Add muldiv_c and muxadd peepopts

Makefile

@@ -854,6 +854,7 @@ SH_TEST_DIRS += tests/bram
 SH_TEST_DIRS += tests/svinterfaces
 SH_TEST_DIRS += tests/xprop
 SH_TEST_DIRS += tests/select
+SH_TEST_DIRS += tests/peepopt
 SH_TEST_DIRS += tests/proc
 SH_TEST_DIRS += tests/blif
 SH_TEST_DIRS += tests/arch

passes/pmgen/Makefile.inc

@@ -57,6 +57,8 @@ 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
+PEEPOPT_PATTERN += passes/pmgen/peepopt_muldiv_c.pmg
+PEEPOPT_PATTERN += passes/pmgen/peepopt_muxadd.pmg
 PEEPOPT_PATTERN += passes/pmgen/peepopt_formal_clockgateff.pmg
 
 passes/pmgen/peepopt_pm.h: passes/pmgen/pmgen.py $(PEEPOPT_PATTERN)

passes/pmgen/peepopt.cc

@@ -28,6 +28,9 @@ bool did_something;
 // scratchpad configurations for pmgen
 int shiftadd_max_ratio;
 
+// Helper function, removes LSB 0s
+SigSpec remove_bottom_padding(SigSpec sig);
+
 #include "passes/pmgen/peepopt_pm.h"
 
 struct PeepoptPass : public Pass {
@@ -44,6 +47,8 @@ struct PeepoptPass : public Pass {
 		log("\n");
 		log("   * muldiv - Replace (A*B)/B with A\n");
 		log("\n");
+		log("   * muldiv_c - Replace (A*B)/C with A*(B/C) when C is a const divisible by B.\n");
+		log("\n");
 		log("   * shiftmul - Replace A>>(B*C) with A'>>(B<<K) where C and K are constants\n");
 		log("                and A' is derived from A by appropriately inserting padding\n");
 		log("                into the signal. (right variant)\n");
@@ -63,20 +68,28 @@ struct PeepoptPass : public Pass {
 		log("                   based pattern to prevent combinational paths from the\n");
 		log("                   output to the enable input after running clk2fflogic.\n");
 		log("\n");
+		log("If -withmuxadd is specified it adds the following rule:\n");
+		log("\n");
+		log("   * muxadd - Replace S?(A+B):A with A+(S?B:0)\n");
+		log("\n");
 	}
 	void execute(std::vector<std::string> args, RTLIL::Design *design) override
 	{
 		log_header(design, "Executing PEEPOPT pass (run peephole optimizers).\n");
 
 		bool formalclk = false;
-
+		bool withmuxadd = false;
 		size_t argidx;
 		for (argidx = 1; argidx < args.size(); argidx++)
 		{
 			if (args[argidx] == "-formalclk") {
 				formalclk = true;
 				continue;
 			}
+			if (args[argidx] == "-withmuxadd") {
+				withmuxadd = true;
+				continue;
+			}
 			break;
 		}
 		extra_args(args, argidx, design);
@@ -105,10 +118,22 @@ struct PeepoptPass : public Pass {
 					pm.run_shiftmul_right();
 					pm.run_shiftmul_left();
 					pm.run_muldiv();
+					pm.run_muldiv_c();
+					if (withmuxadd)
+						pm.run_muxadd();
 				}
 			}
 		}
 	}
 } PeepoptPass;
 
+
+SigSpec remove_bottom_padding(SigSpec sig)
+{
+	int i = 0;
+	for (; i < sig.size() - 1 && sig[i] == State::S0; i++) {
+	}
+	return sig.extract(i, sig.size() - i);
+}
+
 PRIVATE_NAMESPACE_END

passes/pmgen/peepopt_muldiv_c.pmg

@@ -0,0 +1,125 @@
+pattern muldiv_c
+//
+// Authored by Akash Levy and Alain Dargelas of Silimate, Inc. under ISC license.
+// Transforms mul->div into const->mul when b and c are divisible constants:
+// y = (a * b_const) / c_const   ===>   a * eval(b_const / c_const)
+//
+
+state <SigSpec> a b_const mul_y
+
+match mul
+	// Select multiplier
+	select mul->type == $mul
+endmatch
+
+code a b_const mul_y
+	// Get multiplier signals
+	a = port(mul, \A);
+	b_const = port(mul, \B);
+	mul_y = port(mul, \Y);
+
+	// Fanout of each multiplier Y bit should be 1 (no bit-split)
+	if (nusers(mul_y) != 2)
+		reject;
+
+	// A and B can be interchanged
+	branch;
+	std::swap(a, b_const);
+endcode
+
+match div
+	// Select div of form (a * b_const) / c_const
+	select div->type == $div
+
+	// Check that b_const and c_const is constant
+	filter b_const.is_fully_const()
+	filter port(div, \B).is_fully_const()
+	index <SigSpec> remove_bottom_padding(port(div, \A)) === mul_y
+endmatch
+
+code
+	// Get div signals
+	SigSpec div_a = port(div, \A);
+	SigSpec c_const = port(div, \B);
+	SigSpec div_y = port(div, \Y);
+
+	// Get offset of multiplier result chunk in divider
+	int offset = GetSize(div_a) - GetSize(mul_y);
+
+	// Get properties and values of b_const and c_const
+	// b_const may be coming from the A port
+	// But it is an RTLIL invariant that A_SIGNED equals B_SIGNED
+	bool b_const_signed = mul->getParam(ID::B_SIGNED).as_bool();
+	bool c_const_signed = div->getParam(ID::B_SIGNED).as_bool();
+	int b_const_int = b_const.as_int(b_const_signed);
+	int c_const_int = c_const.as_int(c_const_signed);
+	int b_const_int_shifted = b_const_int << offset;
+
+	// Helper lambdas for two's complement math	
+	auto sign2sComplement = [](auto value, int numBits) {
+  		if (value & (1 << (numBits - 1))) {
+   			return -1; 
+  		} else {
+   		    return 1; 
+  		}
+	};
+	auto twosComplement = [](auto value, int numBits) {
+  		if (value & (1 << (numBits - 1))) {
+   			return (~value) + 1; // invert bits before adding 1
+  		} else {
+   		    return value; 
+  		}
+	};
+
+	// Two's complement conversion
+	if (b_const_signed)
+		b_const_int = sign2sComplement(b_const_int, GetSize(b_const)) * twosComplement(b_const_int, GetSize(b_const));
+	if (c_const_signed)
+		c_const_int = sign2sComplement(c_const_int, GetSize(c_const)) * twosComplement(c_const_int, GetSize(c_const));
+	// Calculate the constant and compress the width to fit the value
+	Const const_ratio;
+	Const b_const_actual;
+	// Avoid division by zero
+	if (c_const_int == 0)
+		reject;
+	b_const_actual = b_const_int_shifted;
+	b_const_actual.compress(b_const_signed);
+
+	const_ratio = b_const_int_shifted / c_const_int;
+	const_ratio.compress(b_const_signed | c_const_signed);
+
+	// Integer values should be lesser than 32 bits
+	// This is because we are using C++ types, and int is 32 bits
+	// FIXME: use long long or BigInteger to make pass work with >32 bits
+	if (GetSize(mul->getParam(ID::B_WIDTH)) > 32)
+		reject;
+	if (GetSize(b_const) > 32)
+		reject;
+	if (GetSize(c_const) + offset > 32)
+		reject;
+
+	// Check for potential multiplier overflow
+	if (GetSize(b_const_actual) + GetSize(a) > GetSize(mul_y))
+		reject;
+
+	// Check that there are only zeros before offset
+	if (offset < 0 || !div_a.extract(0, offset).is_fully_zero())
+		reject;
+
+	// Check that b is divisible by c
+	if (b_const_int_shifted % c_const_int != 0)
+		reject;
+
+	// Rewire to only keep multiplier
+	mul->setPort(\A, a);
+	mul->setPort(\B, const_ratio);
+	mul->setPort(\Y, div_y);
+
+	// Remove divider
+	autoremove(div);
+
+	// Log, fixup, accept
+	log("muldiv_const pattern in %s: mul=%s, div=%s\n", log_id(module), log_id(mul), log_id(div));
+	mul->fixup_parameters();
+	accept;
+endcode

passes/pmgen/peepopt_muxadd.pmg

@@ -0,0 +1,123 @@
+pattern muxadd
+//
+// Authored by Akash Levy and Alain Dargelas of Silimate, Inc. under ISC license.
+// Transforms add->mux into mux->add:
+// y = s ? (a + b) : a   ===>   y = a + (s ? b : 0)
+// or
+// y = s ? a : (a + b)   ===>   y = a + (s ? 0 : b)
+
+state <SigSpec> add_a add_b add_y add_a_ext mux_a mux_b mux_y
+state <Const> add_a_signed
+state <IdString> add_a_id add_b_id mux_a_id mux_b_id
+
+match add
+	// Select adder
+	select add->type == $add
+
+	// Set ports, allowing A and B to be swapped
+	choice <IdString> A {\A, \B}
+	define <IdString> B (A == \A ? \B : \A)
+	set add_a port(add, A)
+	set add_b port(add, B)
+	set add_y port(add, \Y)
+
+	// Get signedness
+	set add_a_signed param(add, (A == \A) ? \A_SIGNED : \B_SIGNED)
+
+	// Choice ids
+	set add_a_id A
+	set add_b_id B
+endmatch
+
+code add_y add_a add_b add_a_ext
+	// Get adder signals
+	add_a_ext = SigSpec(port(add, add_a_id));
+	add_a_ext.extend_u0(GetSize(add_y), add_a_signed.as_bool());
+
+	// Fanout of each adder Y bit should be 1 (no bit-split)
+	if (nusers(add_y) != 2)
+		reject;
+endcode
+
+match mux 
+	// Select mux of form: s ? (a + b) : a
+	// Allow leading 0s when A_WIDTH != Y_WIDTH or s ? a : (a + b)
+	select mux->type == $mux
+	choice <IdString> AB {\A, \B}
+	define <IdString> BA (AB == \A ? \B : \A)
+	set mux_y port(mux, \Y)
+	set mux_a port(mux, AB)
+	set mux_b port(mux, BA)
+	set mux_a_id AB
+	set mux_b_id BA
+	index <SigSpec> port(mux, AB) === add_a_ext
+	index <SigSpec> port(mux, BA) === add_y
+endmatch
+
+code add_y add_a add_b add_a_ext add_a_id add_b_id  mux_y mux_a mux_b mux_a_id mux_b_id
+	// Get mux signal
+	SigSpec mid; 
+	std::string adder_y_name;
+	if (add_y.is_wire()) 
+		adder_y_name = add_y.as_wire()->name.c_str();
+	else
+		adder_y_name = add_y.as_string();
+
+	// Start by renaming the LHS of an eventual assign statement
+	// where the RHS is the adder output (that is getting rewired).
+	// Renaming the signal allows equiv_opt to function as it would
+	// otherwise try to match the functionality which would fail
+	// as the LHS signal has indeed changed function.
+
+	// Adder output could be assigned
+	for (auto it = module->connections().begin(); it != module->connections().end(); ++it) {
+		RTLIL::SigSpec rhs = it->second;
+		if (rhs.is_wire()) {
+			const std::string& rhs_name = rhs.as_wire()->name.c_str();
+			if (rhs_name == adder_y_name) {
+				RTLIL::SigSpec lhs = it->first;
+				if (lhs.is_wire()) {
+					const std::string& lhs_name =  lhs.as_wire()->name.c_str();
+					module->rename(lhs_name, module->uniquify("$" + lhs_name));
+					break;
+				}
+			}
+		}
+	}
+	// Alternatively, the port name could be a wire name
+	if (add_y.is_wire()) {
+		if (GetSize(adder_y_name)) {
+			if (adder_y_name[0] != '$') {
+				module->rename(adder_y_name, module->uniquify("$" + adder_y_name));
+			}
+		}
+	} else {
+		for (auto chunk : add_y.chunks()) {
+			if (chunk.is_wire()) {
+				const std::string& name = chunk.wire->name.c_str();
+				if (name[0] != '$') {
+					module->rename(name, module->uniquify("$" + name));
+				}
+			}
+		}
+	}
+
+	// Create new mid wire
+	mid = module->addWire(NEW_ID, GetSize(add_b));
+
+	// Connect ports
+	add->setPort(add_b_id, mid);
+	add->setPort(add_a_id, add_a);
+	add->setPort(\Y, add_y);
+	mux->setPort(mux_a_id, Const(State::S0, GetSize(add_b)));
+	mux->setPort(mux_b_id, add_b);
+	mux->setPort(\Y, mid);
+	module->connect(mux_y, add_y);
+
+	// Log, fixup, accept
+	log("muxadd pattern in %s: mux=%s, add=%s\n", log_id(module), log_id(mux), log_id(add));
+	add->fixup_parameters();
+	mux->fixup_parameters();
+	did_something = true;
+	accept;
+endcode

tests/peepopt/.gitignore

@@ -0,0 +1 @@
+/*.log