Larger instruction window for zinc

1 files changed, 351 insertions, 210 deletions

diff --git a/etc/compile-by-zinc/compile-to-zinc.py b/etc/compile-by-zinc/compile-to-zinc.py
index 632035009..0b91ae073 100755
--- a/etc/compile-by-zinc/compile-to-zinc.py
+++ b/etc/compile-by-zinc/compile-to-zinc.py
@@ -8,7 +8,7 @@ CORES = ('ADD_MUL0', 'ADD_MUL1', 'MUL0', 'LEA_BW0', 'LEA_BW1', 'NOOP_CORE')
 
 OP_NAMES = {'*':'MUL', '+':'ADD', '>>':'SHL', '<<':'SHR', '|':'OR', '&':'AND'}
 
-MAX_INSTRUCTION_WINDOW = 40
+MAX_INSTRUCTION_WINDOW = 53
 
 INSTRUCTIONS_PER_CYCLE = 4
 
@@ -54,9 +54,14 @@ def parse_lines(lines):
     for line in lines[1:-1]:
         datatype, varname, arg1, op, arg2 = re.findall('^(u?int[0-9]*_t) ([^ ]*) = ([^ ]*) ([^ ]*) ([^ ]*);$', line)[0]
         ret['lines'].append({'type':datatype, 'out':varname, 'op':op, 'args':(arg1, arg2), 'source':line})
-    print('Compiling %d of %d lines...' % (min(MAX_INSTRUCTION_WINDOW, len(ret['lines'])), len(ret['lines'])))
-    ret['lines'] = ret['lines'][:MAX_INSTRUCTION_WINDOW]
-    return ret
+    print('Compiling %d lines in groups of %d...' % (len(ret['lines']), min(MAX_INSTRUCTION_WINDOW, len(ret['lines']))))
+    ret['lines'] = tuple(ret['lines'])
+    split_ret = []
+    for start in range(0, len(ret['lines']), MAX_INSTRUCTION_WINDOW):
+        cur_ret = dict(ret)
+        cur_ret['lines'] = ret['lines'][start:][:MAX_INSTRUCTION_WINDOW]
+        split_ret.append(cur_ret)
+    return tuple(split_ret)
 
 def get_var_names(input_data):
     return tuple(line['out'] for line in input_data['lines'])
@@ -75,189 +80,189 @@ def create_set(name, items):
     ret += '\n'
     return ret
 
-######################################################################
-######################################################################
-##                Assign locations to instructions                  ##
-######################################################################
-######################################################################
-def make_assign_locations_to_instructions(data):
-    def make_decls(input_data):
-        var_names = get_var_names(input_data)
-        ret = ''
-        ret += create_set('CORE', CORES)
-        for var in var_names:
-            ret += 'var int: %s_loc;\n' % var
-            ret += 'var int: %s_latency;\n' % var
-            ret += 'var CORE: %s_core;\n' % var
-        ret += 'var int: RET_loc;\n'
-        return ret
-
-    def make_disjoint(input_data):
-        var_names = get_var_names(input_data)
-        ret = ''
-        for var in var_names:
-            ret += 'constraint %s_loc >= 0;\n' % var
-            ret += 'constraint %s_latency >= 0;\n' % var
-            ret += 'constraint %s_loc + %s_latency <= RET_loc;\n' % (var, var)
-        # TODO: Figure out if this next constraint actually helps things....
-        MAX_NUMBER_OF_NOPS_PER_INSTRUCTION = 3
-        APPROXIMATE_MAX_LATENCY = 6 * INSTRUCTIONS_PER_CYCLE
-        max_ret_loc = ('constraint RET_loc <= %d;\n'
-                       % (len(var_names)
-                          * MAX_NUMBER_OF_NOPS_PER_INSTRUCTION
-                          + APPROXIMATE_MAX_LATENCY))
-        #ret += max_ret_loc
-        ret += '\n'
-        for var_i in range(len(var_names)):
-            for var_j in range(var_i+1, len(var_names)):
-                ret += 'constraint %s_loc != %s_loc;\n' % (var_names[var_i], var_names[var_j])
-        ret += '\n'
-        return ret
-
-    def make_dependencies(input_data):
-        var_names = get_var_names(input_data)
-        ret = ''
-        for line in input_data['lines']:
-            for arg in line['args']:
-                if arg in var_names and arg[0] not in '0123456789':
-                    ret += ('constraint %s_loc + %s_latency <= %s_loc;\n'
-                            % (arg, arg, line['out']))
-        ret += '\n'
-        return ret
-
-    def make_cores(input_data):
-        ret = ''
-        for line in input_data['lines']:
-            ret += 'constraint '
-            possible_cores = []
-            cores = MODEL[line['op']]
-            for core in cores:
-                possible_cores.append('(%s_latency == %d /\ %s_core == %s)'
-                                      % (line['out'], core['latency'] * INSTRUCTIONS_PER_CYCLE,
-                                         line['out'], core['core']['name']))
-            ret += ' \/ '.join(possible_cores)
-            ret += ';\n'
-        return ret
-
-    def make_cores_disjoint(input_data):
-        var_names = get_var_names(input_data)
-        ret = ''
-        for core in CORES:
-            for var_i in range(len(var_names)):
-                for var_j in range(var_i+1, len(var_names)):
-                    op_i = input_data['lines'][var_i]['op']
-                    op_j = input_data['lines'][var_j]['op']
-                    latencies_i = [val['core']['latency'] for val in MODEL[op_i] if val['core']['name'] == core]
-                    latencies_j = [val['core']['latency'] for val in MODEL[op_j] if val['core']['name'] == core]
-                    if len(latencies_i) == 0 or len(latencies_j) == 0: continue
-                    assert len(latencies_i) == 1
-                    assert len(latencies_j) == 1
-                    ret += ('constraint (%(vari)s_core != %(core)s \\/ %(varj)s_core != %(core)s) \\/ (%(vari)s_loc + %(latencyi)d <= %(varj)s_loc \\/ %(varj)s_loc + %(latencyj)d <= %(vari)s_loc);\n'
-                            % { 'vari':var_names[var_i] , 'varj':var_names[var_j] , 'core':core ,
-                                'latencyi':latencies_i[0] * INSTRUCTIONS_PER_CYCLE,
-                                'latencyj':latencies_j[0] * INSTRUCTIONS_PER_CYCLE})
-        ret += '\n'
-        return ret
-
-    def make_output(input_data):
-        ret = 'solve minimize RET_loc;\n\n'
-        ret += 'output [ "{\\n" ++\n'
-        for line in input_data['lines']:
-            ret += ('         "  \'%(var)s_loc\': " ++ show(%(var)s_loc) ++ ", \'%(var)s_latency\': " ++ show(%(var)s_latency) ++ ", \'%(var)s_core\': " ++ show(%(var)s_core) ++ ",\\n" ++\n         %% %(source)s\n'
-                    % { 'var':line['out'] , 'source':line['source'] })
-        ret += '         "  \'RET_loc\': " ++ show(RET_loc) ++ "\\n" ++\n'
-        ret += '         "}" ]\n'
-        return ret
-
-    return '\n'.join([
-        make_decls(data),
-        make_disjoint(data),
-        make_dependencies(data),
-        make_cores(data),
-        make_cores_disjoint(data),
-        make_output(data)
-        ])
-
-
-######################################################################
-######################################################################
-##                Assign instructions to locations                  ##
-######################################################################
-######################################################################
-
-def make_assign_instructions_to_locations(data):
-    def make_decls(input_data):
-        var_names = get_var_names(input_data)
-        ret = ''
-        ret += 'include "alldifferent.mzn";\n'
-        ret += create_set('CORE', CORES)
-        ret += create_set('INSTRUCTION', ['NOOP'] + list(var_names))
-        MAX_NUMBER_OF_NOOPS_PER_INSTRUCTION = 3
-        APPROXIMATE_MAX_LATENCY = 6 * INSTRUCTIONS_PER_CYCLE
-        max_loc = len(var_names) * MAX_NUMBER_OF_NOOPS_PER_INSTRUCTION # + APPROXIMATE_MAX_LATENCY
-        ret += 'int: MAX_LOC = %d;\n\n' % max_loc
-        ret += 'array[1..MAX_LOC] of var INSTRUCTION: output_instructions;\n'
-        ret += 'array[1..MAX_LOC] of var CORE: output_cores;\n'
-        ret += 'array[1..MAX_LOC] of var int: output_core_latency;\n'
-        ret += 'array[1..MAX_LOC] of var int: output_data_latency;\n'
-        ret += 'var int: RET_loc;\n'
-        ret += '\n'
-        ret += 'constraint 1 <= RET_loc /\\ RET_loc <= MAX_LOC;\n'
-        ret += '\n'
-        return ret
-
-    def make_disjoint(input_data):
-        var_names = get_var_names(input_data)
-        ret = ''
-        #ret += 'constraint alldifferent( [ output_instructions[i] | i in 1..MAX_LOC where output_instructions[i] != NOOP ] );\n'
-        ret += 'constraint forall (i,j in 1..MAX_LOC where i < j) (output_instructions[i] == NOOP \\/ output_instructions[i] != output_instructions[j]);\n'
-        ret += 'constraint sum( [ 1 | i in 1..MAX_LOC where output_instructions[i] != NOOP ] ) == length(INSTRUCTION) - 1;\n'
-        ret += 'constraint forall (i in 1..MAX_LOC where RET_loc <= i) (output_instructions[i] == NOOP);\n'
-        ret += 'constraint forall (i,j in 1..MAX_LOC where i < j) ((output_instructions[i] != NOOP /\\ output_instructions[j] != NOOP /\\ output_cores[i] == output_cores[j]) -> i + output_core_latency[i] <= j);\n'
-        ret += 'constraint forall (i in 1..MAX_LOC) (output_instructions[i] == NOOP \\/ i + output_data_latency[i] <= RET_loc);\n'
-        ret += 'constraint forall (i in 1..MAX_LOC) (output_instructions[i] == NOOP -> (output_core_latency[i] == 0 /\\ output_data_latency[i] == 0 /\\ output_cores[i] == NOOP_CORE));\n'
-        return ret
-
-    def make_dependencies(input_data):
-        var_names = get_var_names(input_data)
-        ret = ''
-        for line in input_data['lines']:
-            for arg in line['args']:
-                if arg in var_names and arg[0] not in '0123456789':
-                    ret += ('constraint forall (i,j in 1..MAX_LOC) ((output_instructions[i] == %s /\\ output_instructions[j] == %s) -> i + output_data_latency[i] <= j);\n'
-                            % (arg, line['out']))
-        ret += '\n'
-        return ret
-
-    def make_cores(input_data):
-        ret = ''
-        for line in input_data['lines']:
-            ret += 'constraint forall (i in 1..MAX_LOC) (output_instructions[i] == %s -> (' % line['out']
-            possible_cores = []
-            cores = MODEL[line['op']]
-            for core in cores:
-                possible_cores.append(r'(output_core_latency[i] == %d /\ output_data_latency[i] == %d /\ output_cores[i] == %s)'
-                                      % (core['core']['latency'] * INSTRUCTIONS_PER_CYCLE,
-                                         core['latency'] * INSTRUCTIONS_PER_CYCLE,
-                                         core['core']['name']))
-            ret += ' \/ '.join(possible_cores)
-            ret += '));\n'
-        ret += '\n'
-        return ret
-
-    def make_output(input_data):
-        ret = 'solve minimize RET_loc;\n\n'
-        ret += 'output [ "(" ++ show(INSTRUCTION_NAMES[fix(output_instructions[i])]) ++ ", " ++ show(CORE_NAMES[fix(output_cores[i])]) ++ ", " ++ show(output_core_latency[i]) ++ ", " ++ show(output_data_latency[i]) ++ ") ,\\n"\n'
-        ret += '       | i in 1..MAX_LOC ];\n'
-        return ret
-
-    return '\n'.join([
-        make_decls(data),
-        make_disjoint(data),
-        make_dependencies(data),
-        make_cores(data),
-        make_output(data)
-        ])
+########################################################################
+########################################################################
+####                Assign locations to instructions                  ##
+########################################################################
+########################################################################
+##def make_assign_locations_to_instructions(data):
+##    def make_decls(input_data):
+##        var_names = get_var_names(input_data)
+##        ret = ''
+##        ret += create_set('CORE', CORES)
+##        for var in var_names:
+##            ret += 'var int: %s_loc;\n' % var
+##            ret += 'var int: %s_latency;\n' % var
+##            ret += 'var CORE: %s_core;\n' % var
+##        ret += 'var int: RET_loc;\n'
+##        return ret
+##
+##    def make_disjoint(input_data):
+##        var_names = get_var_names(input_data)
+##        ret = ''
+##        for var in var_names:
+##            ret += 'constraint %s_loc >= 0;\n' % var
+##            ret += 'constraint %s_latency >= 0;\n' % var
+##            ret += 'constraint %s_loc + %s_latency <= RET_loc;\n' % (var, var)
+##        # TODO: Figure out if this next constraint actually helps things....
+##        MAX_NUMBER_OF_NOPS_PER_INSTRUCTION = 3
+##        APPROXIMATE_MAX_LATENCY = 6 * INSTRUCTIONS_PER_CYCLE
+##        max_ret_loc = ('constraint RET_loc <= %d;\n'
+##                       % (len(var_names)
+##                          * MAX_NUMBER_OF_NOPS_PER_INSTRUCTION
+##                          + APPROXIMATE_MAX_LATENCY))
+##        #ret += max_ret_loc
+##        ret += '\n'
+##        for var_i in range(len(var_names)):
+##            for var_j in range(var_i+1, len(var_names)):
+##                ret += 'constraint %s_loc != %s_loc;\n' % (var_names[var_i], var_names[var_j])
+##        ret += '\n'
+##        return ret
+##
+##    def make_dependencies(input_data):
+##        var_names = get_var_names(input_data)
+##        ret = ''
+##        for line in input_data['lines']:
+##            for arg in line['args']:
+##                if arg in var_names and arg[0] not in '0123456789':
+##                    ret += ('constraint %s_loc + %s_latency <= %s_loc;\n'
+##                            % (arg, arg, line['out']))
+##        ret += '\n'
+##        return ret
+##
+##    def make_cores(input_data):
+##        ret = ''
+##        for line in input_data['lines']:
+##            ret += 'constraint '
+##            possible_cores = []
+##            cores = MODEL[line['op']]
+##            for core in cores:
+##                possible_cores.append('(%s_latency == %d /\ %s_core == %s)'
+##                                      % (line['out'], core['latency'] * INSTRUCTIONS_PER_CYCLE,
+##                                         line['out'], core['core']['name']))
+##            ret += ' \/ '.join(possible_cores)
+##            ret += ';\n'
+##        return ret
+##
+##    def make_cores_disjoint(input_data):
+##        var_names = get_var_names(input_data)
+##        ret = ''
+##        for core in CORES:
+##            for var_i in range(len(var_names)):
+##                for var_j in range(var_i+1, len(var_names)):
+##                    op_i = input_data['lines'][var_i]['op']
+##                    op_j = input_data['lines'][var_j]['op']
+##                    latencies_i = [val['core']['latency'] for val in MODEL[op_i] if val['core']['name'] == core]
+##                    latencies_j = [val['core']['latency'] for val in MODEL[op_j] if val['core']['name'] == core]
+##                    if len(latencies_i) == 0 or len(latencies_j) == 0: continue
+##                    assert len(latencies_i) == 1
+##                    assert len(latencies_j) == 1
+##                    ret += ('constraint (%(vari)s_core != %(core)s \\/ %(varj)s_core != %(core)s) \\/ (%(vari)s_loc + %(latencyi)d <= %(varj)s_loc \\/ %(varj)s_loc + %(latencyj)d <= %(vari)s_loc);\n'
+##                            % { 'vari':var_names[var_i] , 'varj':var_names[var_j] , 'core':core ,
+##                                'latencyi':latencies_i[0] * INSTRUCTIONS_PER_CYCLE,
+##                                'latencyj':latencies_j[0] * INSTRUCTIONS_PER_CYCLE})
+##        ret += '\n'
+##        return ret
+##
+##    def make_output(input_data):
+##        ret = 'solve minimize RET_loc;\n\n'
+##        ret += 'output [ "{\\n" ++\n'
+##        for line in input_data['lines']:
+##            ret += ('         "  \'%(var)s_loc\': " ++ show(%(var)s_loc) ++ ", \'%(var)s_latency\': " ++ show(%(var)s_latency) ++ ", \'%(var)s_core\': " ++ show(%(var)s_core) ++ ",\\n" ++\n         %% %(source)s\n'
+##                    % { 'var':line['out'] , 'source':line['source'] })
+##        ret += '         "  \'RET_loc\': " ++ show(RET_loc) ++ "\\n" ++\n'
+##        ret += '         "}" ]\n'
+##        return ret
+##
+##    return '\n'.join([
+##        make_decls(data),
+##        make_disjoint(data),
+##        make_dependencies(data),
+##        make_cores(data),
+##        make_cores_disjoint(data),
+##        make_output(data)
+##        ])
+##
+##
+########################################################################
+########################################################################
+####                Assign instructions to locations                  ##
+########################################################################
+########################################################################
+##
+##def make_assign_instructions_to_locations(data):
+##    def make_decls(input_data):
+##        var_names = get_var_names(input_data)
+##        ret = ''
+##        ret += 'include "alldifferent.mzn";\n'
+##        ret += create_set('CORE', CORES)
+##        ret += create_set('INSTRUCTION', ['NOOP'] + list(var_names))
+##        MAX_NUMBER_OF_NOOPS_PER_INSTRUCTION = 3
+##        APPROXIMATE_MAX_LATENCY = 6 * INSTRUCTIONS_PER_CYCLE
+##        max_loc = len(var_names) * MAX_NUMBER_OF_NOOPS_PER_INSTRUCTION # + APPROXIMATE_MAX_LATENCY
+##        ret += 'int: MAX_LOC = %d;\n\n' % max_loc
+##        ret += 'array[1..MAX_LOC] of var INSTRUCTION: output_instructions;\n'
+##        ret += 'array[1..MAX_LOC] of var CORE: output_cores;\n'
+##        ret += 'array[1..MAX_LOC] of var int: output_core_latency;\n'
+##        ret += 'array[1..MAX_LOC] of var int: output_data_latency;\n'
+##        ret += 'var int: RET_loc;\n'
+##        ret += '\n'
+##        ret += 'constraint 1 <= RET_loc /\\ RET_loc <= MAX_LOC;\n'
+##        ret += '\n'
+##        return ret
+##
+##    def make_disjoint(input_data):
+##        var_names = get_var_names(input_data)
+##        ret = ''
+##        #ret += 'constraint alldifferent( [ output_instructions[i] | i in 1..MAX_LOC where output_instructions[i] != NOOP ] );\n'
+##        ret += 'constraint forall (i,j in 1..MAX_LOC where i < j) (output_instructions[i] == NOOP \\/ output_instructions[i] != output_instructions[j]);\n'
+##        ret += 'constraint sum( [ 1 | i in 1..MAX_LOC where output_instructions[i] != NOOP ] ) == length(INSTRUCTION) - 1;\n'
+##        ret += 'constraint forall (i in 1..MAX_LOC where RET_loc <= i) (output_instructions[i] == NOOP);\n'
+##        ret += 'constraint forall (i,j in 1..MAX_LOC where i < j) ((output_instructions[i] != NOOP /\\ output_instructions[j] != NOOP /\\ output_cores[i] == output_cores[j]) -> i + output_core_latency[i] <= j);\n'
+##        ret += 'constraint forall (i in 1..MAX_LOC) (output_instructions[i] == NOOP \\/ i + output_data_latency[i] <= RET_loc);\n'
+##        ret += 'constraint forall (i in 1..MAX_LOC) (output_instructions[i] == NOOP -> (output_core_latency[i] == 0 /\\ output_data_latency[i] == 0 /\\ output_cores[i] == NOOP_CORE));\n'
+##        return ret
+##
+##    def make_dependencies(input_data):
+##        var_names = get_var_names(input_data)
+##        ret = ''
+##        for line in input_data['lines']:
+##            for arg in line['args']:
+##                if arg in var_names and arg[0] not in '0123456789':
+##                    ret += ('constraint forall (i,j in 1..MAX_LOC) ((output_instructions[i] == %s /\\ output_instructions[j] == %s) -> i + output_data_latency[i] <= j);\n'
+##                            % (arg, line['out']))
+##        ret += '\n'
+##        return ret
+##
+##    def make_cores(input_data):
+##        ret = ''
+##        for line in input_data['lines']:
+##            ret += 'constraint forall (i in 1..MAX_LOC) (output_instructions[i] == %s -> (' % line['out']
+##            possible_cores = []
+##            cores = MODEL[line['op']]
+##            for core in cores:
+##                possible_cores.append(r'(output_core_latency[i] == %d /\ output_data_latency[i] == %d /\ output_cores[i] == %s)'
+##                                      % (core['core']['latency'] * INSTRUCTIONS_PER_CYCLE,
+##                                         core['latency'] * INSTRUCTIONS_PER_CYCLE,
+##                                         core['core']['name']))
+##            ret += ' \/ '.join(possible_cores)
+##            ret += '));\n'
+##        ret += '\n'
+##        return ret
+##
+##    def make_output(input_data):
+##        ret = 'solve minimize RET_loc;\n\n'
+##        ret += 'output [ "(" ++ show(INSTRUCTION_NAMES[fix(output_instructions[i])]) ++ ", " ++ show(CORE_NAMES[fix(output_cores[i])]) ++ ", " ++ show(output_core_latency[i]) ++ ", " ++ show(output_data_latency[i]) ++ ") ,\\n"\n'
+##        ret += '       | i in 1..MAX_LOC ];\n'
+##        return ret
+##
+##    return '\n'.join([
+##        make_decls(data),
+##        make_disjoint(data),
+##        make_dependencies(data),
+##        make_cores(data),
+##        make_output(data)
+##        ])
 
 ######################################################################
 ######################################################################
@@ -310,12 +315,14 @@ def make_assign_locations_to_instructions_cumulatively(data):
         ret += 'set of int: LOCATIONS = 1..MAX_LOC;\n'
         ret += 'array[INSTRUCTIONS] of var LOCATIONS: output_locations;\n'
         ret += 'array[INSTRUCTIONS] of var int: output_data_latency;\n'
+        ret += 'array[INSTRUCTIONS] of var int: output_core_latency;\n'
         ret += 'array[INSTRUCTIONS] of var CORE: output_cores;\n'
         ret += 'array[INSTRUCTIONS] of OPS: input_ops = [%s];\n' % ', '.join(OP_NAMES[line['op']] for line in input_data['lines'])
         for core in CORES:
             ret += 'array[INSTRUCTIONS] of var int: output_%s_core_latency;\n' % core
             ret += 'array[INSTRUCTIONS] of var 0..1: output_%s_core_use;\n' % core
             ret += 'constraint forall (i in INSTRUCTIONS) (0 <= output_%s_core_latency[i]);\n' % core
+            ret += 'constraint forall (i in INSTRUCTIONS) (output_%s_core_use[i] == 1 -> output_core_latency[i] == output_%s_core_latency[i]);\n' % (core, core)
         ret += 'var LOCATIONS: RET_loc;\n'
         ret += '\n'
         return ret
@@ -323,23 +330,17 @@ def make_assign_locations_to_instructions_cumulatively(data):
     def make_cores(input_data):
         ret = ''
         for opc, cores in MODEL.items():
-            possible_cores_and = []
-            possible_cores_or = []
-            for core in cores:
-                possible_cores_and.append((r'((output_cores[i] == %s /\ output_%s_core_use[i] == 1 /\ output_%s_core_latency[i] == %d /\ output_data_latency[i] == %d)' +
-                                           r' \/ (output_cores[i] != %s /\ output_%s_core_use[i] == 0 /\ output_%s_core_latency[i] == 0))')
-                                          % (core['core']['name'],
-                                             core['core']['name'],
-                                             core['core']['name'],
-                                             core['core']['latency'] * INSTRUCTIONS_PER_CYCLE,
-                                             core['latency'] * INSTRUCTIONS_PER_CYCLE,
-                                             core['core']['name'],
-                                             core['core']['name'],
-                                             core['core']['name']))
+            possible_cores = []
             for core in cores:
-                possible_cores_or.append('output_cores[i] == %s' % core['core']['name'])
-            ret += ('constraint forall (i in INSTRUCTIONS) (input_ops[i] == %s -> ((%s) /\\ (%s)));\n'
-                    % (OP_NAMES[opc], r' /\ '.join(possible_cores_and), r' \/ '.join(possible_cores_or)))
+                conjuncts = (['output_cores[i] == %s' % core['core']['name'],
+                              'output_%s_core_use[i] == 1' % core['core']['name'],
+                              'output_%s_core_latency[i] == %d' % (core['core']['name'], core['core']['latency'] * INSTRUCTIONS_PER_CYCLE),
+                              'output_data_latency[i] == %d' % (core['latency'] * INSTRUCTIONS_PER_CYCLE)] +
+                             ['output_%s_core_use[i] == 0 /\ output_%s_core_latency[i] == 0' % (other_core, other_core)
+                              for other_core in CORES if other_core != core['core']['name']])
+                possible_cores.append('(%s)' % (r' /\ '.join(conjuncts)))
+            ret += ('constraint forall (i in INSTRUCTIONS) (input_ops[i] == %s -> (%s));\n'
+                    % (OP_NAMES[opc], r' \/ '.join(possible_cores)))
         ret += '\n'
         for core in CORES:
             ret += ('constraint cumulative(output_locations, output_%s_core_latency, output_%s_core_use, %d);\n'
@@ -352,7 +353,7 @@ def make_assign_locations_to_instructions_cumulatively(data):
         ret += 'constraint alldifferent(output_locations);\n'
         return ret
 
-    def make_dependencies(input_data):
+    def make_dependencies_old(input_data):
         var_names = get_var_names(input_data)
         ret = ''
         for line in input_data['lines']:
@@ -365,9 +366,22 @@ def make_assign_locations_to_instructions_cumulatively(data):
         ret += '\n'
         return ret
 
+    def make_dependencies(input_data):
+        var_names = get_var_names(input_data)
+        ret = ''
+        ret += 'array[INSTRUCTIONS,INSTRUCTIONS] of var 0..1: depends_on;\n'
+        dependencies = {}
+        for line in input_data['lines']:
+            dependencies[line['out']] = tuple(arg for arg in line['args']
+                                              if arg in var_names and arg[0] not in '0123456789')
+        # HERE
+        
+        
+    
+
     def make_output(input_data):
         ret = 'solve minimize RET_loc;\n\n'
-        ret += 'output [ "(" ++ show(INSTRUCTIONS_NAMES[i]) ++ ", " ++ show(CORE_NAMES[fix(output_cores[i])]) ++ ", " ++ show(output_locations[i]) ++ ", " ++ show(output_data_latency[i]) ++ ") ,\\n"\n'
+        ret += 'output [ "(" ++ show(INSTRUCTIONS_NAMES[i]) ++ ", " ++ show(CORE_NAMES[fix(output_cores[i])]) ++ ", " ++ show(output_locations[i]) ++ ", " ++ show(output_data_latency[i]) ++ ", " ++ show(output_core_latency[i]) ++ ") ,\\n"\n'
         ret += '       | i in INSTRUCTIONS ];\n'
         ret += 'output [ "RET_loc: " ++ show(RET_loc) ];\n'
         return ret
@@ -381,9 +395,136 @@ def make_assign_locations_to_instructions_cumulatively(data):
         ])
 
 
+RESULTS = [
+"""("x20", "ADD_MUL", 9, 12, 4) ,
+("x21", "ADD_MUL", 21, 12, 4) ,
+("x22", "ADD_MUL", 10, 12, 4) ,
+("x23", "ADD_MUL", 34, 4, 4) ,
+("x24", "ADD_MUL", 47, 12, 4) ,
+("x25", "ADD_MUL", 43, 12, 4) ,
+("x26", "ADD_MUL", 59, 4, 4) ,
+("x27", "ADD_MUL", 51, 12, 4) ,
+("x28", "ADD_MUL", 63, 4, 4) ,
+("x29", "ADD_MUL", 75, 12, 4) ,
+("x30", "ADD_MUL", 71, 12, 4) ,
+("x31", "ADD_MUL", 87, 4, 4) ,
+("x32", "ADD_MUL", 79, 12, 4) ,
+("x33", "ADD_MUL", 91, 4, 4) ,
+("x34", "ADD_MUL", 82, 12, 4) ,
+("x35", "ADD_MUL", 95, 4, 4) ,
+("x36", "ADD_MUL", 58, 12, 4) ,
+("x37", "ADD_MUL", 55, 12, 4) ,
+("x38", "ADD_MUL", 70, 4, 4) ,
+("x39", "ADD_MUL", 62, 12, 4) ,
+("x40", "ADD_MUL", 74, 4, 4) ,
+("x41", "ADD_MUL", 66, 12, 4) ,
+("x42", "ADD_MUL", 78, 4, 4) ,
+("x43", "ADD_MUL", 90, 12, 4) ,
+("x44", "ADD_MUL", 102, 4, 4) ,
+("x45", "ADD_MUL", 1, 12, 4) ,
+("x46", "ADD_MUL", 2, 12, 4) ,
+("x47", "ADD_MUL", 5, 12, 4) ,
+("x48", "ADD_MUL", 6, 12, 4) ,
+("x49", "ADD_MUL", 13, 12, 4) ,
+("x50", "ADD_MUL", 25, 4, 4) ,
+("x51", "ADD_MUL", 14, 12, 4) ,
+("x52", "ADD_MUL", 29, 4, 4) ,
+("x53", "ADD_MUL", 17, 12, 4) ,
+("x54", "ADD_MUL", 33, 4, 4) ,
+("x55", "ADD_MUL", 18, 12, 4) ,
+("x56", "ADD_MUL", 37, 4, 4) ,
+("x57", "ADD_MUL", 22, 12, 4) ,
+("x58", "ADD_MUL", 38, 4, 4) ,
+("x59", "ADD_MUL", 30, 12, 4) ,
+("x60", "ADD_MUL", 42, 4, 4) ,
+("x61", "ADD_MUL", 26, 12, 4) ,
+("x62", "ADD_MUL", 46, 4, 4) ,
+("x63", "ADD_MUL", 54, 12, 4) ,
+("x64", "ADD_MUL", 67, 4, 4) ,
+("x65", "ADD_MUL", 86, 12, 4) ,
+("x66", "ADD_MUL", 98, 4, 4) ,
+("x67", "ADD_MUL", 83, 12, 4) ,
+("x68", "ADD_MUL", 99, 4, 4) ,
+("x69", "LEA_BW", 41, 4, 4) ,
+("x70", "LEA_BW", 44, 4, 4) ,
+("x71", "ADD_MUL", 50, 4, 4) ,
+("x72", "LEA_BW", 56, 4, 4) ,
+RET_loc: 106"""
+,
+"""("x73", "LEA_BW", 2, 4, 4) ,
+("x74", "ADD_MUL", 1, 4, 4) ,
+("x75", "LEA_BW", 5, 4, 4) ,
+("x76", "LEA_BW", 6, 4, 4) ,
+("x77", "ADD_MUL", 9, 4, 4) ,
+("x78", "LEA_BW", 13, 4, 4) ,
+("x79", "LEA_BW", 14, 4, 4) ,
+("x80", "ADD_MUL", 17, 4, 4) ,
+("x81", "LEA_BW", 21, 4, 4) ,
+("x82", "LEA_BW", 22, 4, 4) ,
+("x83", "ADD_MUL", 25, 12, 4) ,
+("x84", "ADD_MUL", 37, 4, 4) ,
+("x85", "LEA_BW", 41, 4, 4) ,
+("x86", "LEA_BW", 42, 4, 4) ,
+("x87", "ADD_MUL", 45, 4, 4) ,
+("x88", "LEA_BW", 49, 4, 4) ,
+("x89", "LEA_BW", 50, 4, 4) ,
+("x90", "ADD_MUL", 53, 4, 4) ,
+RET_loc: 57"""
+]
 
-data = parse_lines(get_lines('femulDisplay.txt'))
-with open('femulDisplay.mzn', 'w') as f:
-    #f.write(make_assign_locations_to_instructions(data))
-    #f.write(make_assign_instructions_to_locations(data))
-    f.write(make_assign_locations_to_instructions_cumulatively(data))
+######################################################################
+######################################################################
+##                     Parsing minizinc output                      ##
+######################################################################
+######################################################################
+
+def assemble_output_and_register_allocate(data_list, result_list):
+    def parse_result(result):
+        def parse_val(val):
+            val = val.strip()
+            if val[0] == '(' and val[-1] == ')':
+                return tuple(map(parse_val, val[1:-1].split(',')))
+            if val[0] in '"\'' and val[-1] in '"\'':
+                return val[1:-1]
+            if val.isdigit():
+                return int(val)
+            print('Unknown value: %s' % val)
+            return val
+        ret = {'schedule':list(map(str.strip, result.split(',\n')))}
+        ret['RET_loc'] = [i[len('RET_loc: '):] for i in ret['schedule'] if i[:len('RET_loc: ')] == 'RET_loc: ']
+        assert len(ret['RET_loc']) == 1
+        ret['RET_loc'] = int(ret['RET_loc'][0])
+        ret['schedule'] = tuple(parse_val(val) for val in ret['schedule'] if val[0] == '(' and val[-1] == ')')
+        return ret
+        
+    def combine_lists(data_list, result_list):
+        data = data_list[0]
+        data['lines'] = list(data['lines'])
+        for cur_data in data_list:
+            data['lines'] += list(cur_data['lines'])
+        data['lines'] = tuple(data['lines'])
+
+        basepoint = 0
+        results = []
+        for result in map(parse_result, result_list):
+            results += [(var, core_type, basepoint+loc, data_latency, core_latency)
+                        for var, core_type, loc, data_latency, core_latency in result['schedule']]
+            basepoint += result['RET_loc']
+        return (data, results)
+
+    def sort_results(data, results):
+        return sorted([(loc, '%s // %s, start: %.2f, end: %.2f' % (line['source'], core_type, loc / 4.0, (loc + data_latency) / 4.0))
+                       for (line, (var, core_type, loc, data_latency, core_latency))
+                       in zip(data['lines'], results)])
+
+    data, results = combine_lists(data_list, result_list)
+    return '\n'.join(code for i, code in sort_results(data, results))
+
+data_list = parse_lines(get_lines('femulDisplay.log'))
+for i, data in enumerate(data_list):
+    with open('femulDisplay_%d.mzn' % i, 'w') as f:
+        #f.write(make_assign_locations_to_instructions(data))
+        #f.write(make_assign_instructions_to_locations(data))
+        f.write(make_assign_locations_to_instructions_cumulatively(data))
+            
+print(assemble_output_and_register_allocate(data_list[:1], RESULTS[:1]))