1 module format.writer; 2 3 import format.chunk; 4 import format.config; 5 6 import std.container.rbtree; 7 8 string write(Chunk[] chunks, Config config) { 9 auto context = Context(config, null); 10 auto w = Writer(BlockSpecifier(chunks, LinePrefix(0, 0)), &context); 11 w.write(); 12 w.buffer ~= '\n'; 13 return w.buffer.data; 14 } 15 16 package: 17 18 struct LinePrefix { 19 uint indent; 20 uint offset; 21 22 size_t toHash() const @safe nothrow { 23 return indent * 0x9e3779b97f4a7c15 + offset + 0xbf4600628f7c64f5; 24 } 25 } 26 27 struct BlockSpecifier { 28 Chunk[] chunks; 29 LinePrefix prefix; 30 31 this(Chunk[] chunks, LinePrefix prefix) { 32 this.chunks = chunks; 33 this.prefix = prefix; 34 } 35 36 bool opEquals(const ref BlockSpecifier rhs) const { 37 return chunks is rhs.chunks && prefix == rhs.prefix; 38 } 39 40 size_t toHash() const @safe nothrow { 41 size_t h = cast(size_t) chunks.ptr; 42 43 h ^= (h >>> 33); 44 h *= 0xff51afd7ed558ccd; 45 h += chunks.length; 46 h ^= (h >>> 33); 47 h *= 0xc4ceb9fe1a85ec53; 48 49 return h + prefix.toHash(); 50 } 51 } 52 53 struct FormatResult { 54 uint cost; 55 uint overflow; 56 string text; 57 } 58 59 struct Context { 60 Config config; 61 FormatResult[BlockSpecifier] cache; 62 } 63 64 struct Writer { 65 Context* context; 66 alias context this; 67 68 uint cost; 69 uint overflow; 70 71 Chunk[] chunks; 72 LinePrefix prefix; 73 74 import std.array; 75 Appender!string buffer; 76 77 this(BlockSpecifier block, Context* context) in { 78 assert(context !is null); 79 } do { 80 chunks = block.chunks; 81 prefix = block.prefix; 82 83 this.context = context; 84 } 85 86 FormatResult write() { 87 cost = 0; 88 overflow = 0; 89 90 import std.array; 91 buffer = appender!string(); 92 93 size_t start = 0; 94 foreach (i, ref c; chunks) { 95 if (i == 0 || !c.startsUnwrappedLine) { 96 continue; 97 } 98 99 LineWriter(&this, chunks[start .. i]).write(); 100 start = i; 101 } 102 103 // Make sure we write the last line too. 104 LineWriter(&this, chunks[start .. $]).write(); 105 106 return FormatResult(cost, overflow, buffer.data); 107 } 108 109 FormatResult formatBlock(Chunk[] chunks, LinePrefix prefix) { 110 auto block = BlockSpecifier(chunks, prefix); 111 return cache.require(block, Writer(block, context).write()); 112 } 113 114 void output(char c) { 115 buffer ~= c; 116 } 117 118 void output(string s) { 119 buffer ~= s; 120 } 121 122 void indent(uint level) { 123 if (config.useTabs) { 124 foreach (_; 0 .. level) { 125 output('\t'); 126 } 127 } else { 128 foreach (_; 0 .. level * config.indentationSize) { 129 output(' '); 130 } 131 } 132 } 133 134 void outputOffset(uint columns) { 135 foreach (_; 0 .. columns) { 136 output(' '); 137 } 138 } 139 } 140 141 enum MAX_ATTEMPT = 5000; 142 143 struct LineWriter { 144 Writer* writer; 145 alias writer this; 146 147 Chunk[] line; 148 149 this(Writer* writer, Chunk[] line) in { 150 assert(line.length > 0, "line must not be empty"); 151 } do { 152 this.writer = writer; 153 this.line = line; 154 } 155 156 void write() { 157 auto state = findBestState(); 158 159 cost += state.cost; 160 overflow += state.overflow; 161 162 writeLine(state); 163 } 164 165 void writeLine(SolveState state) { 166 foreach (i, c; line) { 167 uint lineCount = 0; 168 if (i > 0 || line.ptr !is chunks.ptr) { 169 assert(i == 0 || !c.startsUnwrappedLine, "Line splitting bug"); 170 lineCount = state.newLineCount(line, i); 171 } 172 173 foreach (_; 0 .. lineCount) { 174 output('\n'); 175 } 176 177 if ((i == 0 || lineCount > 0) && !c.glued) { 178 auto prefix = state.getLinePrefix(line, i); 179 indent(prefix.indent); 180 outputOffset(prefix.offset); 181 } 182 183 if (lineCount == 0 && c.separator == Separator.Space) { 184 output(' '); 185 } 186 187 final switch (c.kind) with (ChunkKind) { 188 case Text: 189 output(c.text); 190 break; 191 192 case Block: 193 auto f = 194 formatBlock(c.chunks, state.getLinePrefix(line, i)); 195 196 cost += f.cost; 197 overflow += f.overflow; 198 199 output(f.text); 200 break; 201 } 202 } 203 } 204 205 SolveState findBestState() { 206 auto best = SolveState(this); 207 if (best.overflow == 0 || !best.canExpand) { 208 // Either the line already fit, or it is not breakable. 209 return best; 210 } 211 212 static precisePred(const SolveState a, const SolveState b) { 213 return a < b; 214 } 215 216 auto checkpoints = CheckPoints(&this); 217 return exploreStates!precisePred(best, checkpoints, MAX_ATTEMPT); 218 } 219 220 SolveState exploreStates( 221 alias pred 222 )(SolveState best, ref CheckPoints checkpoints, uint max_attempts) { 223 uint attempts = 0; 224 scope queue = redBlackTree!pred(best); 225 226 // Once we have a solution that fits, or no more things 227 // to try, then we are done. 228 while (!queue.empty) { 229 auto next = queue.front; 230 231 // We found the lowest cost solution that fit on the page. 232 if (next.overflow == 0 || attempts > max_attempts) { 233 break; 234 } 235 236 // We pop from the queue *AFTER* checking for termination condition 237 // so that we do nto lose that state for the next iterration. 238 queue.removeFront(); 239 240 // We reach a dead subtree, there is no point erxploring further. 241 if (next.isDeadSubTree(best)) { 242 break; 243 } 244 245 // There is no point trying to expand this if it cannot 246 // lead to a a solution better than the current best. 247 if (checkpoints.isRedundant(next, false)) { 248 continue; 249 } 250 251 // This algorithm is exponential in nature, so make sure to stop 252 // after some time, even if we haven't found an optimal solution. 253 attempts++; 254 255 bool split = false; 256 foreach (rule; next.ruleValues.frozen .. line.length) { 257 if (split) { 258 // We reach a split point, bail. 259 break; 260 } 261 262 split = next.isSplit(rule); 263 if (!split && !next.canSplit(line, rule)) { 264 // If we cannot split here, move on. 265 continue; 266 } 267 268 auto newRuleValues = next.ruleValues.withFrozenSplit(rule); 269 auto candidate = SolveState(this, newRuleValues); 270 271 if (candidate.isBetterThan(best)) { 272 best = candidate; 273 } 274 275 // We check for redundant path first so that, even in the 276 // case where this candidate is rulled out, it can serve 277 // as a checkpoint. 278 if (checkpoints.isRedundant(candidate, true)) { 279 continue; 280 } 281 282 // This candidate can never expand to something better than the best. 283 if (candidate.isDeadSubTree(best)) { 284 continue; 285 } 286 287 queue.insert(candidate); 288 } 289 } 290 291 return best; 292 } 293 } 294 295 struct CheckPoints { 296 LineWriter* lineWriter; 297 SolveState[ulong][] paths; 298 299 this(LineWriter* lineWriter) { 300 this.lineWriter = lineWriter; 301 this.paths.length = lineWriter.line.length; 302 } 303 304 import format.span; 305 static getSpanStateHash(const ref SolveState s, const Span span, ulong h) { 306 if (span is null) { 307 h ^= (h >>> 33); 308 h *= 0x4cd6944c5cc20b6d; 309 h ^= (h >>> 33); 310 h *= 0xfc12c5b19d3259e9; 311 312 return h; 313 } 314 315 h ^= (h >>> 33); 316 h *= 0x7fb5d329728ea185; 317 h ^= (h >>> 33); 318 h *= 0x81dadef4bc2dd44d; 319 320 h += span.getState(s); 321 322 h ^= (h >>> 33); 323 h *= 0x99bcf6822b23ca35; 324 h ^= (h >>> 33); 325 h *= 0x14020a57acced8b7; 326 327 h += s.isUsed(span); 328 329 return getSpanStateHash(s, span.parent, h); 330 } 331 332 ulong getPathHash(const ref SolveState s, size_t i) { 333 const prefix = s.getLinePrefix(lineWriter.line, i); 334 return getSpanStateHash(s, lineWriter.line[i].span, prefix.toHash()); 335 } 336 337 static isSameSpanState(const ref SolveState a, const ref SolveState b, 338 const Span span) { 339 if (span is null) { 340 return true; 341 } 342 343 if (a.isUsed(span) != b.isUsed(span)) { 344 return false; 345 } 346 347 if (span.getState(a) != span.getState(b)) { 348 return false; 349 } 350 351 return isSameSpanState(a, b, span.parent); 352 } 353 354 bool isSamePath(const ref SolveState a, const ref SolveState b, size_t i) { 355 if (a == b) { 356 return false; 357 } 358 359 if (a.getLinePrefix(lineWriter.line, i) 360 != b.getLinePrefix(lineWriter.line, i)) { 361 return false; 362 } 363 364 return isSameSpanState(a, b, lineWriter.line[i].span); 365 } 366 367 bool isRedundant(const ref SolveState s, bool isRedundantWithSelf) { 368 if (!s.canExpand || s.ruleValues.frozen >= lineWriter.line.length) { 369 // There is nothing more to explore down this path. 370 return true; 371 } 372 373 const i = s.ruleValues.frozen - 1; 374 const h = getPathHash(s, i); 375 376 const c = h in paths[i]; 377 const cmp = s.opCmp(c); 378 if (cmp == 0) { 379 // A state is always redundant with itself. 380 return isRedundantWithSelf; 381 } 382 383 if (cmp < 0) { 384 // We have a new best on that specific path. 385 paths[i][h] = cast() s; 386 return false; 387 } 388 389 // We are on a path that is worse than the checkpoint. 390 // If this isn't a collision, this path is redundant. 391 return isSamePath(s, *c, i); 392 } 393 } 394 395 struct SolveState { 396 uint cost = 0; 397 uint overflow = 0; 398 uint sunk = 0; 399 400 LinePrefix prefix; 401 402 import format.rulevalues; 403 RuleValues ruleValues; 404 405 import format.span, std.bitmanip; 406 mixin(taggedClassRef!( 407 // sdfmt off 408 // Spans that require indentation. 409 RedBlackTree!(const Span), "usedSpans", 410 bool, "canExpand", 1, 411 // sdfmt on 412 )); 413 414 this(ref LineWriter lineWriter) { 415 this(lineWriter, RuleValues(1, lineWriter.line.length)); 416 } 417 418 this(ref LineWriter lineWriter, RuleValues ruleValues) { 419 this.ruleValues = ruleValues; 420 this.prefix = lineWriter.prefix; 421 422 computeCost(lineWriter.line, lineWriter.writer); 423 } 424 425 void computeCost(Chunk[] line, Writer* writer) { 426 sunk = 0; 427 overflow = 0; 428 cost = 0; 429 430 // If there is nothing to be done, just skip. 431 if (line.length == 0) { 432 return; 433 } 434 435 // Preparatory work. 436 computeUsedSpans(line); 437 438 // All the span which do not fit on one line. 439 RedBlackTree!Span brokenSpans; 440 441 uint length = 0; 442 uint previousColumn = 0; 443 Span previousSpan = null; 444 size_t regionStart = 0; 445 446 const indentationSize = writer.config.indentationSize; 447 const pageWidth = writer.config.pageWidth; 448 449 foreach (i, ref c; line) { 450 uint lineLength = 0; 451 uint column = 0; 452 453 if (c.startsRegion) { 454 regionStart = i; 455 } 456 457 if (c.kind == ChunkKind.Block) { 458 auto f = writer.formatBlock(c.chunks, getLinePrefix(line, i)); 459 460 // Compute the column at which the block starts. 461 auto text = f.text; 462 463 foreach (n; 0 .. f.text.length) { 464 if (text[n] == ' ') { 465 column++; 466 continue; 467 } 468 469 if (text[n] == '\t') { 470 column += indentationSize; 471 continue; 472 } 473 474 break; 475 } 476 477 cost += f.cost; 478 overflow += f.overflow; 479 updateSunk(line, i); 480 } else { 481 if (newLineCount(line, i) == 0) { 482 length += (c.separator == Separator.Space) + c.length; 483 continue; 484 } 485 486 column = 0; 487 lineLength = c.length; 488 489 if (!c.glued) { 490 const prefix = getLinePrefix(line, i); 491 column = prefix.indent * indentationSize + prefix.offset; 492 } 493 } 494 495 if (i > 0) { 496 // Try to avoid subsequent line to have the same indentation 497 // level if they belong to a different span. 498 uint penality = computeNewLinePenality( 499 c, column, length, previousColumn, previousSpan); 500 501 // End the previous line if there is one. 502 endLine(line, i, length, pageWidth, penality); 503 } 504 505 previousSpan = c.span; 506 previousColumn = column; 507 length = column + lineLength; 508 509 if (c.continuation) { 510 continue; 511 } 512 513 cost += 1; 514 515 auto span = c.span; 516 bool needInsert = true; 517 518 // Make sure to keep track of the span that cross over line breaks. 519 while (span !is null && needInsert) { 520 scope(success) span = span.parent; 521 522 if (brokenSpans is null) { 523 brokenSpans = redBlackTree!Span(); 524 } 525 526 needInsert = brokenSpans.insert(span) > 0; 527 } 528 } 529 530 endLine(line, line.length, length, pageWidth); 531 532 // Account for the cost of breaking spans. 533 if (brokenSpans !is null) { 534 foreach (s; brokenSpans) { 535 cost += s.getCost(this); 536 } 537 } 538 } 539 540 void computeUsedSpans(Chunk[] line) { 541 auto freezePoint = ruleValues.frozen; 542 ruleValues.frozen = 1; 543 544 scope(success) { 545 ruleValues.frozen = freezePoint; 546 } 547 548 foreach (i, ref c; line) { 549 // Continuation are not considered line splits. 550 if (c.continuation) { 551 continue; 552 } 553 554 if (!isSplit(i)) { 555 if (!mustSplit(line, i)) { 556 continue; 557 } 558 559 // Mark this as split. 560 ruleValues[i] = true; 561 } 562 563 // If there are no spans to break, move on. 564 if (c.span is null) { 565 continue; 566 } 567 568 if (usedSpans is null) { 569 usedSpans = redBlackTree!(const(Span))(); 570 } 571 572 usedSpans.insert(c.span); 573 } 574 } 575 576 uint computeNewLinePenality( 577 const ref Chunk c, 578 uint column, 579 uint length, 580 uint previousColumn, 581 const(Span) previousSpan 582 ) const { 583 // No penality for top level. 584 if (column == 0) { 585 return 0; 586 } 587 588 // Avoid line break that make the next lien starts past the previous line. 589 if (column >= length && c.kind != ChunkKind.Block) { 590 return 1; 591 } 592 593 // No penality for mismatching levels. 594 if (column != previousColumn) { 595 return 0; 596 } 597 598 // No penality for double line breaks. 599 if (c.separator == Separator.TwoNewLines) { 600 return 0; 601 } 602 603 // If both spans are expected to be on the same level, 604 // then it's all good. 605 if (previousSpan.isSameLevel(c.span)) { 606 return 0; 607 } 608 609 // This new line is at the same level as the previous line, yet belong to another span. 610 // This tends to make the code confusing to read, so we penalize this solution. 611 return 1; 612 } 613 614 void updateSunk(const Chunk[] line, size_t i) { 615 if (canExpand) { 616 return; 617 } 618 619 foreach (j; ruleValues.frozen .. i) { 620 if (canSplit(line, j)) { 621 canExpand = true; 622 return; 623 } 624 } 625 626 // If the line overflow, but has no split point, it is sunk. 627 sunk = overflow; 628 } 629 630 void endLine(const Chunk[] line, size_t i, uint length, uint pageWidth, 631 uint penality = 0) { 632 if (length > pageWidth) { 633 penality += length - pageWidth; 634 } 635 636 if (penality > 0) { 637 overflow += penality; 638 updateSunk(line, i); 639 } 640 } 641 642 bool canSplit(const Chunk[] line, size_t i) const { 643 if (isSplit(i)) { 644 return false; 645 } 646 647 auto c = line[i]; 648 if (!c.canSplit()) { 649 return false; 650 } 651 652 return c.span.canSplit(this, i); 653 } 654 655 bool mustSplit(const Chunk[] line, size_t i) const { 656 auto c = line[i]; 657 return c.newLineCount() > 0 || c.span.mustSplit(this, i); 658 } 659 660 bool isSplit(size_t i) const { 661 return ruleValues[i]; 662 } 663 664 bool isSplit(size_t from, size_t to) const { 665 foreach (i; from .. to) { 666 if (isSplit(i)) { 667 return true; 668 } 669 } 670 671 return false; 672 } 673 674 uint newLineCount(const Chunk[] line, size_t i) const { 675 if (auto c = line[i].newLineCount()) { 676 return c; 677 } 678 679 return isSplit(i); 680 } 681 682 bool isUsed(const Span span) const { 683 return usedSpans !is null && span in usedSpans; 684 } 685 686 LinePrefix getLinePrefix(Chunk[] line, size_t i) const { 687 uint offset = 0; 688 const c = line[i]; 689 690 // Find the preceding line break. 691 size_t r = c.span.getAlignIndex(this, i); 692 while (r > 0 && newLineCount(line, r) == 0) { 693 offset += line[r].separator == Separator.Space; 694 offset += line[--r].length; 695 } 696 697 if (line[r].glued) { 698 return LinePrefix(0, offset); 699 } 700 701 if (r == 0 || r == i) { 702 // We don't need to do any alignement magic. 703 const indent = c.indentation + c.span.getIndent(this, i); 704 return LinePrefix(prefix.indent + indent, prefix.offset + offset); 705 } 706 707 const base = getLinePrefix(line, r); 708 const indent = line[i].span.getExtraIndent(line[r].span, this, i); 709 710 return LinePrefix(base.indent + indent, base.offset + offset); 711 } 712 713 // Return if this solve state must be chosen over rhs as a solution. 714 bool isDeadSubTree(const ref SolveState best) const { 715 if (sunk > best.overflow) { 716 // We already have comitted to an overflow greater than the best. 717 return true; 718 } 719 720 if (sunk == best.overflow && cost >= best.cost) { 721 // We already comitted to a cost greater than the best. 722 return true; 723 } 724 725 // There is still hope to find a better solution down that path. 726 return false; 727 } 728 729 // Return if this solve state must be chosen over rhs as a solution. 730 bool isBetterThan(const ref SolveState rhs) const { 731 if (overflow != rhs.overflow) { 732 return overflow < rhs.overflow; 733 } 734 735 if (cost != rhs.cost) { 736 return cost < rhs.cost; 737 } 738 739 return ruleValues < rhs.ruleValues; 740 } 741 742 // lhs < rhs => lhs.opCmp(rhs) < 0 743 int opCmp(const ref SolveState rhs) const { 744 if (sunk != rhs.sunk) { 745 return sunk - rhs.sunk; 746 } 747 748 if (cost != rhs.cost) { 749 return cost - rhs.cost; 750 } 751 752 if (overflow != rhs.overflow) { 753 return overflow - rhs.overflow; 754 } 755 756 return ruleValues.opCmp(rhs.ruleValues); 757 } 758 759 int opCmp(const SolveState* rhs) const { 760 if (rhs is null) { 761 return -1; 762 } 763 764 return this.opCmp(*rhs); 765 } 766 }