Merge branch 'master' into mwk/xilinx_bufgmap

8 files changed, 405 insertions, 60 deletions

diff --git a/CHANGELOG b/CHANGELOG
index b68f0bbfa..f7774cb9e 100644
--- a/CHANGELOG
+++ b/CHANGELOG
@@ -12,7 +12,10 @@ Yosys 0.9 .. Yosys 0.9-dev
     - Added "synth_xilinx -abc9" (experimental)
     - Added "synth_ice40 -abc9" (experimental)
     - Added "synth -abc9" (experimental)
-    - Added "script -scriptwire
+    - Added "script -scriptwire"
+    - Added "synth_xilinx -nocarry"
+    - Added "synth_xilinx -nowidelut"
+    - Added "synth_ecp5 -nowidelut"
     - "synth_xilinx" to now infer wide multiplexers (-widemux <min> to enable)
     - Renamed labels/options in synth_ice40 (e.g. dram -> map_lutram; -nodram -> -nolutram)
     - Renamed labels/options in synth_ecp5 (e.g. dram -> map_lutram; -nodram -> -nolutram)
@@ -32,33 +35,126 @@ Yosys 0.9 .. Yosys 0.9-dev
     - Added "opt_share" pass, run as part of "opt -full"
     - Added "ice40_wrapcarry" to encapsulate SB_LUT+SB_CARRY pairs for techmapping
     - Removed "ice40_unlut"
+    - Improvements in pmgen: slices, choices, define, generate
 
-Yosys 0.8 .. Yosys 0.8-dev
---------------------------
+Yosys 0.8 .. Yosys 0.9
+----------------------
 
  * Various
-    - Added $changed support to read_verilog
+    - Many bugfixes and small improvements
+    - Added support for SystemVerilog interfaces and modports
     - Added "write_edif -attrprop"
-    - Added "ice40_unlut" pass
     - Added "opt_lut" pass
-    - Added "synth_ice40 -relut"
-    - Added "synth_ice40 -noabc"
     - Added "gate2lut.v" techmap rule
     - Added "rename -src"
     - Added "equiv_opt" pass
-    - Added "shregmap -tech xilinx"
+    - Added "flowmap" LUT mapping pass
+    - Added "rename -wire" to rename cells based on the wires they drive
+    - Added "bugpoint" for creating minimised testcases
+    - Added "write_edif -gndvccy"
+    - "write_verilog" to escape Verilog keywords
+    - Fixed sign handling of real constants
+    - "write_verilog" to write initial statement for initial flop state
+    - Added pmgen pattern matcher generator
+    - Fixed opt_rmdff handling of $_DFFSR_???_ and $_DLATCHSR_???_
+    - Added "setundef -params" to replace undefined cell parameters
+    - Renamed "yosys -D" to "yosys -U", added "yosys -D" to set Verilog defines
+    - Fixed handling of defparam when default_nettype is none
+    - Fixed "wreduce" flipflop handling
+    - Fixed FIRRTL to Verilog process instance subfield assignment
+    - Added "write_verilog -siminit"
+    - Several fixes and improvements for mem2reg memories
+    - Fixed handling of task output ports in clocked always blocks
+    - Improved handling of and-with-1 and or-with-0 in "opt_expr"
     - Added "read_aiger" frontend
+    - Added "mutate" pass
+    - Added "hdlname" attribute
+    - Added "rename -output"
+    - Added "read_ilang -lib"
+    - Improved "proc" full_case detection and handling
+    - Added "whitebox" and "lib_whitebox" attributes
+    - Added "read_verilog -nowb", "flatten -wb" and "wbflip"
+    - Added Python bindings and support for Python plug-ins
+    - Added "pmux2shiftx"
+    - Added log_debug framework for reduced default verbosity
+    - Improved "opt_expr" and "opt_clean" handling of (partially) undriven and/or unused wires
+    - Added "peepopt" peephole optimisation pass using pmgen
+    - Added approximate support for SystemVerilog "var" keyword
+    - Added parsing of "specify" blocks into $specrule and $specify[23]
+    - Added support for attributes on parameters and localparams
+    - Added support for parsing attributes on port connections
+    - Added "wreduce -keepdc"
+    - Added support for optimising $dffe and $_DFFE_* cells in "opt_rmdff"
+    - Added Verilog wand/wor wire type support
+    - Added support for elaboration system tasks
     - Added "muxcover -mux{4,8,16}=<cost>"
     - Added "muxcover -dmux=<cost>"
     - Added "muxcover -nopartial"
     - Added "muxpack" pass
     - Added "pmux2shiftx -norange"
+    - Added support for "~" in filename parsing
+    - Added "read_verilog -pwires" feature to turn parameters into wires
+    - Fixed sign extension of unsized constants with 'bx and 'bz MSB
+    - Fixed genvar to be a signed type
+    - Added support for attributes on case rules
+    - Added "upto" and "offset" to JSON frontend and backend
+    - Several liberty file parser improvements
+    - Fixed handling of more complex BRAM patterns
+    - Add "write_aiger -I -O -B"
+
+ * Formal Verification
+    - Added $changed support to read_verilog
+    - Added "read_verilog -noassert -noassume -assert-assumes"
+    - Added btor ops for $mul, $div, $mod and $concat
+    - Added yosys-smtbmc support for btor witnesses
+    - Added "supercover" pass
+    - Fixed $global_clock handling vs autowire
+    - Added $dffsr support to "async2sync"
+    - Added "fmcombine" pass
+    - Added memory init support in "write_btor"
+    - Added "cutpoint" pass
+    - Changed "ne" to "neq" in btor2 output
+    - Added support for SVA "final" keyword
+    - Added "fmcombine -initeq -anyeq"
+    - Added timescale and generated-by header to yosys-smtbmc vcd output
+    - Improved BTOR2 handling of undriven wires
+
+ * Verific support
+    - Enabled Verific flags vhdl_support_variable_slice and veri_elaborate_top_level_modules_having_interface_ports
+    - Improved support for asymmetric memories
+    - Added "verific -chparam"
+    - Fixed "verific -extnets" for more complex situations
+    - Added "read -verific" and "read -noverific"
+    - Added "hierarchy -chparam"
+
+ * New back-ends
+    - Added initial Anlogic support
+    - Added initial SmartFusion2 and IGLOO2 support
+
+ * ECP5 support
+    - Added "synth_ecp5 -nowidelut"
+    - Added BRAM inference support to "synth_ecp5"
+    - Added support for transforming Diamond IO and flipflop primitives
+
+ * iCE40 support
+    - Added "ice40_unlut" pass
+    - Added "synth_ice40 -relut"
+    - Added "synth_ice40 -noabc"
+    - Added "synth_ice40 -dffe_min_ce_use"
+    - Added DSP inference support using pmgen
+    - Added support for initialising BRAM primitives from a file
+    - Added iCE40 Ultra RGB LED driver cells
+
+ * Xilinx support
+    - Use "write_edif -pvector bra" for Xilinx EDIF files
+    - Fixes for VPR place and route support with "synth_xilinx"
+    - Added more cell simulation models
+    - Added "synth_xilinx -family"
+    - Added "stat -tech xilinx" to estimate logic cell usage
     - Added "synth_xilinx -nocarry"
     - Added "synth_xilinx -nowidelut"
-    - Added "synth_ecp5 -nowidelut"
     - "synth_xilinx" to now infer hard shift registers (-nosrl to disable)
-    - Fixed sign extension of unsized constants with 'bx and 'bz MSB
-
+    - Added support for mapping RAM32X1D
 
 Yosys 0.7 .. Yosys 0.8
 ----------------------
diff --git a/Makefile b/Makefile
index a742f2e50..48a4f3a1c 100644
--- a/Makefile
+++ b/Makefile
@@ -115,7 +115,7 @@ LDFLAGS += -rdynamic
 LDLIBS += -lrt
 endif
 
-YOSYS_VER := 0.8+$(shell cd $(YOSYS_SRC) && test -e .git && { git log --author=clifford@clifford.at --oneline 4d4665b.. 2> /dev/null | wc -l; })
+YOSYS_VER := 0.9+1
 GIT_REV := $(shell cd $(YOSYS_SRC) && git rev-parse --short HEAD 2> /dev/null || echo UNKNOWN)
 OBJS = kernel/version_$(GIT_REV).o
 
diff --git a/misc/py_wrap_generator.py b/misc/py_wrap_generator.py
index 66d661fa1..2bf364470 100644
--- a/misc/py_wrap_generator.py
+++ b/misc/py_wrap_generator.py
@@ -508,23 +508,17 @@ class TupleTranslator(PythonDictTranslator):
 	#Generate c++ code to translate to a boost::python::tuple
 	@classmethod
 	def translate_cpp(c, varname, types, prefix, ref):
-		text  = prefix + TupleTranslator.typename + " " + varname + "___tmp = boost::python::make_tuple(" + varname + ".first, " + varname + ".second);"
-		return text
-		tmp_name = "tmp_" + str(Translator.tmp_cntr)
-		Translator.tmp_cntr = Translator.tmp_cntr + 1
-		if ref:
-			text += prefix + "for(auto " + tmp_name + " : *" + varname + ")"
+		# if the tuple is a pair of SigSpecs (aka SigSig), then we need
+		# to call get_py_obj() on each item in the tuple
+		if types[0].name in classnames:
+			first_var = types[0].name + "::get_py_obj(" + varname + ".first)"
 		else:
-			text += prefix + "for(auto " + tmp_name + " : " + varname + ")"
-		text += prefix + "{"
-		if types[0].name.split(" ")[-1] in primitive_types or types[0].name in enum_names:
-			text += prefix + "\t" + varname + "___tmp.append(" + tmp_name + ");"
-		elif types[0].name in known_containers:
-			text += known_containers[types[0].name].translate_cpp(tmp_name, types[0].cont.args, prefix + "\t", types[1].attr_type == attr_types.star)
-			text += prefix + "\t" + varname + "___tmp.append(" + types[0].name + "::get_py_obj(" + tmp_name + "___tmp);"
-		elif types[0].name in classnames:
-			text += prefix + "\t" + varname + "___tmp.append(" + types[0].name + "::get_py_obj(" + tmp_name + "));"
-		text += prefix + "}"
+			first_var = varname + ".first"
+		if types[1].name in classnames:
+			second_var = types[1].name + "::get_py_obj(" + varname + ".second)"
+		else:
+			second_var = varname + ".second"
+		text  = prefix + TupleTranslator.typename + " " + varname + "___tmp = boost::python::make_tuple(" + first_var + ", " + second_var + ");"
 		return text
 
 #Associate the Translators with their c++ type
diff --git a/passes/pmgen/README.md b/passes/pmgen/README.md
index 5f6a8ab1b..0856c9ba3 100644
--- a/passes/pmgen/README.md
+++ b/passes/pmgen/README.md
@@ -178,6 +178,45 @@ evaluates to `false`.
 The `semioptional` statement marks matches that must match if at least one
 matching cell exists, but if no matching cell exists it is set to `nullptr`.
 
+Slices and choices
+------------------
+
+Cell matches can contain "slices" and "choices". Slices can be used to
+create matches for different sections of a cell. For example:
+
+    state <int> pmux_slice
+
+    match pmux
+        select pmux->type == $pmux
+        slice idx GetSize(port(pmux, \S))
+        index <SigBit> port(pmux, \S)[idx] === port(eq, \Y)
+	set pmux_slice idx
+    endmatch
+
+The first argument to `slice` is the local variable name used to identify the
+slice. The second argument is the number of slices that should be created for
+this cell. The `set` statement can be used to copy that index into a state
+variable so that later matches and/or code blocks can refer to it.
+
+A similar mechanism is "choices", where a list of options is given as
+second argument, and the matcher will iterate over those options:
+
+    state <SigSpec> foo bar
+    state <IdString> eq_ab eq_ba
+
+    match eq
+        select eq->type == $eq
+        choice <IdString> AB {\A, \B}
+        define <IdString> BA (AB == \A ? \B : \A)
+        index <SigSpec> port(eq, AB) === foo
+        index <SigSpec> port(eq, BA) === bar
+        set eq_ab AB
+        set eq_ba BA
+    generate
+
+Notice how `define` can be used to define additional local variables similar
+to the loop variables defined by `slice` and `choice`.
+
 Additional code
 ---------------
 
@@ -326,7 +365,7 @@ test-case generation. For example:
 
     match mul
         ...
-    generate 10
+    generate 10 0
         SigSpec Y = port(ff, \D);
         SigSpec A = module->addWire(NEW_ID, GetSize(Y) - rng(GetSize(Y)/2));
         SigSpec B = module->addWire(NEW_ID, GetSize(Y) - rng(GetSize(Y)/2));
@@ -335,8 +374,11 @@ test-case generation. For example:
 
 The expression `rng(n)` returns a non-negative integer less than `n`.
 
-The argument to `generate` is the chance of this generate block being executed
-when the match block did not match anything, in percent.
+The first argument to `generate` is the chance of this generate block being
+executed when the match block did not match anything, in percent.
+
+The second argument to `generate` is the chance of this generate block being
+executed when the match block did match something, in percent.
 
 The special statement `finish` can be used within generate blocks to terminate
 the current pattern matcher run.
diff --git a/passes/pmgen/pmgen.py b/passes/pmgen/pmgen.py
index 18c3bf5a5..573722d68 100644
--- a/passes/pmgen/pmgen.py
+++ b/passes/pmgen/pmgen.py
@@ -207,9 +207,10 @@ def process_pmgfile(f, filename):
             state_types[current_pattern][line[1]] = "Cell*";
 
             block["if"] = list()
-            block["select"] = list()
+            block["setup"] = list()
             block["index"] = list()
             block["filter"] = list()
+            block["sets"] = list()
             block["optional"] = False
             block["semioptional"] = False
 
@@ -228,7 +229,22 @@ def process_pmgfile(f, filename):
 
                 if a[0] == "select":
                     b = l.lstrip()[6:]
-                    block["select"].append(rewrite_cpp(b.strip()))
+                    block["setup"].append(("select", rewrite_cpp(b.strip())))
+                    continue
+
+                if a[0] == "slice":
+                    m = re.match(r"^\s*slice\s+(\S+)\s+(.*?)\s*$", l)
+                    block["setup"].append(("slice", m.group(1), rewrite_cpp(m.group(2))))
+                    continue
+
+                if a[0] == "choice":
+                    m = re.match(r"^\s*choice\s+<(.*?)>\s+(\S+)\s+(.*?)\s*$", l)
+                    block["setup"].append(("choice", m.group(1), m.group(2), rewrite_cpp(m.group(3))))
+                    continue
+
+                if a[0] == "define":
+                    m = re.match(r"^\s*define\s+<(.*?)>\s+(\S+)\s+(.*?)\s*$", l)
+                    block["setup"].append(("define", m.group(1), m.group(2), rewrite_cpp(m.group(3))))
                     continue
 
                 if a[0] == "index":
@@ -242,6 +258,11 @@ def process_pmgfile(f, filename):
                     block["filter"].append(rewrite_cpp(b.strip()))
                     continue
 
+                if a[0] == "set":
+                    m = re.match(r"^\s*set\s+(\S+)\s+(.*?)\s*$", l)
+                    block["sets"].append((m.group(1), rewrite_cpp(m.group(2))))
+                    continue
+
                 if a[0] == "optional":
                     block["optional"] = True
                     continue
@@ -252,14 +273,16 @@ def process_pmgfile(f, filename):
 
                 if a[0] == "generate":
                     block["genargs"] = list([int(s) for s in a[1:]])
+                    if len(block["genargs"]) == 0: block["genargs"].append(100)
+                    if len(block["genargs"]) == 1: block["genargs"].append(0)
+                    assert len(block["genargs"]) == 2
                     block["gencode"] = list()
-                    assert len(block["genargs"]) < 2
                     while True:
                         linenr += 1
                         l = f.readline()
                         assert l != ""
                         a = l.split()
-                        if a[0] == "endmatch": break
+                        if len(a) == 1 and a[0] == "endmatch": break
                         block["gencode"].append(rewrite_cpp(l.rstrip()))
                     break
 
@@ -357,8 +380,17 @@ with open(outfile, "w") as f:
             index_types = list()
             for entry in block["index"]:
                 index_types.append(entry[0])
+            value_types = ["Cell*"]
+            for entry in block["setup"]:
+                if entry[0] == "slice":
+                    value_types.append("int")
+                if entry[0] == "choice":
+                    value_types.append(entry[1])
+                if entry[0] == "define":
+                    value_types.append(entry[1])
             print("  typedef std::tuple<{}> index_{}_key_type;".format(", ".join(index_types), index), file=f)
-            print("  dict<index_{}_key_type, vector<Cell*>> index_{};".format(index, index), file=f)
+            print("  typedef std::tuple<{}> index_{}_value_type;".format(", ".join(value_types), index), file=f)
+            print("  dict<index_{}_key_type, vector<index_{}_value_type>> index_{};".format(index, index, index), file=f)
     print("  dict<SigBit, pool<Cell*>> sigusers;", file=f)
     print("  pool<Cell*> blacklist_cells;", file=f)
     print("  pool<Cell*> autoremove_cells;", file=f)
@@ -390,8 +422,6 @@ with open(outfile, "w") as f:
     print("  void add_siguser(const SigSpec &sig, Cell *cell) {", file=f)
     print("    for (auto bit : sigmap(sig)) {", file=f)
     print("      if (bit.wire == nullptr) continue;", file=f)
-    print("      if (sigusers.count(bit) == 0 && bit.wire->port_id)", file=f)
-    print("        sigusers[bit].insert(nullptr);", file=f)
     print("      sigusers[bit].insert(cell);", file=f)
     print("    }", file=f)
     print("  }", file=f)
@@ -446,10 +476,11 @@ with open(outfile, "w") as f:
             else:
                 print("    ud_{}.{} = {}();".format(current_pattern, s, t), file=f)
     current_pattern = None
-    print("    for (auto cell : module->cells()) {", file=f)
+    print("    for (auto port : module->ports)", file=f)
+    print("      add_siguser(module->wire(port), nullptr);", file=f)
+    print("    for (auto cell : module->cells())", file=f)
     print("      for (auto &conn : cell->connections())", file=f)
     print("        add_siguser(conn.second, cell);", file=f)
-    print("    }", file=f)
     print("    for (auto cell : cells) {", file=f)
 
     for index in range(len(blocks)):
@@ -457,12 +488,34 @@ with open(outfile, "w") as f:
         if block["type"] == "match":
             print("      do {", file=f)
             print("        Cell *{} = cell;".format(block["cell"]), file=f)
-            for expr in block["select"]:
-                print("        if (!({})) break;".format(expr), file=f)
+            print("        index_{}_value_type value;".format(index), file=f)
+            print("        std::get<0>(value) = cell;", file=f)
+            loopcnt = 0
+            valueidx = 1
+            for item in block["setup"]:
+                if item[0] == "select":
+                    print("        if (!({})) continue;".format(item[1]), file=f)
+                if item[0] == "slice":
+                    print("        int &{} = std::get<{}>(value);".format(item[1], valueidx), file=f)
+                    print("        for ({} = 0; {} < {}; {}++) {{".format(item[1], item[1], item[2], item[1]), file=f)
+                    valueidx += 1
+                    loopcnt += 1
+                if item[0] == "choice":
+                    print("        vector<{}> _pmg_choices_{} = {};".format(item[1], item[2], item[3]), file=f)
+                    print("        for (const {} &{} : _pmg_choices_{}) {{".format(item[1], item[2], item[2]), file=f)
+                    print("        std::get<{}>(value) = {};".format(valueidx, item[2]), file=f)
+                    valueidx += 1
+                    loopcnt += 1
+                if item[0] == "define":
+                    print("        {} &{} = std::get<{}>(value);".format(item[1], item[2], valueidx), file=f)
+                    print("        {} = {};".format(item[2], item[3]), file=f)
+                    valueidx += 1
             print("        index_{}_key_type key;".format(index), file=f)
             for field, entry in enumerate(block["index"]):
                 print("        std::get<{}>(key) = {};".format(field, entry[1]), file=f)
-            print("        index_{}[key].push_back(cell);".format(index), file=f)
+            print("        index_{}[key].push_back(value);".format(index), file=f)
+            for i in range(loopcnt):
+                print("        }", file=f)
             print("      } while (0);", file=f)
 
     print("    }", file=f)
@@ -535,6 +588,8 @@ with open(outfile, "w") as f:
                     const_st.add(s)
             elif blocks[i]["type"] == "match":
                 const_st.add(blocks[i]["cell"])
+                for item in blocks[i]["sets"]:
+                    const_st.add(item[0])
             else:
                 assert False
 
@@ -548,6 +603,10 @@ with open(outfile, "w") as f:
             s = block["cell"]
             assert s not in const_st
             nonconst_st.add(s)
+            for item in block["sets"]:
+                if item[0] in const_st:
+                    const_st.remove(item[0])
+                nonconst_st.add(item[0])
         else:
             assert False
 
@@ -570,7 +629,7 @@ with open(outfile, "w") as f:
             print("", file=f)
             for s in sorted(restore_st):
                 t = state_types[current_pattern][s]
-                print("    {} backup_{} = {};".format(t, s, s), file=f)
+                print("    {} _pmg_backup_{} = {};".format(t, s, s), file=f)
 
         if block["type"] == "code":
             print("", file=f)
@@ -610,7 +669,7 @@ with open(outfile, "w") as f:
                 print("", file=f)
                 for s in sorted(restore_st):
                     t = state_types[current_pattern][s]
-                    print("    {} = backup_{};".format(s, s), file=f)
+                    print("    {} = _pmg_backup_{};".format(s, s), file=f)
                 for s in sorted(nonconst_st):
                     if s not in restore_st:
                         t = state_types[current_pattern][s]
@@ -622,7 +681,7 @@ with open(outfile, "w") as f:
         elif block["type"] == "match":
             assert len(restore_st) == 0
 
-            print("    Cell* backup_{} = {};".format(block["cell"], block["cell"]), file=f)
+            print("    Cell* _pmg_backup_{} = {};".format(block["cell"], block["cell"]), file=f)
 
             if len(block["if"]):
                 for expr in block["if"]:
@@ -630,7 +689,7 @@ with open(outfile, "w") as f:
                     print("    if (!({})) {{".format(expr), file=f)
                     print("      {} = nullptr;".format(block["cell"]), file=f)
                     print("      block_{}(recursion+1);".format(index+1), file=f)
-                    print("      {} = backup_{};".format(block["cell"], block["cell"]), file=f)
+                    print("      {} = _pmg_backup_{};".format(block["cell"], block["cell"]), file=f)
                     print("      return;", file=f)
                     print("    }", file=f)
 
@@ -645,21 +704,37 @@ with open(outfile, "w") as f:
 
             print("", file=f)
             print("    if (cells_ptr != index_{}.end()) {{".format(index), file=f)
-            print("      const vector<Cell*> &cells = cells_ptr->second;".format(index), file=f)
-            print("      for (int idx = 0; idx < GetSize(cells); idx++) {", file=f)
-            print("        {} = cells[idx];".format(block["cell"]), file=f)
+            print("      const vector<index_{}_value_type> &cells = cells_ptr->second;".format(index), file=f)
+            print("      for (int _pmg_idx = 0; _pmg_idx < GetSize(cells); _pmg_idx++) {", file=f)
+            print("        {} = std::get<0>(cells[_pmg_idx]);".format(block["cell"]), file=f)
+            valueidx = 1
+            for item in block["setup"]:
+                if item[0] == "slice":
+                    print("        const int &{} YS_ATTRIBUTE(unused) = std::get<{}>(cells[_pmg_idx]);".format(item[1], valueidx), file=f)
+                    valueidx += 1
+                if item[0] == "choice":
+                    print("        const {} &{} YS_ATTRIBUTE(unused) = std::get<{}>(cells[_pmg_idx]);".format(item[1], item[2], valueidx), file=f)
+                    valueidx += 1
+                if item[0] == "define":
+                    print("        const {} &{} YS_ATTRIBUTE(unused) = std::get<{}>(cells[_pmg_idx]);".format(item[1], item[2], valueidx), file=f)
+                    valueidx += 1
             print("        if (blacklist_cells.count({})) continue;".format(block["cell"]), file=f)
             for expr in block["filter"]:
                 print("        if (!({})) continue;".format(expr), file=f)
             if block["semioptional"] or block["genargs"] is not None:
                 print("        found_any_match = true;", file=f)
-            print("        auto rollback_ptr = rollback_cache.insert(make_pair(cells[idx], recursion));", file=f)
+            for item in block["sets"]:
+                print("        auto _pmg_backup_{} = {};".format(item[0], item[0]), file=f)
+                print("        {} = {};".format(item[0], item[1]), file=f)
+            print("        auto rollback_ptr = rollback_cache.insert(make_pair(std::get<0>(cells[_pmg_idx]), recursion));", file=f)
             print("        block_{}(recursion+1);".format(index+1), file=f)
+            for item in block["sets"]:
+                print("        {} = _pmg_backup_{};".format(item[0], item[0]), file=f)
             print("        if (rollback_ptr.second)", file=f)
             print("          rollback_cache.erase(rollback_ptr.first);", file=f)
             print("        if (rollback) {", file=f)
             print("          if (rollback != recursion) {{".format(index+1), file=f)
-            print("            {} = backup_{};".format(block["cell"], block["cell"]), file=f)
+            print("            {} = _pmg_backup_{};".format(block["cell"], block["cell"]), file=f)
             print("            return;", file=f)
             print("          }", file=f)
             print("          rollback = 0;", file=f)
@@ -676,13 +751,11 @@ with open(outfile, "w") as f:
             if block["semioptional"]:
                 print("    if (!found_any_match) block_{}(recursion+1);".format(index+1), file=f)
 
-            print("    {} = backup_{};".format(block["cell"], block["cell"]), file=f)
+            print("    {} = _pmg_backup_{};".format(block["cell"], block["cell"]), file=f)
 
             if block["genargs"] is not None:
                 print("#define finish do { rollback = -1; return; } while(0)", file=f)
-                print("    if (generate_mode && !found_any_match) {", file=f)
-                if len(block["genargs"]) == 1:
-                    print("    if (rng(100) >= {}) return;".format(block["genargs"][0]), file=f)
+                print("    if (generate_mode && rng(100) < (found_any_match ? {} : {})) {{".format(block["genargs"][1], block["genargs"][0]), file=f)
                 for line in block["gencode"]:
                     print("      " + line, file=f)
                 print("    }", file=f)
diff --git a/passes/pmgen/test_pmgen.cc b/passes/pmgen/test_pmgen.cc
index 9f42a95d0..0ad769dfd 100644
--- a/passes/pmgen/test_pmgen.cc
+++ b/passes/pmgen/test_pmgen.cc
@@ -99,6 +99,24 @@ void reduce_tree(test_pmgen_pm &pm)
 	log("    -> %s (%s)\n", log_id(c), log_id(c->type));
 }
 
+void opt_eqpmux(test_pmgen_pm &pm)
+{
+	auto &st = pm.st_eqpmux;
+
+	SigSpec Y = st.pmux->getPort(ID::Y);
+	int width = GetSize(Y);
+
+	SigSpec EQ = st.pmux->getPort(ID::B).extract(st.pmux_slice_eq*width, width);
+	SigSpec NE = st.pmux->getPort(ID::B).extract(st.pmux_slice_ne*width, width);
+
+	log("Found eqpmux circuit driving %s (eq=%s, ne=%s, pmux=%s).\n",
+			log_signal(Y), log_id(st.eq), log_id(st.ne), log_id(st.pmux));
+
+	pm.autoremove(st.pmux);
+	Cell *c = pm.module->addMux(NEW_ID, NE, EQ, st.eq->getPort(ID::Y), Y);
+	log("    -> %s (%s)\n", log_id(c), log_id(c->type));
+}
+
 #define GENERATE_PATTERN(pmclass, pattern) \
 	generate_pattern<pmclass>([](pmclass &pm, std::function<void()> f){ return pm.run_ ## pattern(f); }, #pmclass, #pattern, design)
 
@@ -149,16 +167,17 @@ void generate_pattern(std::function<void(pm&,std::function<void()>)> run, const
 	log("Generating \"%s\" patterns for pattern matcher \"%s\".\n", pattern, pmclass);
 
 	int modcnt = 0;
+	int maxmodcnt = 100;
 	int maxsubcnt = 4;
 	int timeout = 0;
 	vector<Module*> mods;
 
-	while (modcnt < 100)
+	while (modcnt < maxmodcnt)
 	{
 		int submodcnt = 0, itercnt = 0, cellcnt = 0;
 		Module *mod = design->addModule(NEW_ID);
 
-		while (modcnt < 100 && submodcnt < maxsubcnt && itercnt++ < 1000)
+		while (modcnt < maxmodcnt && submodcnt < maxsubcnt && itercnt++ < 1000)
 		{
 			if (timeout++ > 10000)
 				log_error("pmgen generator is stuck: 10000 iterations an no matching module generated.\n");
@@ -233,6 +252,12 @@ struct TestPmgenPass : public Pass {
 		log("\n");
 
 		log("\n");
+		log("    test_pmgen -eqpmux [options] [selection]\n");
+		log("\n");
+		log("Demo for recursive pmgen patterns. Optimize EQ/NE/PMUX circuits.\n");
+		log("\n");
+
+		log("\n");
 		log("    test_pmgen -generate [options] <pattern_name>\n");
 		log("\n");
 		log("Create modules that match the specified pattern.\n");
@@ -277,6 +302,25 @@ struct TestPmgenPass : public Pass {
 			test_pmgen_pm(module, module->selected_cells()).run_reduce(reduce_tree);
 	}
 
+	void execute_eqpmux(std::vector<std::string> args, RTLIL::Design *design)
+	{
+		log_header(design, "Executing TEST_PMGEN pass (-eqpmux).\n");
+
+		size_t argidx;
+		for (argidx = 2; argidx < args.size(); argidx++)
+		{
+			// if (args[argidx] == "-singleton") {
+			// 	singleton_mode = true;
+			// 	continue;
+			// }
+			break;
+		}
+		extra_args(args, argidx, design);
+
+		for (auto module : design->selected_modules())
+			test_pmgen_pm(module, module->selected_cells()).run_eqpmux(opt_eqpmux);
+	}
+
 	void execute_generate(std::vector<std::string> args, RTLIL::Design *design)
 	{
 		log_header(design, "Executing TEST_PMGEN pass (-generate).\n");
@@ -299,6 +343,9 @@ struct TestPmgenPass : public Pass {
 		if (pattern == "reduce")
 			return GENERATE_PATTERN(test_pmgen_pm, reduce);
 
+		if (pattern == "eqpmux")
+			return GENERATE_PATTERN(test_pmgen_pm, eqpmux);
+
 		if (pattern == "ice40_dsp")
 			return GENERATE_PATTERN(ice40_dsp_pm, ice40_dsp);
 
@@ -319,6 +366,8 @@ struct TestPmgenPass : public Pass {
 				return execute_reduce_chain(args, design);
 			if (args[1] == "-reduce_tree")
 				return execute_reduce_tree(args, design);
+			if (args[1] == "-eqpmux")
+				return execute_eqpmux(args, design);
 			if (args[1] == "-generate")
 				return execute_generate(args, design);
 		}
diff --git a/passes/pmgen/test_pmgen.pmg b/passes/pmgen/test_pmgen.pmg
index 211477a62..287ed97d8 100644
--- a/passes/pmgen/test_pmgen.pmg
+++ b/passes/pmgen/test_pmgen.pmg
@@ -60,8 +60,8 @@ code portname
 endcode
 
 match next
-	select nusers(port(next, \Y)) == 2
 	select next->type.in($_AND_, $_OR_, $_XOR_)
+	select nusers(port(next, \Y)) == 2
 	index <IdString> next->type === first->type
 	index <SigSpec> port(next, \Y) === port(first, portname)
 endmatch
@@ -77,8 +77,8 @@ arg first
 
 match next
 	semioptional
-	select nusers(port(next, \Y)) == 2
 	select next->type.in($_AND_, $_OR_, $_XOR_)
+	select nusers(port(next, \Y)) == 2
 	index <IdString> next->type === chain.back().first->type
 	index <SigSpec> port(next, \Y) === port(chain.back().first, chain.back().second)
 generate 10
@@ -104,3 +104,86 @@ finally
 	if (next)
 		chain.pop_back();
 endcode
+
+// ==================================================================
+
+pattern eqpmux
+
+state <bool> eq_ne_signed
+state <SigSpec> eq_inA eq_inB
+state <int> pmux_slice_eq pmux_slice_ne
+
+match eq
+	select eq->type == $eq
+	choice <IdString> AB {\A, \B}
+	define <IdString> BA AB == \A ? \B : \A
+	set eq_inA port(eq, \A)
+	set eq_inB port(eq, \B)
+	set eq_ne_signed param(eq, \A_SIGNED).as_bool()
+generate 100 10
+	SigSpec A = module->addWire(NEW_ID, rng(7)+1);
+	SigSpec B = module->addWire(NEW_ID, rng(7)+1);
+	SigSpec Y = module->addWire(NEW_ID);
+	module->addEq(NEW_ID, A, B, Y, rng(2));
+endmatch
+
+match pmux
+	select pmux->type == $pmux
+	slice idx GetSize(port(pmux, \S))
+	index <SigBit> port(pmux, \S)[idx] === port(eq, \Y)
+	set pmux_slice_eq idx
+generate 100 10
+	int width = rng(7) + 1;
+	int numsel = rng(4) + 1;
+	int idx = rng(numsel);
+
+	SigSpec A = module->addWire(NEW_ID, width);
+	SigSpec Y = module->addWire(NEW_ID, width);
+
+	SigSpec B, S;
+	for (int i = 0; i < numsel; i++) {
+		B.append(module->addWire(NEW_ID, width));
+		S.append(i == idx ? port(eq, \Y) : module->addWire(NEW_ID));
+	}
+
+	module->addPmux(NEW_ID, A, B, S, Y);
+endmatch
+
+match ne
+	select ne->type == $ne
+	choice <IdString> AB {\A, \B}
+	define <IdString> BA (AB == \A ? \B : \A)
+	index <SigSpec> port(ne, AB) === eq_inA
+	index <SigSpec> port(ne, BA) === eq_inB
+	index <int> param(ne, \A_SIGNED).as_bool() === eq_ne_signed
+generate 100 10
+	SigSpec A = eq_inA, B = eq_inB, Y;
+	if (rng(2)) {
+		std::swap(A, B);
+	}
+	if (rng(2)) {
+		for (auto bit : port(pmux, \S)) {
+			if (nusers(bit) < 2)
+				Y.append(bit);
+		}
+		if (GetSize(Y))
+			Y = Y[rng(GetSize(Y))];
+		else
+			Y = module->addWire(NEW_ID);
+	} else {
+		Y = module->addWire(NEW_ID);
+	}
+	module->addNe(NEW_ID, A, B, Y, rng(2));
+endmatch
+
+match pmux2
+	select pmux2->type == $pmux
+	slice idx GetSize(port(pmux2, \S))
+	index <Cell*> pmux2 === pmux
+	index <SigBit> port(pmux2, \S)[idx] === port(ne, \Y)
+	set pmux_slice_ne idx
+endmatch
+
+code
+	accept;
+endcode
diff --git a/techlibs/xilinx/synth_xilinx.cc b/techlibs/xilinx/synth_xilinx.cc
index 00849ad00..f058da83d 100644
--- a/techlibs/xilinx/synth_xilinx.cc
+++ b/techlibs/xilinx/synth_xilinx.cc
@@ -117,6 +117,7 @@ struct SynthXilinxPass : public ScriptPass
 
 	std::string top_opt, edif_file, blif_file, family;
 	bool flatten, retime, vpr, ise, iopad, noiopad, noclkbuf, nobram, nolutram, nosrl, nocarry, nowidelut, abc9;
+	bool flatten_before_abc;
 	int widemux;
 
 	void clear_flags() YS_OVERRIDE
@@ -139,6 +140,7 @@ struct SynthXilinxPass : public ScriptPass
 		nocarry = false;
 		nowidelut = false;
 		abc9 = false;
+		flatten_before_abc = false;
 		widemux = 0;
 	}
 
@@ -178,6 +180,10 @@ struct SynthXilinxPass : public ScriptPass
 				flatten = true;
 				continue;
 			}
+			if (args[argidx] == "-flatten_before_abc") {
+				flatten_before_abc = true;
+				continue;
+			}
 			if (args[argidx] == "-retime") {
 				retime = true;
 				continue;
@@ -417,6 +423,8 @@ struct SynthXilinxPass : public ScriptPass
 
 		if (check_label("map_luts")) {
 			run("opt_expr -mux_undef");
+			if (flatten_before_abc)
+				run("flatten");
 			if (help_mode)
 				run("abc -luts 2:2,3,6:5[,10,20] [-dff]", "(option for 'nowidelut', option for '-retime')");
 			else if (abc9) {