pattern shiftpow2
//
// Optimize shifts that result from expressions such as foo[s*W+:W]
// when W is a power of two and the multiply has been folded away.
//

match shift
	select shift->type.in($shift, $shiftx, $shr)
	filter !port(shift, \B).empty()
endmatch

code
{
	// make sure the shift amount cannot be negative
	SigSpec amount = port(shift, \B);
	bool b_signed = shift->type.in($shift, $shiftx) && param(shift, \B_SIGNED).as_bool();
	if (!b_signed)
		amount.append(State::S0);
	if (amount.bits().back() != State::S0)
		reject;

	while (GetSize(amount) > 1 && amount.bits().back() == State::S0)
		amount.remove(GetSize(amount) - 1);

	// low zero bits encode the power-of-two scale
	int log2scale = 0;
	while (!amount.empty() && amount[0] == State::S0) {
		amount.remove(0);
		log2scale++;
	}

	if (log2scale < 1)
		reject;

	if (amount.empty() || amount.is_fully_const())
		reject;

	SigSpec sel = amount;
	int sel_width = GetSize(sel);
	int width = param(shift, \Y_WIDTH).as_int();
	if (log2scale >= 8 * (int)sizeof(int) - 1)
		reject;
	int stride = 1 << log2scale;

	// avoid overlapping selections
	if (width > stride)
		reject;

	if (sel_width > 20)
		reject;
	long long ways = 1LL << sel_width;

	SigSpec A = port(shift, \A);
	int a_width = GetSize(A);
	bool a_signed = !shift->type.in($shiftx) && param(shift, \A_SIGNED).as_bool();
	int extended_a_width = a_signed ? std::max(a_width, width) : a_width;

	// limit padding for out-of-range select values
	int max_ratio = module->design->scratchpad_get_int("peepopt.shiftpow2.max_data_multiple", 2);
	if (ways * (long long)width > (long long)max_ratio * std::max(a_width, width))
		reject;

	did_something = true;
	log("shiftpow2 pattern in %s: shift=%s, index=%s, stride=%d, width=%d, ways=%lld\n",
			module, shift, log_signal(sel), stride, width, ways);

	// way m holds A[m*stride +: width], way 0 in the LSBs
	State fill = shift->type.in($shiftx) ? State::Sx : State::S0;
	SigSpec bmux_a;
	for (long long m = 0; m < ways; m++) {
		long long base = m * (long long)stride;
		for (int b = 0; b < width; b++) {
			long long idx = base + b;
			if (idx < a_width)
				bmux_a.append(A[idx]);
			else if (idx < extended_a_width)
				bmux_a.append(A.back());
			else
				bmux_a.append(fill);
		}
	}

	module->addBmux(NEW_ID, bmux_a, sel, port(shift, \Y));
	autoremove(shift);
	accept;
}
endcode