Merge pull request #3883 from phsauter/peepopt-shiftadd

peepopt: Add `shiftadd` pattern

passes/pmgen/Makefile.inc

@@ -44,6 +44,7 @@ $(eval $(call add_extra_objs,passes/pmgen/peepopt_pm.h))
 
 PEEPOPT_PATTERN  = passes/pmgen/peepopt_shiftmul_right.pmg
 PEEPOPT_PATTERN += passes/pmgen/peepopt_shiftmul_left.pmg
+PEEPOPT_PATTERN += passes/pmgen/peepopt_shiftadd.pmg
 PEEPOPT_PATTERN += passes/pmgen/peepopt_muldiv.pmg
 
 passes/pmgen/peepopt_pm.h: passes/pmgen/pmgen.py $(PEEPOPT_PATTERN)

passes/pmgen/peepopt.cc

@@ -48,6 +48,9 @@ struct PeepoptPass : public Pass {
 		log("                Analogously, replace A<<(B*C) with appropriate selection of\n");
 		log("                output bits from A<<(B<<K). (left variant)\n");
 		log("\n");
+		log("   * shiftadd - Replace A>>(B+D) with (A'>>D)>>(B) where D is constant and\n");
+		log("                A' is derived from A by padding or cutting inaccessible bits.\n");
+		log("\n");
 	}
 	void execute(std::vector<std::string> args, RTLIL::Design *design) override
 	{
@@ -72,6 +75,7 @@ struct PeepoptPass : public Pass {
 
 				pm.setup(module->selected_cells());
 
+				pm.run_shiftadd();
 				pm.run_shiftmul_right();
 				pm.run_shiftmul_left();
 				pm.run_muldiv();

passes/pmgen/peepopt_shiftadd.pmg

@@ -0,0 +1,121 @@
+pattern shiftadd
+//
+// Transforms add/sub+shift pairs that result from expressions such as data[s*W +C +:W2]
+// specifically something like: out[W2-1:0] = data >> (s*W +C)
+// will be transformed into: out[W2-1:0] = (data >> C) >> (s*W)
+// this can then be optimized using peepopt_shiftmul_right.pmg
+//
+
+match shift
+	select shift->type.in($shift, $shiftx, $shr)
+	filter !port(shift, \B).empty()
+endmatch
+
+// the right shift amount
+state <SigSpec> shift_amount
+// log2 scale factor in interpreting of shift_amount
+// due to zero padding on the shift cell's B port
+state <int> log2scale
+// zeros at the MSB position make it unsigned
+state <bool> msb_zeros
+
+code shift_amount log2scale msb_zeros
+	shift_amount = port(shift, \B);
+
+	log2scale = 0;
+	while (shift_amount[0] == State::S0) {
+		shift_amount.remove(0);
+		if (shift_amount.empty()) reject;
+		log2scale++;
+	}
+
+	msb_zeros = 0;
+	while (shift_amount.bits().back() == State::S0) {
+		msb_zeros = true;
+		shift_amount.remove(GetSize(shift_amount) - 1);
+		if (shift_amount.empty()) reject;
+	}
+endcode
+
+state <bool>    var_signed
+state <SigSpec> var_signal
+// offset: signed constant value c in data[var+c +:W1] (constant shift-right amount)
+state <int>     offset
+
+match add
+	// either data[var+c +:W1] or data[var-c +:W1]
+	select add->type.in($add, $sub)
+	index <SigSpec> port(add, \Y) === shift_amount
+
+	// one must be constant, the other is variable
+	choice <IdString> constport {\A, \B}
+	select !port(add, constport).empty()
+	select port(add, constport).is_fully_const()
+	define <IdString> varport (constport == \A ? \B : \A)
+
+	// if a value of var is able to wrap the output, the transformation might give wrong results
+	// an addition/substraction can at most flip one more bit than the largest operand (the carry bit)
+	// as long as the output can show this bit, no wrap should occur (assuming all signed-ness make sense)
+	select ( GetSize(port(add, \Y)) > max(GetSize(port(add, \A)), GetSize(port(add, \B))) )
+
+	define <bool> varport_A (varport == \A)
+	define <bool> is_sub add->type.in($sub)
+
+	define <bool> constport_signed param(add, !varport_A ? \A_SIGNED : \B_SIGNED).as_bool()
+	define <bool> varport_signed param(add, varport_A ? \A_SIGNED : \B_SIGNED).as_bool();
+	define <bool> offset_negative ((port(add, constport).bits().back() == State::S1) ^ (is_sub && varport_A))
+
+	// checking some value boundaries as well:
+	// data[...-c +:W1] is fine for +/-var (pad at LSB, all data still accessible)
+	// data[...+c +:W1] is only fine for +var(add) and var unsigned
+	// (+c cuts lower C bits, making them inaccessible, a signed var could try to access them)
+	// either its an add or the variable port is A (it must be positive)
+	select (add->type.in($add) || varport == \A)
+
+	// -> data[var+c +:W1] (with var signed) is illegal
+	filter !(!offset_negative && varport_signed)
+
+	// state-variables are assigned at the end only:
+	// shift the log2scale offset in-front of add to get true value: (var+c)<<N -> (var<<N)+(c<<N)
+	set offset ( (port(add, constport).as_int(constport_signed) << log2scale) * ( (is_sub && varport_A) ? -1 : 1 ) )
+
+	set var_signed varport_signed
+	set var_signal add->getPort(varport)
+endmatch
+
+code
+{
+	// positive constant offset with a signed variable (index) cannot be handled
+	// the above filter should get rid of this case but 'offset' is calculated differently
+	// due to limitations of state-variables in pmgen
+	// it should only differ if previous passes create invalid data
+	log_assert(!(offset>0 && var_signed));
+
+	did_something = true;
+	log("shiftadd pattern in %s: shift=%s, add/sub=%s, offset: %d\n", \
+			log_id(module), log_id(shift), log_id(add), offset);
+
+	SigSpec old_a = port(shift, \A), new_a;
+	if(offset<0) {
+		// data >> (...-c) transformed to {data, c'X} >> (...)
+		SigSpec padding( (shift->type.in($shiftx) ? State::Sx : State::S0), -offset );
+		new_a.append(padding);
+		new_a.append(old_a);
+	} else {
+		// data >> (...+c) transformed to data[MAX:c] >> (...)
+		new_a.append(old_a.extract_end(offset));
+
+	}
+
+	SigSpec new_b = {var_signal, SigSpec(State::S0, log2scale)};
+	if (msb_zeros || !var_signed)
+		new_b.append(State::S0);
+
+	shift->setPort(\A, new_a);
+	shift->setParam(\A_WIDTH, GetSize(new_a));
+	shift->setPort(\B, new_b);
+	shift->setParam(\B_WIDTH, GetSize(new_b));
+	blacklist(add);
+	accept;
+}
+endcode

tests/simple/partsel.v

@@ -110,3 +110,42 @@ module partsel_test007 (
 		dout[n+1] = din[n];
 	end
 endmodule
+
+
+module partsel_test008 (
+	input [127:0] din,
+	input [3:0] idx,
+	input        [4:0] uoffset,
+	input signed [4:0] soffset,
+	output [ 7:0] dout0,
+	output [ 7:0] dout1,
+	output [ 7:0] dout2,
+	output [ 7:0] dout3,
+	output [ 3:0] dout4,
+	output [ 3:0] dout5,
+	output [ 3:0] dout6,
+	output [ 3:0] dout7,
+	output [ 3:0] dout8,
+	output [11:0] dout9,
+	output [11:0] dout10,
+	output [11:0] dout11
+);
+
+// common: block-select with offsets
+assign dout0  = din[idx*8 +uoffset +:8];
+assign dout1  = din[idx*8 -uoffset +:8];
+assign dout2  = din[idx*8 +soffset +:8];
+assign dout3  = din[idx*8 -soffset +:8];
+
+// only partial block used
+assign dout4  = din[idx*8 +uoffset +:4];
+assign dout5  = din[idx*8 -uoffset +:4];
+assign dout6  = din[idx*8 +soffset +:4];
+assign dout7  = din[idx*8 -soffset +:4];
+
+// uncommon: more than one block used
+assign dout8  = din[idx*8 +uoffset +:12];
+assign dout9  = din[idx*8 -uoffset +:12];
+assign dout10 = din[idx*8 +soffset +:12];
+assign dout11 = din[idx*8 -soffset +:12];
+endmodule