1 #!/usr/bin/env python 2 3 """ 4 Optimise object layouts and generate access instruction plans. 5 6 Copyright (C) 2014, 2015, 2016, 2017 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 self.accessor_kinds = {} 61 62 # Object structure information. 63 64 self.structures = {} 65 self.attr_table = {} 66 67 # Parameter list information. 68 69 self.parameters = {} 70 self.param_table = {} 71 72 # Constant literal information. 73 74 self.constants = [] 75 self.constant_numbers = {} 76 77 # Optimiser activities. 78 79 self.populate_objects() 80 self.position_attributes() 81 self.populate_parameters() 82 self.position_parameters() 83 self.populate_tables() 84 self.populate_constants() 85 self.initialise_access_instructions() 86 87 def to_output(self): 88 89 "Write the output files using optimisation information." 90 91 if not exists(self.output): 92 makedirs(self.output) 93 94 self.write_objects() 95 96 def write_objects(self): 97 98 """ 99 Write object-related output. 100 101 The locations are a list of positions indicating the attributes residing 102 at each position in the different structures in a program. 103 104 ---- 105 106 The parameter locations are a list of positions indicating the parameters 107 residing at each position in the different parameter lists in a program. 108 109 ---- 110 111 Each attribute plan provides attribute details in the following format: 112 113 location " " name " " test " " test type " " base 114 " " traversed attributes " " traversed attribute ambiguity 115 " " traversal access modes 116 " " attributes to traverse " " attribute ambiguity 117 " " context " " access method " " static attribute 118 119 Locations have the following format: 120 121 qualified name of scope "." local name ":" name version 122 123 Traversal access modes are either "class" (obtain accessor class to 124 access attribute) or "object" (obtain attribute directly from accessor). 125 126 ---- 127 128 The structures are presented as a table in the following format: 129 130 qualified name " " attribute names 131 132 The attribute names are separated by ", " characters and indicate the 133 attribute provided at each position in the structure associated with the 134 given type. Where no attribute is provided at a particular location 135 within a structure, "-" is given. 136 137 ---- 138 139 The parameters are presented as a table in the following format: 140 141 qualified name " " parameter details 142 143 The parameter details are separated by ", " characters and indicate 144 the parameter name and list position for each parameter described at 145 each location in the parameter table associated with the given 146 function. Where no parameter details are provided at a particular 147 location within a parameter table, "-" is given. The name and list 148 position are separated by a colon (":"). 149 150 ---- 151 152 The attribute table is presented as a table in the following format: 153 154 qualified name " " attribute identifiers 155 156 Instead of attribute names, identifiers defined according to the order 157 given in the "attrnames" file are employed to denote the attributes 158 featured in each type's structure. Where no attribute is provided at a 159 particular location within a structure, "-" is given. 160 161 ---- 162 163 The parameter table is presented as a table in the following format: 164 165 qualified name " " parameter details 166 167 Instead of parameter names, identifiers defined according to the order 168 given in the "paramnames" file are employed to denote the parameters 169 featured in each function's parameter table. Where no parameter is 170 provided at a particular location within a table, "-" is given. 171 172 ---- 173 174 The ordered list of attribute names is given in the "attrnames" file. 175 176 ---- 177 178 The ordered list of parameter names is given in the "paramnames" file. 179 180 ---- 181 182 The ordered list of constant literals is given in the "constants" file. 183 """ 184 185 f = open(join(self.output, "locations"), "w") 186 try: 187 for attrnames in self.locations: 188 print >>f, sorted_output(attrnames) 189 190 finally: 191 f.close() 192 193 f = open(join(self.output, "parameter_locations"), "w") 194 try: 195 for argnames in self.arg_locations: 196 print >>f, sorted_output(argnames) 197 198 finally: 199 f.close() 200 201 f = open(join(self.output, "instruction_plans"), "w") 202 try: 203 access_instructions = self.access_instructions.items() 204 access_instructions.sort() 205 206 for location, instructions in access_instructions: 207 print >>f, encode_access_location(location), "..." 208 for instruction in instructions: 209 print >>f, encode_instruction(instruction) 210 print >>f 211 212 finally: 213 f.close() 214 215 f = open(join(self.output, "structures"), "w") 216 try: 217 structures = self.structures.items() 218 structures.sort() 219 220 for name, attrnames in structures: 221 print >>f, name, ", ".join([s or "-" for s in attrnames]) 222 223 finally: 224 f.close() 225 226 f = open(join(self.output, "parameters"), "w") 227 try: 228 parameters = self.parameters.items() 229 parameters.sort() 230 231 for name, argnames in parameters: 232 print >>f, name, ", ".join([s and ("%s:%d" % s) or "-" for s in argnames]) 233 234 finally: 235 f.close() 236 237 f = open(join(self.output, "attrtable"), "w") 238 try: 239 attr_table = self.attr_table.items() 240 attr_table.sort() 241 242 for name, attrcodes in attr_table: 243 print >>f, name, ", ".join([i is not None and str(i) or "-" for i in attrcodes]) 244 245 finally: 246 f.close() 247 248 f = open(join(self.output, "paramtable"), "w") 249 try: 250 param_table = self.param_table.items() 251 param_table.sort() 252 253 for name, paramcodes in param_table: 254 print >>f, name, ", ".join([s and ("%d:%d" % s) or "-" for s in paramcodes]) 255 256 finally: 257 f.close() 258 259 f = open(join(self.output, "attrnames"), "w") 260 try: 261 for name in self.all_attrnames: 262 print >>f, name 263 264 finally: 265 f.close() 266 267 f = open(join(self.output, "paramnames"), "w") 268 try: 269 for name in self.all_paramnames: 270 print >>f, name 271 272 finally: 273 f.close() 274 275 f = open(join(self.output, "constants"), "w") 276 try: 277 constants = [] 278 for (value, value_type, encoding), n in self.constants.items(): 279 constants.append((n, value_type, encoding, value)) 280 constants.sort() 281 for n, value_type, encoding, value in constants: 282 print >>f, value_type, encoding or "{}", repr(value) 283 284 finally: 285 f.close() 286 287 def populate_objects(self): 288 289 "Populate objects using attribute and usage information." 290 291 self.all_attrs = {} 292 293 # Partition attributes into separate sections so that class and instance 294 # attributes are treated separately. 295 296 for source, objkind in [ 297 (self.importer.all_class_attrs, "<class>"), 298 (self.importer.all_instance_attrs, "<instance>"), 299 (self.importer.all_module_attrs, "<module>") 300 ]: 301 302 for name, attrnames in source.items(): 303 304 # Remove temporary names from structures. 305 306 attrnames = filter(lambda x: not x.startswith("$t"), attrnames) 307 self.all_attrs[(objkind, name)] = attrnames 308 309 self.locations = get_allocated_locations(self.all_attrs, get_attributes_and_sizes) 310 311 def populate_parameters(self): 312 313 "Populate parameter tables using parameter information." 314 315 self.arg_locations = [set()] + get_allocated_locations(self.importer.function_parameters, get_parameters_and_sizes) 316 317 def position_attributes(self): 318 319 "Position specific attribute references." 320 321 # Reverse the location mappings. 322 323 attr_locations = self.attr_locations = {} 324 325 for i, attrnames in enumerate(self.locations): 326 for attrname in attrnames: 327 attr_locations[attrname] = i 328 329 # Record the structures. 330 331 for (objkind, name), attrnames in self.all_attrs.items(): 332 key = Reference(objkind, name) 333 l = self.structures[key] = [None] * len(attrnames) 334 for attrname in attrnames: 335 position = attr_locations[attrname] 336 if position >= len(l): 337 l.extend([None] * (position - len(l) + 1)) 338 l[position] = attrname 339 340 def initialise_access_instructions(self): 341 342 "Expand access plans into instruction sequences." 343 344 for access_location, access_plan in self.deducer.access_plans.items(): 345 346 # Obtain the access details. 347 348 name, test, test_type, base, \ 349 traversed, traversal_modes, attrnames, \ 350 context, context_test, \ 351 first_method, final_method, \ 352 origin, accessor_kinds = access_plan 353 354 instructions = [] 355 emit = instructions.append 356 357 if base: 358 original_accessor = base 359 else: 360 original_accessor = "<expr>" # use a generic placeholder 361 362 # Prepare for any first attribute access. 363 364 if traversed: 365 attrname = traversed[0] 366 del traversed[0] 367 elif attrnames: 368 attrname = attrnames[0] 369 del attrnames[0] 370 371 # Perform the first access explicitly if at least one operation 372 # requires it. 373 374 access_first_attribute = final_method in ("access", "assign") or traversed or attrnames 375 376 # Determine whether the first access involves assignment. 377 378 assigning = not traversed and not attrnames and final_method == "assign" 379 set_accessor = assigning and "<set_target_accessor>" or "<set_accessor>" 380 stored_accessor = assigning and "<target_accessor>" or "<accessor>" 381 382 # Set the context if already available. 383 384 if context == "base": 385 accessor = context_var = (base,) 386 elif context == "original-accessor": 387 388 # Prevent re-evaluation of any dynamic expression by storing it. 389 390 if original_accessor == "<expr>": 391 emit((set_accessor, original_accessor)) 392 accessor = context_var = (stored_accessor,) 393 else: 394 accessor = context_var = (original_accessor,) 395 396 # Assigning does not set the context. 397 398 elif context in ("final-accessor", "unset") and access_first_attribute: 399 400 # Prevent re-evaluation of any dynamic expression by storing it. 401 402 if original_accessor == "<expr>": 403 emit((set_accessor, original_accessor)) 404 accessor = (stored_accessor,) 405 else: 406 accessor = (original_accessor,) 407 408 # Apply any test. 409 410 if test[0] == "test": 411 accessor = ("__%s_%s_%s" % test, accessor, test_type) 412 413 # Perform the first or final access. 414 # The access only needs performing if the resulting accessor is used. 415 416 remaining = len(traversed + attrnames) 417 418 if access_first_attribute: 419 420 if first_method == "relative-class": 421 if assigning: 422 emit(("__store_via_class", accessor, attrname, "<assexpr>")) 423 else: 424 accessor = ("__load_via_class", accessor, attrname) 425 426 elif first_method == "relative-object": 427 if assigning: 428 emit(("__store_via_object", accessor, attrname, "<assexpr>")) 429 else: 430 accessor = ("__load_via_object", accessor, attrname) 431 432 elif first_method == "relative-object-class": 433 if assigning: 434 emit(("__get_class_and_store", accessor, attrname, "<assexpr>")) 435 else: 436 accessor = ("__get_class_and_load", accessor, attrname) 437 438 elif first_method == "check-class": 439 if assigning: 440 emit(("__check_and_store_via_class", accessor, attrname, "<assexpr>")) 441 else: 442 accessor = ("__check_and_load_via_class", accessor, attrname) 443 444 elif first_method == "check-object": 445 if assigning: 446 emit(("__check_and_store_via_object", accessor, attrname, "<assexpr>")) 447 else: 448 accessor = ("__check_and_load_via_object", accessor, attrname) 449 450 elif first_method == "check-object-class": 451 if assigning: 452 emit(("__check_and_store_via_any", accessor, attrname, "<assexpr>")) 453 else: 454 accessor = ("__check_and_load_via_any", accessor, attrname) 455 456 # Traverse attributes using the accessor. 457 458 if traversed: 459 for attrname, traversal_mode in zip(traversed, traversal_modes): 460 assigning = remaining == 1 and final_method == "assign" 461 462 # Set the context, if appropriate. 463 464 if remaining == 1 and final_method != "assign" and context == "final-accessor": 465 emit(("__set_context", accessor)) 466 accessor = context_var = "<context>" 467 468 # Perform the access only if not achieved directly. 469 470 if remaining > 1 or final_method in ("access", "assign"): 471 472 if traversal_mode == "class": 473 if assigning: 474 emit(("__store_via_class", accessor, attrname, "<assexpr>")) 475 else: 476 accessor = ("__load_via_class", accessor, attrname) 477 else: 478 if assigning: 479 emit(("__store_via_object", accessor, attrname, "<assexpr>")) 480 else: 481 accessor = ("__load_via_object", accessor, attrname) 482 483 remaining -= 1 484 485 if attrnames: 486 for attrname in attrnames: 487 assigning = remaining == 1 and final_method == "assign" 488 489 # Set the context, if appropriate. 490 491 if remaining == 1 and final_method != "assign" and context == "final-accessor": 492 emit(("__set_context", accessor)) 493 accessor = context_var = "<context>" 494 495 # Perform the access only if not achieved directly. 496 497 if remaining > 1 or final_method in ("access", "assign"): 498 499 if assigning: 500 emit(("__check_and_store_via_any", accessor, attrname, "<assexpr>")) 501 else: 502 accessor = ("__check_and_load_via_any", accessor, attrname) 503 504 remaining -= 1 505 506 # Define or emit the means of accessing the actual target. 507 508 if final_method == "static-assign": 509 parent, attrname = origin.rsplit(".", 1) 510 emit(("__store_via_object", parent, attrname, "<assexpr>")) 511 512 elif final_method in ("static", "static-invoke"): 513 parent, attrname = origin.rsplit(".", 1) 514 accessor = ("__load_static", parent, origin) 515 516 # Wrap accesses in context operations. 517 518 if context_test == "test": 519 emit(("__test_context", context_var, accessor)) 520 521 elif context_test == "replace": 522 emit(("__update_context", context_var, accessor)) 523 524 elif final_method not in ("assign", "static-assign"): 525 emit(accessor) 526 527 self.access_instructions[access_location] = instructions 528 self.accessor_kinds[access_location] = accessor_kinds 529 530 def get_ambiguity_for_attributes(self, attrnames): 531 532 """ 533 Return a list of attribute position alternatives corresponding to each 534 of the given 'attrnames'. 535 """ 536 537 ambiguity = [] 538 539 for attrname in attrnames: 540 position = self.attr_locations[attrname] 541 ambiguity.append(len(self.locations[position])) 542 543 return ambiguity 544 545 def position_parameters(self): 546 547 "Position the parameters for each function's parameter table." 548 549 # Reverse the location mappings. 550 551 param_locations = self.param_locations = {} 552 553 for i, argnames in enumerate(self.arg_locations): 554 555 # Position the arguments. 556 557 for argname in argnames: 558 param_locations[argname] = i 559 560 for name, argnames in self.importer.function_parameters.items(): 561 562 # Allocate an extra context parameter in the table. 563 564 l = self.parameters[name] = [None] + [None] * len(argnames) 565 566 # Store an entry for the name along with the name's position in the 567 # parameter list. 568 569 for pos, argname in enumerate(argnames): 570 571 # Position the argument in the table. 572 573 position = param_locations[argname] 574 if position >= len(l): 575 l.extend([None] * (position - len(l) + 1)) 576 577 # Indicate an argument list position starting from 1 (after the 578 # initial context argument). 579 580 l[position] = (argname, pos + 1) 581 582 def populate_tables(self): 583 584 """ 585 Assign identifiers to attributes and encode structure information using 586 these identifiers. 587 """ 588 589 self.all_attrnames, d = self._get_name_mapping(self.attr_locations) 590 591 # Record the numbers indicating the locations of the names. 592 593 for key, attrnames in self.structures.items(): 594 l = self.attr_table[key] = [] 595 for attrname in attrnames: 596 if attrname is None: 597 l.append(None) 598 else: 599 l.append(d[attrname]) 600 601 self.all_paramnames, d = self._get_name_mapping(self.param_locations) 602 603 # Record the numbers indicating the locations of the names. 604 605 for key, values in self.parameters.items(): 606 l = self.param_table[key] = [] 607 for value in values: 608 if value is None: 609 l.append(None) 610 else: 611 name, pos = value 612 l.append((d[name], pos)) 613 614 def _get_name_mapping(self, locations): 615 616 """ 617 Get a sorted list of names from 'locations', then map them to 618 identifying numbers. Return the list and the mapping. 619 """ 620 621 all_names = locations.keys() 622 all_names.sort() 623 d = {} 624 for i, name in enumerate(all_names): 625 d[name] = i 626 return all_names, d 627 628 def populate_constants(self): 629 630 """ 631 Obtain a collection of distinct constant literals, building a mapping 632 from constant references to those in this collection. 633 """ 634 635 # Obtain mappings from constant values to identifiers. 636 637 self.constants = {} 638 639 for path, constants in self.importer.all_constants.items(): 640 641 # Record constants and obtain a number for them. 642 # Each constant is actually (value, value_type, encoding). 643 644 for constant, n in constants.items(): 645 add_counter_item(self.constants, constant) 646 647 self.constant_numbers = {} 648 649 for name, constant in self.importer.all_constant_values.items(): 650 self.constant_numbers[name] = self.constants[constant] 651 652 def combine_rows(a, b): 653 c = [] 654 for i, j in zip(a, b): 655 if i is None or j is None: 656 c.append(i or j) 657 else: 658 return None 659 return c 660 661 def get_attributes_and_sizes(d): 662 663 """ 664 Return a matrix of attributes, a list of type names corresponding to columns 665 in the matrix, and a list of ranked sizes each indicating... 666 667 * a weighted size depending on the kind of object 668 * the minimum size of an object employing an attribute 669 * the number of free columns in the matrix for the attribute 670 * the attribute name itself 671 """ 672 673 attrs = {} 674 sizes = {} 675 objkinds = {} 676 677 for name, attrnames in d.items(): 678 objkind, _name = name 679 680 for attrname in attrnames: 681 682 # Record each type supporting the attribute. 683 684 init_item(attrs, attrname, set) 685 attrs[attrname].add(name) 686 687 # Maintain a record of the smallest object size supporting the given 688 # attribute. 689 690 if not sizes.has_key(attrname): 691 sizes[attrname] = len(attrnames) 692 else: 693 sizes[attrname] = min(sizes[attrname], len(attrnames)) 694 695 # Record the object types/kinds supporting the attribute. 696 697 init_item(objkinds, attrname, set) 698 objkinds[attrname].add(objkind) 699 700 # Obtain attribute details in order of size and occupancy. 701 702 names = d.keys() 703 704 rsizes = [] 705 for attrname, size in sizes.items(): 706 priority = "<instance>" in objkinds[attrname] and 0.5 or 1 707 occupied = len(attrs[attrname]) 708 key = (priority * size, size, len(names) - occupied, attrname) 709 rsizes.append(key) 710 711 rsizes.sort() 712 713 # Make a matrix of attributes. 714 715 matrix = {} 716 for attrname, types in attrs.items(): 717 row = [] 718 for name in names: 719 if name in types: 720 row.append(attrname) 721 else: 722 row.append(None) 723 matrix[attrname] = row 724 725 return matrix, names, rsizes 726 727 def get_parameters_and_sizes(d): 728 729 """ 730 Return a matrix of parameters, a list of functions corresponding to columns 731 in the matrix, and a list of ranked sizes each indicating... 732 733 * a weighted size depending on the kind of object 734 * the minimum size of a parameter list employing a parameter 735 * the number of free columns in the matrix for the parameter 736 * the parameter name itself 737 738 This is a slightly simpler version of the above 'get_attributes_and_sizes' 739 function. 740 """ 741 742 params = {} 743 sizes = {} 744 745 for name, argnames in d.items(): 746 for argname in argnames: 747 748 # Record each function supporting the parameter. 749 750 init_item(params, argname, set) 751 params[argname].add(name) 752 753 # Maintain a record of the smallest parameter list supporting the 754 # given parameter. 755 756 if not sizes.has_key(argname): 757 sizes[argname] = len(argnames) 758 else: 759 sizes[argname] = min(sizes[argname], len(argnames)) 760 761 # Obtain attribute details in order of size and occupancy. 762 763 names = d.keys() 764 765 rsizes = [] 766 for argname, size in sizes.items(): 767 occupied = len(params[argname]) 768 key = (size, size, len(names) - occupied, argname) 769 rsizes.append(key) 770 771 rsizes.sort() 772 773 # Make a matrix of parameters. 774 775 matrix = {} 776 for argname, types in params.items(): 777 row = [] 778 for name in names: 779 if name in types: 780 row.append(argname) 781 else: 782 row.append(None) 783 matrix[argname] = row 784 785 return matrix, names, rsizes 786 787 def get_allocated_locations(d, fn): 788 789 """ 790 Return a list where each element corresponds to a structure location and 791 contains a set of attribute names that may be stored at that location, given 792 a mapping 'd' whose keys are (object kind, object name) tuples and whose 793 values are collections of attributes. 794 """ 795 796 matrix, names, rsizes = fn(d) 797 allocated = [] 798 799 x = 0 800 while x < len(rsizes): 801 weight, size, free, attrname = rsizes[x] 802 base = matrix[attrname] 803 y = x + 1 804 while y < len(rsizes): 805 _weight, _size, _free, _attrname = rsizes[y] 806 occupied = len(names) - _free 807 if occupied > free: 808 break 809 new = combine_rows(base, matrix[_attrname]) 810 if new: 811 del matrix[_attrname] 812 del rsizes[y] 813 base = new 814 free -= occupied 815 else: 816 y += 1 817 allocated.append(base) 818 x += 1 819 820 # Return the list of attribute names from each row of the allocated 821 # attributes table. 822 823 locations = [] 824 for attrnames in allocated: 825 l = set() 826 for attrname in attrnames: 827 if attrname: 828 l.add(attrname) 829 locations.append(l) 830 return locations 831 832 # vim: tabstop=4 expandtab shiftwidth=4