// This file describes the second of three pattern matcher setups that
//   forms the `xilinx_dsp` pass described in xilinx_dsp.cc
// At a high level, it works as follows:
//   (1) Starting from a DSP48* cell that (a) doesn't have a CREG already,
//       and (b) uses the 'C' port
//   (2) Match the driver of the 'C' input to a possible $dff cell (CREG)
//       (attached to at most two $mux cells that implement clock-enable or
//        reset functionality, using a subpattern discussed below)
// Notes:
//   - Running CREG packing after xilinx_dsp_pack is necessary since there is no
//     guarantee that the cell ordering corresponds to the "expected" case (i.e.
//     the order in which they appear in the source) thus the possiblity existed
//     that a register got packed as a CREG into a downstream DSP that should
//     have otherwise been a PREG of an upstream DSP that had not been visited
//     yet
//   - The reason this is separated out from the xilinx_dsp.pmg file is
//     for efficiency --- each *.pmg file creates a class of the same basename,
//     which when constructed, creates a custom database tailored to the
//     pattern(s) contained within. Since the pattern in this file must be
//     executed after the pattern contained in xilinx_dsp.pmg, it is necessary
//     to reconstruct this database. Separating the two patterns into
//     independent files causes two smaller, more specific, databases.

pattern xilinx_dsp_packC

udata <std::function<SigSpec(const SigSpec&)>> unextend
state <SigBit> clock
state <SigSpec> sigC sigP
state <Cell*> ffC

// Variables used for subpatterns
state <SigSpec> argQ argD
state <int> ffoffset
udata <SigSpec> dffD dffQ
udata <SigBit> dffclock
udata <Cell*> dff

// (1) Starting from a DSP48* cell that (a) doesn't have a CREG already,
//     and (b) uses the 'C' port
match dsp
	select dsp->type.in(\DSP48A, \DSP48A1, \DSP48E1)
	select param(dsp, \CREG).as_int() == 0
	select nusers(port(dsp, \C, SigSpec())) > 1
endmatch

code sigC sigP clock
	unextend = [](const SigSpec &sig) {
		int i;
		for (i = GetSize(sig)-1; i > 0; i--)
			if (sig[i] != sig[i-1])
				break;
		// Do not remove non-const sign bit
		if (sig[i].wire)
			++i;
		return sig.extract(0, i);
	};
	sigC = unextend(port(dsp, \C, SigSpec()));

	SigSpec P = port(dsp, \P);
	if (!dsp->type.in(\DSP48E1) ||
            param(dsp, \USE_MULT).decode_string() == "MULTIPLY") {
		// Only care about those bits that are used
		int i;
		for (i = GetSize(P)-1; i >= 0; i--)
			if (nusers(P[i]) > 1)
				break;
		i++;
		log_assert(nusers(P.extract_end(i)) <= 1);
		sigP = P.extract(0, i);
	}
	else
		sigP = P;

	clock = port(dsp, \CLK, SigBit());
endcode

// (2) Match the driver of the 'C' input to a possible $dff cell (CREG)
//     (attached to at most two $mux cells that implement clock-enable or
//      reset functionality, using the in_dffe subpattern)
code argQ ffC sigC clock
	argQ = sigC;
	subpattern(in_dffe);
	if (dff) {
		ffC = dff;
		clock = dffclock;
		sigC = dffD;
	}
endcode

code
	if (ffC)
		accept;
endcode

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

// Subpattern for matching against input registers, based on knowledge of the
//   'Q' input.
subpattern in_dffe
arg argQ clock

code
	dff = nullptr;
	if (argQ.empty())
		reject;
	for (const auto &c : argQ.chunks()) {
		// Abandon matches when 'Q' is a constant
		if (!c.wire)
			reject;
		// Abandon matches when 'Q' has the keep attribute set
		if (c.wire->get_bool_attribute(\keep))
			reject;
		// Abandon matches when 'Q' has a non-zero init attribute set
		// (not supported by DSP48E1)
		Const init = c.wire->attributes.at(\init, Const());
		if (!init.empty())
			for (auto b : init.extract(c.offset, c.width))
				if (b != State::Sx && b != State::S0)
					reject;
	}
endcode

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

	// Check that reset value, if present, is fully 0.
	filter ff->type.in($dff, $dffe) || param(ff, \SRST_VALUE).is_fully_zero()

	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

	filter clock == SigBit() || port(ff, \CLK) == clock
endmatch

code argQ
	SigSpec Q = port(ff, \Q);
	dff = ff;
	dffclock = port(ff, \CLK);
	dffD = argQ;
	SigSpec D = port(ff, \D);
	argQ = Q;
	dffD.replace(argQ, D);
endcode