1 #!/usr/bin/env python 2 3 """ 4 Optimise object layouts and generate access instruction plans. 5 6 Copyright (C) 2014, 2015, 2016 Paul Boddie <paul@boddie.org.uk> 7 8 This program is free software; you can redistribute it and/or modify it under 9 the terms of the GNU General Public License as published by the Free Software 10 Foundation; either version 3 of the License, or (at your option) any later 11 version. 12 13 This program is distributed in the hope that it will be useful, but WITHOUT 14 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS 15 FOR A PARTICULAR PURPOSE. See the GNU General Public License for more 16 details. 17 18 You should have received a copy of the GNU General Public License along with 19 this program. If not, see <http://www.gnu.org/licenses/>. 20 """ 21 22 from common import add_counter_item, get_attrname_from_location, init_item, \ 23 sorted_output 24 from encoders import encode_access_location, encode_instruction, get_kinds 25 from os.path import exists, join 26 from os import makedirs 27 from referencing import Reference 28 29 class Optimiser: 30 31 "Optimise objects in a program." 32 33 def __init__(self, importer, deducer, output): 34 35 """ 36 Initialise an instance using the given 'importer' and 'deducer' that 37 will perform the arrangement of attributes for program objects, writing 38 the results to the given 'output' directory. 39 """ 40 41 self.importer = importer 42 self.deducer = deducer 43 self.output = output 44 45 # Locations/offsets of attributes in objects. 46 47 self.locations = None 48 self.attr_locations = None 49 self.all_attrnames = None 50 51 # Locations of parameters in parameter tables. 52 53 self.arg_locations = None 54 self.param_locations = None 55 self.all_paramnames = None 56 57 # Specific attribute access information. 58 59 self.access_instructions = {} 60 61 # Object structure information. 62 63 self.structures = {} 64 self.attr_table = {} 65 66 # Parameter list information. 67 68 self.parameters = {} 69 self.param_table = {} 70 71 # Constant literal information. 72 73 self.constants = [] 74 self.constant_numbers = {} 75 76 # Optimiser activities. 77 78 self.populate_objects() 79 self.position_attributes() 80 self.populate_parameters() 81 self.position_parameters() 82 self.populate_tables() 83 self.populate_constants() 84 self.initialise_access_instructions() 85 86 def to_output(self): 87 88 "Write the output files using optimisation information." 89 90 if not exists(self.output): 91 makedirs(self.output) 92 93 self.write_objects() 94 95 def write_objects(self): 96 97 """ 98 Write object-related output. 99 100 The locations are a list of positions indicating the attributes residing 101 at each position in the different structures in a program. 102 103 ---- 104 105 The parameter locations are a list of positions indicating the parameters 106 residing at each position in the different parameter lists in a program. 107 108 ---- 109 110 Each attribute plan provides attribute details in the following format: 111 112 location " " name " " test " " test type " " base 113 " " traversed attributes " " traversed attribute ambiguity 114 " " traversal access modes 115 " " attributes to traverse " " attribute ambiguity 116 " " context " " access method " " static attribute 117 118 Locations have the following format: 119 120 qualified name of scope "." local name ":" name version 121 122 Traversal access modes are either "class" (obtain accessor class to 123 access attribute) or "object" (obtain attribute directly from accessor). 124 125 ---- 126 127 The structures are presented as a table in the following format: 128 129 qualified name " " attribute names 130 131 The attribute names are separated by ", " characters and indicate the 132 attribute provided at each position in the structure associated with the 133 given type. Where no attribute is provided at a particular location 134 within a structure, "-" is given. 135 136 ---- 137 138 The parameters are presented as a table in the following format: 139 140 qualified name " " parameter details 141 142 The parameter details are separated by ", " characters and indicate 143 the parameter name and list position for each parameter described at 144 each location in the parameter table associated with the given 145 function. Where no parameter details are provided at a particular 146 location within a parameter table, "-" is given. The name and list 147 position are separated by a colon (":"). 148 149 ---- 150 151 The attribute table is presented as a table in the following format: 152 153 qualified name " " attribute identifiers 154 155 Instead of attribute names, identifiers defined according to the order 156 given in the "attrnames" file are employed to denote the attributes 157 featured in each type's structure. Where no attribute is provided at a 158 particular location within a structure, "-" is given. 159 160 ---- 161 162 The parameter table is presented as a table in the following format: 163 164 qualified name " " parameter details 165 166 Instead of parameter names, identifiers defined according to the order 167 given in the "paramnames" file are employed to denote the parameters 168 featured in each function's parameter table. Where no parameter is 169 provided at a particular location within a table, "-" is given. 170 171 ---- 172 173 The ordered list of attribute names is given in the "attrnames" file. 174 175 ---- 176 177 The ordered list of parameter names is given in the "paramnames" file. 178 179 ---- 180 181 The ordered list of constant literals is given in the "constants" file. 182 """ 183 184 f = open(join(self.output, "locations"), "w") 185 try: 186 for attrnames in self.locations: 187 print >>f, sorted_output(attrnames) 188 189 finally: 190 f.close() 191 192 f = open(join(self.output, "parameter_locations"), "w") 193 try: 194 for argnames in self.arg_locations: 195 print >>f, sorted_output(argnames) 196 197 finally: 198 f.close() 199 200 f = open(join(self.output, "instruction_plans"), "w") 201 try: 202 access_instructions = self.access_instructions.items() 203 access_instructions.sort() 204 205 for location, instructions in access_instructions: 206 print >>f, encode_access_location(location), "..." 207 for instruction in instructions: 208 print >>f, encode_instruction(instruction) 209 print >>f 210 211 finally: 212 f.close() 213 214 f = open(join(self.output, "structures"), "w") 215 try: 216 structures = self.structures.items() 217 structures.sort() 218 219 for name, attrnames in structures: 220 print >>f, name, ", ".join([s or "-" for s in attrnames]) 221 222 finally: 223 f.close() 224 225 f = open(join(self.output, "parameters"), "w") 226 try: 227 parameters = self.parameters.items() 228 parameters.sort() 229 230 for name, argnames in parameters: 231 print >>f, name, ", ".join([s and ("%s:%d" % s) or "-" for s in argnames]) 232 233 finally: 234 f.close() 235 236 f = open(join(self.output, "attrtable"), "w") 237 try: 238 attr_table = self.attr_table.items() 239 attr_table.sort() 240 241 for name, attrcodes in attr_table: 242 print >>f, name, ", ".join([i is not None and str(i) or "-" for i in attrcodes]) 243 244 finally: 245 f.close() 246 247 f = open(join(self.output, "paramtable"), "w") 248 try: 249 param_table = self.param_table.items() 250 param_table.sort() 251 252 for name, paramcodes in param_table: 253 print >>f, name, ", ".join([s and ("%d:%d" % s) or "-" for s in paramcodes]) 254 255 finally: 256 f.close() 257 258 f = open(join(self.output, "attrnames"), "w") 259 try: 260 for name in self.all_attrnames: 261 print >>f, name 262 263 finally: 264 f.close() 265 266 f = open(join(self.output, "paramnames"), "w") 267 try: 268 for name in self.all_paramnames: 269 print >>f, name 270 271 finally: 272 f.close() 273 274 f = open(join(self.output, "constants"), "w") 275 try: 276 constants = [(n, value) for (value, n) in self.constants.items()] 277 constants.sort() 278 for n, value in constants: 279 print >>f, repr(value) 280 281 finally: 282 f.close() 283 284 def populate_objects(self): 285 286 "Populate objects using attribute and usage information." 287 288 all_attrs = {} 289 290 # Partition attributes into separate sections so that class and instance 291 # attributes are treated separately. 292 293 for source, objtype in [ 294 (self.importer.all_class_attrs, "<class>"), 295 (self.importer.all_instance_attrs, "<instance>"), 296 (self.importer.all_module_attrs, "<module>") 297 ]: 298 for name, attrs in source.items(): 299 all_attrs[(objtype, name)] = attrs 300 301 self.locations = get_allocated_locations(all_attrs, get_attributes_and_sizes) 302 303 def populate_parameters(self): 304 305 "Populate parameter tables using parameter information." 306 307 self.arg_locations = get_allocated_locations(self.importer.function_parameters, get_parameters_and_sizes) 308 309 def position_attributes(self): 310 311 "Position specific attribute references." 312 313 # Reverse the location mappings. 314 315 attr_locations = self.attr_locations = {} 316 317 for i, attrnames in enumerate(self.locations): 318 for attrname in attrnames: 319 attr_locations[attrname] = i 320 321 # Record the structures. 322 323 for source, objtype in [ 324 (self.importer.all_class_attrs, "<class>"), 325 (self.importer.all_instance_attrs, "<instance>"), 326 (self.importer.all_module_attrs, "<module>") 327 ]: 328 329 for name, attrnames in source.items(): 330 key = Reference(objtype, name) 331 l = self.structures[key] = [None] * len(attrnames) 332 for attrname in attrnames: 333 position = attr_locations[attrname] 334 if position >= len(l): 335 l.extend([None] * (position - len(l) + 1)) 336 l[position] = attrname 337 338 def initialise_access_instructions(self): 339 340 "Expand access plans into instruction sequences." 341 342 for access_location, access_plan in self.deducer.access_plans.items(): 343 344 # Obtain the access details. 345 346 name, test, test_type, base, traversed, traversal_modes, \ 347 attrnames, context, first_method, final_method, origin = access_plan 348 349 instructions = [] 350 emit = instructions.append 351 352 if base: 353 original_accessor = base 354 else: 355 original_accessor = "<expr>" # use a generic placeholder 356 357 # Prepare for any first attribute access. 358 359 if traversed: 360 attrname = traversed[0] 361 del traversed[0] 362 elif attrnames: 363 attrname = attrnames[0] 364 del attrnames[0] 365 366 access_first_attribute = final_method == "access" or traversed or attrnames 367 368 # Set the context if already available. 369 370 if context == "original-accessor": 371 emit(("set_context", original_accessor)) 372 accessor = "context" 373 elif context == "base": 374 emit(("set_context", base)) 375 accessor = "context" 376 elif context == "final-accessor" or access_first_attribute: 377 emit(("set_accessor", original_accessor)) 378 accessor = "accessor" 379 380 # Apply any test. 381 382 if test == "specific-type": 383 emit(("test_specific_type", accessor, test_type)) 384 elif test == "specific-instance": 385 emit(("test_specific_instance", accessor, test_type)) 386 elif test == "specific-object": 387 emit(("test_specific_object", accessor, test_type)) 388 elif test == "common-type": 389 emit(("test_common_type", accessor, test_type)) 390 elif test == "common-instance": 391 emit(("test_common_instance", accessor, test_type)) 392 elif test == "common-object": 393 emit(("test_common_object", accessor, test_type)) 394 395 # Perform the first or final access. 396 # The access only needs performing if the resulting accessor is used. 397 398 if access_first_attribute: 399 400 if first_method == "relative-class": 401 emit(("set_accessor", ("load_via_class", accessor, attrname))) 402 elif first_method == "relative-object": 403 emit(("set_accessor", ("load_via_object", accessor, attrname))) 404 elif first_method == "relative-object-class": 405 emit(("set_accessor", ("get_class_and_load", accessor, attrname))) 406 elif first_method == "check-class": 407 emit(("set_accessor", ("check_and_load_via_class", accessor, attrname))) 408 elif first_method == "check-object": 409 emit(("set_accessor", ("check_and_load_via_object", accessor, attrname))) 410 elif first_method == "check-object-class": 411 emit(("set_accessor", ("check_and_load_via_any", accessor, attrname))) 412 413 # Obtain an accessor. 414 415 remaining = len(traversed + attrnames) 416 417 if traversed: 418 for attrname, traversal_mode in zip(traversed, traversal_modes): 419 420 # Set the context, if appropriate. 421 422 if remaining == 1 and context == "final-accessor": 423 emit(("set_context", "accessor")) 424 425 # Perform the access only if not achieved directly. 426 427 if remaining > 1 or final_method == "access": 428 if traversal_mode == "class": 429 emit(("set_accessor", ("load_via_class", "accessor", attrname))) 430 else: 431 emit(("set_accessor", ("load_via_object", "accessor", attrname))) 432 433 remaining -= 1 434 435 if attrnames: 436 for attrname in attrnames: 437 438 # Set the context, if appropriate. 439 440 if remaining == 1 and context == "final-accessor": 441 emit(("set_context", "accessor")) 442 443 # Perform the access only if not achieved directly. 444 445 if remaining > 1 or final_method == "access": 446 emit(("set_accessor", ("check_and_load_via_any", "accessor", attrname))) 447 448 remaining -= 1 449 450 if final_method == "assign": 451 emit(("store_member", origin, "<expr>")) 452 elif final_method == "static": 453 emit(("load_static", origin)) 454 455 self.access_instructions[access_location] = instructions 456 457 def get_ambiguity_for_attributes(self, attrnames): 458 459 """ 460 Return a list of attribute position alternatives corresponding to each 461 of the given 'attrnames'. 462 """ 463 464 ambiguity = [] 465 466 for attrname in attrnames: 467 position = self.attr_locations[attrname] 468 ambiguity.append(len(self.locations[position])) 469 470 return ambiguity 471 472 def position_parameters(self): 473 474 "Position the parameters for each function's parameter table." 475 476 # Reverse the location mappings. 477 478 param_locations = self.param_locations = {} 479 480 for i, argnames in enumerate(self.arg_locations): 481 for argname in argnames: 482 param_locations[argname] = i 483 484 for name, argnames in self.importer.function_parameters.items(): 485 l = self.parameters[name] = [None] * len(argnames) 486 487 # Store an entry for the name along with the name's position in the 488 # parameter list. 489 490 for pos, argname in enumerate(argnames): 491 position = param_locations[argname] 492 if position >= len(l): 493 l.extend([None] * (position - len(l) + 1)) 494 l[position] = (argname, pos) 495 496 def populate_tables(self): 497 498 """ 499 Assign identifiers to attributes and encode structure information using 500 these identifiers. 501 """ 502 503 self.all_attrnames, d = self._get_name_mapping(self.attr_locations) 504 505 # Record the numbers indicating the locations of the names. 506 507 for key, attrnames in self.structures.items(): 508 l = self.attr_table[key] = [] 509 for attrname in attrnames: 510 if attrname is None: 511 l.append(None) 512 else: 513 l.append(d[attrname]) 514 515 self.all_paramnames, d = self._get_name_mapping(self.param_locations) 516 517 # Record the numbers indicating the locations of the names. 518 519 for key, values in self.parameters.items(): 520 l = self.param_table[key] = [] 521 for value in values: 522 if value is None: 523 l.append(None) 524 else: 525 name, pos = value 526 l.append((d[name], pos)) 527 528 def _get_name_mapping(self, locations): 529 530 """ 531 Get a sorted list of names from 'locations', then map them to 532 identifying numbers. Return the list and the mapping. 533 """ 534 535 all_names = locations.keys() 536 all_names.sort() 537 return all_names, dict([(name, i) for i, name in enumerate(all_names)]) 538 539 def populate_constants(self): 540 541 """ 542 Obtain a collection of distinct constant literals, building a mapping 543 from constant references to those in this collection. 544 """ 545 546 # Obtain mappings from constant values to identifiers. 547 548 self.constants = {} 549 550 for path, constants in self.importer.all_constants.items(): 551 for constant, n in constants.items(): 552 553 # Record constants and obtain a number for them. 554 555 add_counter_item(self.constants, constant) 556 557 self.constant_numbers = {} 558 559 for name, (value, value_type) in self.importer.all_constant_values.items(): 560 self.constant_numbers[name] = self.constants[value] 561 562 def combine_rows(a, b): 563 c = [] 564 for i, j in zip(a, b): 565 if i is None or j is None: 566 c.append(i or j) 567 else: 568 return None 569 return c 570 571 def get_attributes_and_sizes(d): 572 573 """ 574 Return a matrix of attributes, a list of type names corresponding to columns 575 in the matrix, and a list of ranked sizes each indicating... 576 577 * a weighted size depending on the kind of object 578 * the minimum size of an object employing an attribute 579 * the number of free columns in the matrix for the attribute 580 * the attribute name itself 581 """ 582 583 attrs = {} 584 sizes = {} 585 objtypes = {} 586 587 for name, attrnames in d.items(): 588 objtype, _name = name 589 590 for attrname in attrnames: 591 592 # Record each type supporting the attribute. 593 594 init_item(attrs, attrname, set) 595 attrs[attrname].add(name) 596 597 # Maintain a record of the smallest object size supporting the given 598 # attribute. 599 600 if not sizes.has_key(attrname): 601 sizes[attrname] = len(attrnames) 602 else: 603 sizes[attrname] = min(sizes[attrname], len(attrnames)) 604 605 # Record the object types/kinds supporting the attribute. 606 607 init_item(objtypes, attrname, set) 608 objtypes[attrname].add(objtype) 609 610 # Obtain attribute details in order of size and occupancy. 611 612 names = d.keys() 613 614 rsizes = [] 615 for attrname, size in sizes.items(): 616 priority = "<instance>" in objtypes[attrname] and 0.5 or 1 617 occupied = len(attrs[attrname]) 618 key = (priority * size, size, len(names) - occupied, attrname) 619 rsizes.append(key) 620 621 rsizes.sort() 622 623 # Make a matrix of attributes. 624 625 matrix = {} 626 for attrname, types in attrs.items(): 627 row = [] 628 for name in names: 629 if name in types: 630 row.append(attrname) 631 else: 632 row.append(None) 633 matrix[attrname] = row 634 635 return matrix, names, rsizes 636 637 def get_parameters_and_sizes(d): 638 639 """ 640 Return a matrix of parameters, a list of functions corresponding to columns 641 in the matrix, and a list of ranked sizes each indicating... 642 643 * a weighted size depending on the kind of object 644 * the minimum size of a parameter list employing a parameter 645 * the number of free columns in the matrix for the parameter 646 * the parameter name itself 647 648 This is a slightly simpler version of the above 'get_attributes_and_sizes' 649 function. 650 """ 651 652 params = {} 653 sizes = {} 654 655 for name, argnames in d.items(): 656 for argname in argnames: 657 658 # Record each function supporting the parameter. 659 660 init_item(params, argname, set) 661 params[argname].add(name) 662 663 # Maintain a record of the smallest parameter list supporting the 664 # given parameter. 665 666 if not sizes.has_key(argname): 667 sizes[argname] = len(argnames) 668 else: 669 sizes[argname] = min(sizes[argname], len(argnames)) 670 671 # Obtain attribute details in order of size and occupancy. 672 673 names = d.keys() 674 675 rsizes = [] 676 for argname, size in sizes.items(): 677 occupied = len(params[argname]) 678 key = (size, size, len(names) - occupied, argname) 679 rsizes.append(key) 680 681 rsizes.sort() 682 683 # Make a matrix of parameters. 684 685 matrix = {} 686 for argname, types in params.items(): 687 row = [] 688 for name in names: 689 if name in types: 690 row.append(argname) 691 else: 692 row.append(None) 693 matrix[argname] = row 694 695 return matrix, names, rsizes 696 697 def get_allocated_locations(d, fn): 698 699 """ 700 Return a list where each element corresponds to a structure location and 701 contains a set of attribute names that may be stored at that location, given 702 a mapping 'd' whose keys are (object type, object name) tuples and whose 703 values are collections of attributes. 704 """ 705 706 matrix, names, rsizes = fn(d) 707 allocated = [] 708 709 x = 0 710 while x < len(rsizes): 711 weight, size, free, attrname = rsizes[x] 712 base = matrix[attrname] 713 y = x + 1 714 while y < len(rsizes): 715 _weight, _size, _free, _attrname = rsizes[y] 716 occupied = len(names) - _free 717 if occupied > free: 718 break 719 new = combine_rows(base, matrix[_attrname]) 720 if new: 721 del matrix[_attrname] 722 del rsizes[y] 723 base = new 724 free -= occupied 725 else: 726 y += 1 727 allocated.append(base) 728 x += 1 729 730 # Return the list of attribute names from each row of the allocated 731 # attributes table. 732 733 return [set([attrname for attrname in attrnames if attrname]) for attrnames in allocated] 734 735 # vim: tabstop=4 expandtab shiftwidth=4