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 # Perform the first access explicitly if at least one operation 367 # requires it. 368 369 access_first_attribute = final_method in ("access", "assign") or traversed or attrnames 370 371 # Determine whether the first access involves assignment. 372 373 assigning = not traversed and not attrnames and final_method == "assign" 374 375 # Set the context if already available. 376 377 if context == "original-accessor": 378 emit(("set_context", original_accessor)) 379 accessor = "context" 380 elif context == "base": 381 emit(("set_context", base)) 382 accessor = "context" 383 384 # Assigning does not set the context. 385 386 elif context in ("final-accessor", "unset"): 387 emit(("set_accessor", original_accessor)) 388 accessor = "accessor" 389 390 # Apply any test. 391 392 if test == "specific-type": 393 emit(("test_specific_type", accessor, test_type)) 394 elif test == "specific-instance": 395 emit(("test_specific_instance", accessor, test_type)) 396 elif test == "specific-object": 397 emit(("test_specific_object", accessor, test_type)) 398 elif test == "common-type": 399 emit(("test_common_type", accessor, test_type)) 400 elif test == "common-instance": 401 emit(("test_common_instance", accessor, test_type)) 402 elif test == "common-object": 403 emit(("test_common_object", accessor, test_type)) 404 405 # Perform the first or final access. 406 # The access only needs performing if the resulting accessor is used. 407 408 if access_first_attribute: 409 410 if first_method == "relative-class": 411 if assigning: 412 emit(("store_via_class", accessor, attrname, "<assexpr>")) 413 else: 414 emit(("set_accessor", ("load_via_class", accessor, attrname))) 415 416 elif first_method == "relative-object": 417 if assigning: 418 emit(("store_via_object", accessor, attrname, "<assexpr>")) 419 else: 420 emit(("set_accessor", ("load_via_object", accessor, attrname))) 421 422 elif first_method == "relative-object-class": 423 if assigning: 424 emit(("get_class_and_store", accessor, attrname, "<assexpr>")) 425 else: 426 emit(("set_accessor", ("get_class_and_load", accessor, attrname))) 427 428 elif first_method == "check-class": 429 if assigning: 430 emit(("check_and_store_via_class", accessor, attrname, "<assexpr>")) 431 else: 432 emit(("set_accessor", ("check_and_load_via_class", accessor, attrname))) 433 434 elif first_method == "check-object": 435 if assigning: 436 emit(("check_and_store_via_object", accessor, attrname, "<assexpr>")) 437 else: 438 emit(("set_accessor", ("check_and_load_via_object", accessor, attrname))) 439 440 elif first_method == "check-object-class": 441 if assigning: 442 emit(("check_and_store_via_any", accessor, attrname, "<assexpr>")) 443 else: 444 emit(("set_accessor", ("check_and_load_via_any", accessor, attrname))) 445 446 # Obtain an accessor. 447 448 remaining = len(traversed + attrnames) 449 450 if traversed: 451 for attrname, traversal_mode in zip(traversed, traversal_modes): 452 assigning = remaining == 1 and final_method == "assign" 453 454 # Set the context, if appropriate. 455 456 if remaining == 1 and final_method != "assign" and context == "final-accessor": 457 emit(("set_context", "accessor")) 458 459 # Perform the access only if not achieved directly. 460 461 if remaining > 1 or final_method in ("access", "assign"): 462 463 if traversal_mode == "class": 464 if assigning: 465 emit(("store_via_class", "accessor", attrname, "<assexpr>")) 466 else: 467 emit(("set_accessor", ("load_via_class", "accessor", attrname))) 468 else: 469 if assigning: 470 emit(("store_via_object", "accessor", attrname, "<assexpr>")) 471 else: 472 emit(("set_accessor", ("load_via_object", "accessor", attrname))) 473 474 remaining -= 1 475 476 if attrnames: 477 for attrname in attrnames: 478 assigning = remaining == 1 and final_method == "assign" 479 480 # Set the context, if appropriate. 481 482 if remaining == 1 and final_method != "assign" and context == "final-accessor": 483 emit(("set_context", "accessor")) 484 485 # Perform the access only if not achieved directly. 486 487 if remaining > 1 or final_method in ("access", "assign"): 488 489 if assigning: 490 emit(("check_and_store_via_any", "accessor", attrname, "<assexpr>")) 491 else: 492 emit(("set_accessor", ("check_and_load_via_any", "accessor", attrname))) 493 494 remaining -= 1 495 496 if final_method == "static-assign": 497 emit(("store_member", origin, "<assexpr>")) 498 elif final_method == "static": 499 emit(("load_static", origin)) 500 501 self.access_instructions[access_location] = instructions 502 503 def get_ambiguity_for_attributes(self, attrnames): 504 505 """ 506 Return a list of attribute position alternatives corresponding to each 507 of the given 'attrnames'. 508 """ 509 510 ambiguity = [] 511 512 for attrname in attrnames: 513 position = self.attr_locations[attrname] 514 ambiguity.append(len(self.locations[position])) 515 516 return ambiguity 517 518 def position_parameters(self): 519 520 "Position the parameters for each function's parameter table." 521 522 # Reverse the location mappings. 523 524 param_locations = self.param_locations = {} 525 526 for i, argnames in enumerate(self.arg_locations): 527 for argname in argnames: 528 param_locations[argname] = i 529 530 for name, argnames in self.importer.function_parameters.items(): 531 l = self.parameters[name] = [None] * len(argnames) 532 533 # Store an entry for the name along with the name's position in the 534 # parameter list. 535 536 for pos, argname in enumerate(argnames): 537 position = param_locations[argname] 538 if position >= len(l): 539 l.extend([None] * (position - len(l) + 1)) 540 l[position] = (argname, pos) 541 542 def populate_tables(self): 543 544 """ 545 Assign identifiers to attributes and encode structure information using 546 these identifiers. 547 """ 548 549 self.all_attrnames, d = self._get_name_mapping(self.attr_locations) 550 551 # Record the numbers indicating the locations of the names. 552 553 for key, attrnames in self.structures.items(): 554 l = self.attr_table[key] = [] 555 for attrname in attrnames: 556 if attrname is None: 557 l.append(None) 558 else: 559 l.append(d[attrname]) 560 561 self.all_paramnames, d = self._get_name_mapping(self.param_locations) 562 563 # Record the numbers indicating the locations of the names. 564 565 for key, values in self.parameters.items(): 566 l = self.param_table[key] = [] 567 for value in values: 568 if value is None: 569 l.append(None) 570 else: 571 name, pos = value 572 l.append((d[name], pos)) 573 574 def _get_name_mapping(self, locations): 575 576 """ 577 Get a sorted list of names from 'locations', then map them to 578 identifying numbers. Return the list and the mapping. 579 """ 580 581 all_names = locations.keys() 582 all_names.sort() 583 return all_names, dict([(name, i) for i, name in enumerate(all_names)]) 584 585 def populate_constants(self): 586 587 """ 588 Obtain a collection of distinct constant literals, building a mapping 589 from constant references to those in this collection. 590 """ 591 592 # Obtain mappings from constant values to identifiers. 593 594 self.constants = {} 595 596 for path, constants in self.importer.all_constants.items(): 597 for constant, n in constants.items(): 598 599 # Record constants and obtain a number for them. 600 601 add_counter_item(self.constants, constant) 602 603 self.constant_numbers = {} 604 605 for name, (value, value_type) in self.importer.all_constant_values.items(): 606 self.constant_numbers[name] = self.constants[value] 607 608 def combine_rows(a, b): 609 c = [] 610 for i, j in zip(a, b): 611 if i is None or j is None: 612 c.append(i or j) 613 else: 614 return None 615 return c 616 617 def get_attributes_and_sizes(d): 618 619 """ 620 Return a matrix of attributes, a list of type names corresponding to columns 621 in the matrix, and a list of ranked sizes each indicating... 622 623 * a weighted size depending on the kind of object 624 * the minimum size of an object employing an attribute 625 * the number of free columns in the matrix for the attribute 626 * the attribute name itself 627 """ 628 629 attrs = {} 630 sizes = {} 631 objtypes = {} 632 633 for name, attrnames in d.items(): 634 objtype, _name = name 635 636 for attrname in attrnames: 637 638 # Record each type supporting the attribute. 639 640 init_item(attrs, attrname, set) 641 attrs[attrname].add(name) 642 643 # Maintain a record of the smallest object size supporting the given 644 # attribute. 645 646 if not sizes.has_key(attrname): 647 sizes[attrname] = len(attrnames) 648 else: 649 sizes[attrname] = min(sizes[attrname], len(attrnames)) 650 651 # Record the object types/kinds supporting the attribute. 652 653 init_item(objtypes, attrname, set) 654 objtypes[attrname].add(objtype) 655 656 # Obtain attribute details in order of size and occupancy. 657 658 names = d.keys() 659 660 rsizes = [] 661 for attrname, size in sizes.items(): 662 priority = "<instance>" in objtypes[attrname] and 0.5 or 1 663 occupied = len(attrs[attrname]) 664 key = (priority * size, size, len(names) - occupied, attrname) 665 rsizes.append(key) 666 667 rsizes.sort() 668 669 # Make a matrix of attributes. 670 671 matrix = {} 672 for attrname, types in attrs.items(): 673 row = [] 674 for name in names: 675 if name in types: 676 row.append(attrname) 677 else: 678 row.append(None) 679 matrix[attrname] = row 680 681 return matrix, names, rsizes 682 683 def get_parameters_and_sizes(d): 684 685 """ 686 Return a matrix of parameters, a list of functions corresponding to columns 687 in the matrix, and a list of ranked sizes each indicating... 688 689 * a weighted size depending on the kind of object 690 * the minimum size of a parameter list employing a parameter 691 * the number of free columns in the matrix for the parameter 692 * the parameter name itself 693 694 This is a slightly simpler version of the above 'get_attributes_and_sizes' 695 function. 696 """ 697 698 params = {} 699 sizes = {} 700 701 for name, argnames in d.items(): 702 for argname in argnames: 703 704 # Record each function supporting the parameter. 705 706 init_item(params, argname, set) 707 params[argname].add(name) 708 709 # Maintain a record of the smallest parameter list supporting the 710 # given parameter. 711 712 if not sizes.has_key(argname): 713 sizes[argname] = len(argnames) 714 else: 715 sizes[argname] = min(sizes[argname], len(argnames)) 716 717 # Obtain attribute details in order of size and occupancy. 718 719 names = d.keys() 720 721 rsizes = [] 722 for argname, size in sizes.items(): 723 occupied = len(params[argname]) 724 key = (size, size, len(names) - occupied, argname) 725 rsizes.append(key) 726 727 rsizes.sort() 728 729 # Make a matrix of parameters. 730 731 matrix = {} 732 for argname, types in params.items(): 733 row = [] 734 for name in names: 735 if name in types: 736 row.append(argname) 737 else: 738 row.append(None) 739 matrix[argname] = row 740 741 return matrix, names, rsizes 742 743 def get_allocated_locations(d, fn): 744 745 """ 746 Return a list where each element corresponds to a structure location and 747 contains a set of attribute names that may be stored at that location, given 748 a mapping 'd' whose keys are (object type, object name) tuples and whose 749 values are collections of attributes. 750 """ 751 752 matrix, names, rsizes = fn(d) 753 allocated = [] 754 755 x = 0 756 while x < len(rsizes): 757 weight, size, free, attrname = rsizes[x] 758 base = matrix[attrname] 759 y = x + 1 760 while y < len(rsizes): 761 _weight, _size, _free, _attrname = rsizes[y] 762 occupied = len(names) - _free 763 if occupied > free: 764 break 765 new = combine_rows(base, matrix[_attrname]) 766 if new: 767 del matrix[_attrname] 768 del rsizes[y] 769 base = new 770 free -= occupied 771 else: 772 y += 1 773 allocated.append(base) 774 x += 1 775 776 # Return the list of attribute names from each row of the allocated 777 # attributes table. 778 779 return [set([attrname for attrname in attrnames if attrname]) for attrnames in allocated] 780 781 # vim: tabstop=4 expandtab shiftwidth=4