index <SigBit> port(ff, \Q)[offset] === argQ[0]

	// Check that the rest of argQ is present
	filter GetSize(port(ff, \Q)) >= offset + GetSize(argQ)
	filter port(ff, \Q).extract(offset, GetSize(argQ)) == argQ
endmatch

code argQ argD
{
	if (clock != SigBit()) {
		if (port(ff, \CLK)[0] != clock)
			reject;
		if (param(ff, \CLK_POLARITY).as_bool() != clock_pol)
			reject;
	}

	SigSpec Q = port(ff, \Q);
	dff = ff;
	dffclock = port(ff, \CLK);
	dffclock_pol = param(ff, \CLK_POLARITY).as_bool();
	dffD = argQ;
	argD = port(ff, \D);
	argQ = Q;
	dffD.replace(argQ, argD);
}
endcode

// #######################

subpattern out_dffe
arg argD argSdff argQ clock clock_pol

code
	dff = nullptr;
	for (auto c : argD.chunks())
		if (c.wire->get_bool_attribute(\keep))
			reject;
endcode

match ff
	select ff->type.in($dff, $dffe, $sdff, $sdffce)
	// SB_MAC16 does not support clock inversion
	select param(ff, \CLK_POLARITY).as_bool()

	slice offset GetSize(port(ff, \D))
	index <SigBit> port(ff, \D)[offset] === argD[0]

	// Only allow sync reset if requested.
	filter argSdff || ff->type.in($dff, $dffe)
	// Check that the rest of argD is present
	filter GetSize(port(ff, \D)) >= offset + GetSize(argD)
	filter port(ff, \D).extract(offset, GetSize(argD)) == argD
endmatch

code argQ
	if (ff) {
		if (clock != SigBit()) {
			if (port(ff, \CLK)[0] != clock)
				reject;
			if (param(ff, \CLK_POLARITY).as_bool() != clock_pol)
				reject;
		}
		SigSpec D = port(ff, \D);
		SigSpec Q = port(ff, \Q);
		argQ = argD;
		argQ.replace(D, Q);

		for (auto c : argQ.chunks()) {
			Const init = c.wire->attributes.at(\init, State::Sx);
			if (!init.is_fully_undef() && !init.is_fully_zero())
				reject;
		}

		dff = ff;
		dffQ = argQ;
		dffclock = port(ff, \CLK);
		dffclock_pol = param(ff, \CLK_POLARITY).as_bool();
	}
endcode