mono/profiler/decode.c

   1 /*
   2  * decode.c: mprof-report program source: decode and analyze the log profiler data
   3  *
   4  * Author:
   5  *   Paolo Molaro (lupus@ximian.com)
   6  *
   7  * Copyright 2010 Novell, Inc (http://www.novell.com)
   8  */
   9 #include <config.h>
  10 #include "utils.c"
  11 #include "proflog.h"
  12 #include <string.h>
  13 #include <assert.h>
  14 #include <stdio.h>
  15 #if !defined(__APPLE__)
  16 #include <malloc.h>
  17 #endif
  18 #include <unistd.h>
  19 #include <stdlib.h>
  20 #if defined (HAVE_SYS_ZLIB)
  21 #include <zlib.h>
  22 #endif
  23 #include <mono/metadata/profiler.h>
  24 #include <mono/metadata/object.h>
  25 #include <mono/metadata/debug-helpers.h>
  26
  27 #define HASH_SIZE 9371
  28 #define SMALL_HASH_SIZE 31
  29
  30 static int debug = 0;
  31 static int collect_traces = 0;
  32 static int show_traces = 0;
  33 static int trace_max = 6;
  34 static int verbose = 0;
  35 static uintptr_t *tracked_objects = 0;
  36 static int num_tracked_objects = 0;
  37 static uintptr_t thread_filter = 0;
  38 static uint64_t find_size = 0;
  39 static const char* find_name = NULL;
  40 static uint64_t time_from = 0;
  41 static uint64_t time_to = 0xffffffffffffffffULL;
  42 static uint64_t startup_time = 0;
  43 static FILE* outfile = NULL;
  44
  45 static int32_t
  46 read_int32 (unsigned char *p)
  47 {
  48         int32_t value = *p++;
  49         value |= (*p++) << 8;
  50         value |= (*p++) << 16;
  51         value |= (uint32_t)(*p++) << 24;
  52         return value;
  53 }
  54
  55 static int64_t
  56 read_int64 (unsigned char *p)
  57 {
  58         uint64_t value = *p++;
  59         value |= (*p++) << 8;
  60         value |= (*p++) << 16;
  61         value |= (uint64_t)(*p++) << 24;
  62         value |= (uint64_t)(*p++) << 32;
  63         value |= (uint64_t)(*p++) << 40;
  64         value |= (uint64_t)(*p++) << 48;
  65         value |= (uint64_t)(*p++) << 54;
  66         return value;
  67 }
  68
  69 static char*
  70 pstrdup (const char *s)
  71 {
  72         int len = strlen (s) + 1;
  73         char *p = malloc (len);
  74         memcpy (p, s, len);
  75         return p;
  76 }
  77
  78 static int num_images;
  79 typedef struct _ImageDesc ImageDesc;
  80 struct _ImageDesc {
  81         ImageDesc *next;
  82         intptr_t image;
  83         char *filename;
  84 };
  85
  86 static ImageDesc* image_hash [SMALL_HASH_SIZE] = {0};
  87
  88 static void
  89 add_image (intptr_t image, char *name)
  90 {
  91         int slot = ((image >> 2) & 0xffff) % SMALL_HASH_SIZE;
  92         ImageDesc *cd = malloc (sizeof (ImageDesc));
  93         cd->image = image;
  94         cd->filename = pstrdup (name);
  95         cd->next = image_hash [slot];
  96         image_hash [slot] = cd;
  97         num_images++;
  98 }
  99
 100 typedef struct _BackTrace BackTrace;
 101 typedef struct {
 102         uint64_t count;
 103         BackTrace *bt;
 104 } CallContext;
 105
 106 typedef struct {
 107         int count;
 108         int size;
 109         CallContext *traces;
 110 } TraceDesc;
 111
 112 typedef struct _ClassDesc ClassDesc;
 113 struct _ClassDesc {
 114         ClassDesc *next;
 115         intptr_t klass;
 116         char *name;
 117         intptr_t allocs;
 118         uint64_t alloc_size;
 119         TraceDesc traces;
 120 };
 121
 122 static ClassDesc* class_hash [HASH_SIZE] = {0};
 123 static int num_classes = 0;
 124
 125 static ClassDesc*
 126 add_class (intptr_t klass, const char *name)
 127 {
 128         int slot = ((klass >> 2) & 0xffff) % HASH_SIZE;
 129         ClassDesc *cd;
 130         cd = class_hash [slot];
 131         while (cd && cd->klass != klass)
 132                 cd = cd->next;
 133         /* we resolved an unknown class (unless we had the code unloaded) */
 134         if (cd) {
 135                 /*printf ("resolved unknown: %s\n", name);*/
 136                 free (cd->name);
 137                 cd->name = pstrdup (name);
 138                 return cd;
 139         }
 140         cd = calloc (sizeof (ClassDesc), 1);
 141         cd->klass = klass;
 142         cd->name = pstrdup (name);
 143         cd->next = class_hash [slot];
 144         cd->allocs = 0;
 145         cd->alloc_size = 0;
 146         cd->traces.count = 0;
 147         cd->traces.size = 0;
 148         cd->traces.traces = NULL;
 149         class_hash [slot] = cd;
 150         num_classes++;
 151         return cd;
 152 }
 153
 154 static ClassDesc *
 155 lookup_class (intptr_t klass)
 156 {
 157         int slot = ((klass >> 2) & 0xffff) % HASH_SIZE;
 158         ClassDesc *cd = class_hash [slot];
 159         while (cd && cd->klass != klass)
 160                 cd = cd->next;
 161         if (!cd)
 162                 return add_class (klass, "unresolved class");
 163         return cd;
 164 }
 165
 166 typedef struct _MethodDesc MethodDesc;
 167 struct _MethodDesc {
 168         MethodDesc *next;
 169         intptr_t method;
 170         char *name;
 171         intptr_t code;
 172         int len;
 173         int recurse_count;
 174         uint64_t calls;
 175         uint64_t total_time;
 176         uint64_t callee_time;
 177         uint64_t self_time;
 178         TraceDesc traces;
 179 };
 180
 181 static MethodDesc* method_hash [HASH_SIZE] = {0};
 182 static int num_methods = 0;
 183
 184 static MethodDesc*
 185 add_method (intptr_t method, const char *name, intptr_t code, int len)
 186 {
 187         int slot = ((method >> 2) & 0xffff) % HASH_SIZE;
 188         MethodDesc *cd;
 189         cd = method_hash [slot];
 190         while (cd && cd->method != method)
 191                 cd = cd->next;
 192         /* we resolved an unknown method (unless we had the code unloaded) */
 193         if (cd) {
 194                 cd->code = code;
 195                 cd->len = len;
 196                 /*printf ("resolved unknown: %s\n", name);*/
 197                 free (cd->name);
 198                 cd->name = pstrdup (name);
 199                 return cd;
 200         }
 201         cd = calloc (sizeof (MethodDesc), 1);
 202         cd->method = method;
 203         cd->name = pstrdup (name);
 204         cd->code = code;
 205         cd->len = len;
 206         cd->calls = 0;
 207         cd->total_time = 0;
 208         cd->traces.count = 0;
 209         cd->traces.size = 0;
 210         cd->traces.traces = NULL;
 211         cd->next = method_hash [slot];
 212         method_hash [slot] = cd;
 213         num_methods++;
 214         return cd;
 215 }
 216
 217 static MethodDesc *
 218 lookup_method (intptr_t method)
 219 {
 220         int slot = ((method >> 2) & 0xffff) % HASH_SIZE;
 221         MethodDesc *cd = method_hash [slot];
 222         while (cd && cd->method != method)
 223                 cd = cd->next;
 224         if (!cd)
 225                 return add_method (method, "unknown method", 0, 0);
 226         return cd;
 227 }
 228
 229 typedef struct _HeapClassDesc HeapClassDesc;
 230 typedef struct {
 231         HeapClassDesc *klass;
 232         uint64_t count;
 233 } HeapClassRevRef;
 234
 235 struct _HeapClassDesc {
 236         ClassDesc *klass;
 237         int64_t count;
 238         int64_t total_size;
 239         HeapClassRevRef *rev_hash;
 240         int rev_hash_size;
 241         int rev_count;
 242 };
 243
 244 static int
 245 add_rev_class_hashed (HeapClassRevRef *rev_hash, uintptr_t size, HeapClassDesc *hklass, uint64_t value)
 246 {
 247         uintptr_t i;
 248         uintptr_t start_pos;
 249         start_pos = (hklass->klass->klass >> 2) % size;
 250         assert (start_pos < size);
 251         i = start_pos;
 252         do {
 253                 if (rev_hash [i].klass == hklass) {
 254                         rev_hash [i].count += value;
 255                         return 0;
 256                 } else if (!rev_hash [i].klass) {
 257                         rev_hash [i].klass = hklass;
 258                         rev_hash [i].count += value;
 259                         start_pos = 0;
 260                         for (i = 0; i < size; ++i)
 261                                 if (rev_hash [i].klass && rev_hash [i].klass->klass == hklass->klass)
 262                                         start_pos ++;
 263                         assert (start_pos == 1);
 264                         return 1;
 265                 }
 266                 /* wrap around */
 267                 if (++i == size)
 268                         i = 0;
 269         } while (i != start_pos);
 270         /* should not happen */
 271         printf ("failed revref store\n");
 272         return 0;
 273 }
 274
 275 static void
 276 add_heap_class_rev (HeapClassDesc *from, HeapClassDesc *to)
 277 {
 278         uintptr_t i;
 279         if (to->rev_count * 2 >= to->rev_hash_size) {
 280                 HeapClassRevRef *n;
 281                 uintptr_t old_size = to->rev_hash_size;
 282                 to->rev_hash_size *= 2;
 283                 if (to->rev_hash_size == 0)
 284                         to->rev_hash_size = 4;
 285                 n = calloc (sizeof (HeapClassRevRef) * to->rev_hash_size, 1);
 286                 for (i = 0; i < old_size; ++i) {
 287                         if (to->rev_hash [i].klass)
 288                                 add_rev_class_hashed (n, to->rev_hash_size, to->rev_hash [i].klass, to->rev_hash [i].count);
 289                 }
 290                 if (to->rev_hash)
 291                         free (to->rev_hash);
 292                 to->rev_hash = n;
 293         }
 294         to->rev_count += add_rev_class_hashed (to->rev_hash, to->rev_hash_size, from, 1);
 295 }
 296
 297 typedef struct {
 298         uintptr_t objaddr;
 299         HeapClassDesc *hklass;
 300         uintptr_t num_refs;
 301         uintptr_t refs [0];
 302 } HeapObjectDesc;
 303
 304 typedef struct _HeapShot HeapShot;
 305 struct _HeapShot {
 306         HeapShot *next;
 307         uint64_t timestamp;
 308         int class_count;
 309         int hash_size;
 310         HeapClassDesc **class_hash;
 311         HeapClassDesc **sorted;
 312         HeapObjectDesc **objects_hash;
 313         uintptr_t objects_count;
 314         uintptr_t objects_hash_size;
 315 };
 316
 317 static HeapShot *heap_shots = NULL;
 318 static int num_heap_shots = 0;
 319
 320 static HeapShot*
 321 new_heap_shot (uint64_t timestamp)
 322 {
 323         HeapShot *hs = calloc (sizeof (HeapShot), 1);
 324         hs->hash_size = 4;
 325         hs->class_hash = calloc (sizeof (void*), hs->hash_size);
 326         hs->timestamp = timestamp;
 327         num_heap_shots++;
 328         hs->next = heap_shots;
 329         heap_shots = hs;
 330         return hs;
 331 }
 332
 333 static HeapClassDesc*
 334 heap_class_lookup (HeapShot *hs, ClassDesc *klass)
 335 {
 336         int i;
 337         unsigned int start_pos;
 338         start_pos = ((uintptr_t)klass->klass >> 2) % hs->hash_size;
 339         i = start_pos;
 340         do {
 341                 HeapClassDesc* cd = hs->class_hash [i];
 342                 if (!cd)
 343                         return NULL;
 344                 if (cd->klass == klass)
 345                         return cd;
 346                 /* wrap around */
 347                 if (++i == hs->hash_size)
 348                         i = 0;
 349         } while (i != start_pos);
 350         return NULL;
 351 }
 352
 353 static int
 354 add_heap_hashed (HeapClassDesc **hash, HeapClassDesc **retv, uintptr_t hsize, ClassDesc *klass, uint64_t size, uint64_t count)
 355 {
 356         uintptr_t i;
 357         uintptr_t start_pos;
 358         start_pos = ((uintptr_t)klass->klass >> 2) % hsize;
 359         i = start_pos;
 360         do {
 361                 if (hash [i] && hash [i]->klass == klass) {
 362                         hash [i]->total_size += size;
 363                         hash [i]->count += count;
 364                         *retv = hash [i];
 365                         return 0;
 366                 } else if (!hash [i]) {
 367                         if (*retv) {
 368                                 hash [i] = *retv;
 369                                 return 1;
 370                         }
 371                         hash [i] = calloc (sizeof (HeapClassDesc), 1);
 372                         hash [i]->klass = klass;
 373                         hash [i]->total_size += size;
 374                         hash [i]->count += count;
 375                         *retv = hash [i];
 376                         return 1;
 377                 }
 378                 /* wrap around */
 379                 if (++i == hsize)
 380                         i = 0;
 381         } while (i != start_pos);
 382         /* should not happen */
 383         printf ("failed heap class store\n");
 384         return 0;
 385 }
 386
 387 static HeapClassDesc*
 388 add_heap_shot_class (HeapShot *hs, ClassDesc *klass, uint64_t size)
 389 {
 390         HeapClassDesc *res;
 391         int i;
 392         if (hs->class_count * 2 >= hs->hash_size) {
 393                 HeapClassDesc **n;
 394                 int old_size = hs->hash_size;
 395                 hs->hash_size *= 2;
 396                 if (hs->hash_size == 0)
 397                         hs->hash_size = 4;
 398                 n = calloc (sizeof (void*) * hs->hash_size, 1);
 399                 for (i = 0; i < old_size; ++i) {
 400                         res = hs->class_hash [i];
 401                         if (hs->class_hash [i])
 402                                 add_heap_hashed (n, &res, hs->hash_size, hs->class_hash [i]->klass, hs->class_hash [i]->total_size, hs->class_hash [i]->count);
 403                 }
 404                 if (hs->class_hash)
 405                         free (hs->class_hash);
 406                 hs->class_hash = n;
 407         }
 408         res = NULL;
 409         hs->class_count += add_heap_hashed (hs->class_hash, &res, hs->hash_size, klass, size, 1);
 410         //if (res->count == 1)
 411         //      printf ("added heap class: %s\n", res->klass->name);
 412         return res;
 413 }
 414
 415 static HeapObjectDesc*
 416 alloc_heap_obj (uintptr_t objaddr, HeapClassDesc *hklass, uintptr_t num_refs)
 417 {
 418         HeapObjectDesc* ho = calloc (sizeof (HeapObjectDesc) + num_refs * sizeof (uintptr_t), 1);
 419         ho->objaddr = objaddr;
 420         ho->hklass = hklass;
 421         ho->num_refs = num_refs;
 422         return ho;
 423 }
 424
 425 static uintptr_t
 426 heap_shot_find_obj_slot (HeapShot *hs, uintptr_t objaddr)
 427 {
 428         uintptr_t i;
 429         uintptr_t start_pos;
 430         HeapObjectDesc **hash = hs->objects_hash;
 431         start_pos = ((uintptr_t)objaddr >> 3) % hs->objects_hash_size;
 432         i = start_pos;
 433         do {
 434                 if (hash [i] && hash [i]->objaddr == objaddr) {
 435                         return i;
 436                 } else if (!hash [i]) {
 437                         break; /* fail */
 438                 }
 439                 /* wrap around */
 440                 if (++i == hs->objects_hash_size)
 441                         i = 0;
 442         } while (i != start_pos);
 443         /* should not happen */
 444         printf ("failed heap obj update\n");
 445         return -1;
 446 }
 447
 448 static HeapObjectDesc*
 449 heap_shot_obj_add_refs (HeapShot *hs, uintptr_t objaddr, uintptr_t num, uintptr_t *ref_offset)
 450 {
 451         HeapObjectDesc **hash = hs->objects_hash;
 452         uintptr_t i = heap_shot_find_obj_slot (hs, objaddr);
 453         if (i >= 0) {
 454                 HeapObjectDesc* ho = alloc_heap_obj (objaddr, hash [i]->hklass, hash [i]->num_refs + num);
 455                 *ref_offset = hash [i]->num_refs;
 456                 memcpy (ho->refs, hash [i]->refs, hash [i]->num_refs * sizeof (uintptr_t));
 457                 free (hash [i]);
 458                 hash [i] = ho;
 459                 return ho;
 460         }
 461         /* should not happen */
 462         printf ("failed heap obj update\n");
 463         return NULL;
 464
 465 }
 466
 467 static uintptr_t
 468 add_heap_hashed_obj (HeapObjectDesc **hash, uintptr_t hsize, HeapObjectDesc *obj)
 469 {
 470         uintptr_t i;
 471         uintptr_t start_pos;
 472         start_pos = ((uintptr_t)obj->objaddr >> 3) % hsize;
 473         i = start_pos;
 474         do {
 475                 if (hash [i] && hash [i]->objaddr == obj->objaddr) {
 476                         printf ("duplicate object!\n");
 477                         return 0;
 478                 } else if (!hash [i]) {
 479                         hash [i] = obj;
 480                         return 1;
 481                 }
 482                 /* wrap around */
 483                 if (++i == hsize)
 484                         i = 0;
 485         } while (i != start_pos);
 486         /* should not happen */
 487         printf ("failed heap obj store\n");
 488         return 0;
 489 }
 490
 491 static void
 492 add_heap_shot_obj (HeapShot *hs, HeapObjectDesc *obj)
 493 {
 494         uintptr_t i;
 495         if (hs->objects_count * 2 >= hs->objects_hash_size) {
 496                 HeapObjectDesc **n;
 497                 uintptr_t old_size = hs->objects_hash_size;
 498                 hs->objects_hash_size *= 2;
 499                 if (hs->objects_hash_size == 0)
 500                         hs->objects_hash_size = 4;
 501                 n = calloc (sizeof (void*) * hs->objects_hash_size, 1);
 502                 for (i = 0; i < old_size; ++i) {
 503                         if (hs->objects_hash [i])
 504                                 add_heap_hashed_obj (n, hs->objects_hash_size, hs->objects_hash [i]);
 505                 }
 506                 if (hs->objects_hash)
 507                         free (hs->objects_hash);
 508                 hs->objects_hash = n;
 509         }
 510         hs->objects_count += add_heap_hashed_obj (hs->objects_hash, hs->objects_hash_size, obj);
 511 }
 512
 513 static void
 514 heap_shot_resolve_reverse_refs (HeapShot *hs)
 515 {
 516         uintptr_t i;
 517         for (i = 0; i < hs->objects_hash_size; ++i) {
 518                 uintptr_t r;
 519                 HeapObjectDesc *ho = hs->objects_hash [i];
 520                 if (!ho)
 521                         continue;
 522                 for (r = 0; r < ho->num_refs; ++r) {
 523                         uintptr_t oi = heap_shot_find_obj_slot (hs, ho->refs [r]);
 524                         add_heap_class_rev (ho->hklass, hs->objects_hash [oi]->hklass);
 525                 }
 526         }
 527 }
 528
 529 static void
 530 heap_shot_free_objects (HeapShot *hs)
 531 {
 532         uintptr_t i;
 533         for (i = 0; i < hs->objects_hash_size; ++i) {
 534                 HeapObjectDesc *ho = hs->objects_hash [i];
 535                 if (ho)
 536                         free (ho);
 537         }
 538         if (hs->objects_hash)
 539                 free (hs->objects_hash);
 540         hs->objects_hash = NULL;
 541         hs->objects_hash_size = 0;
 542         hs->objects_count = 0;
 543 }
 544
 545 struct _BackTrace {
 546         BackTrace *next;
 547         unsigned int hash;
 548         int count;
 549         int id;
 550         MethodDesc *methods [1];
 551 };
 552
 553 static BackTrace *backtrace_hash [HASH_SIZE];
 554 static BackTrace **backtraces = NULL;
 555 static int num_backtraces = 0;
 556 static int next_backtrace = 0;
 557
 558 static int
 559 hash_backtrace (int count, MethodDesc **methods)
 560 {
 561         int hash = count;
 562         int i;
 563         for (i = 0; i < count; ++i) {
 564                 hash = (hash << 5) - hash + methods [i]->method;
 565         }
 566         return hash;
 567 }
 568
 569 static int
 570 compare_backtrace (BackTrace *bt, int count, MethodDesc **methods)
 571 {
 572         int i;
 573         if (bt->count != count)
 574                 return 0;
 575         for (i = 0; i < count; ++i)
 576                 if (methods [i] != bt->methods [i])
 577                         return 0;
 578         return 1;
 579 }
 580
 581 static BackTrace*
 582 add_backtrace (int count, MethodDesc **methods)
 583 {
 584         int hash = hash_backtrace (count, methods);
 585         int slot = (hash & 0xffff) % HASH_SIZE;
 586         BackTrace *bt = backtrace_hash [slot];
 587         while (bt) {
 588                 if (bt->hash == hash && compare_backtrace (bt, count, methods))
 589                         return bt;
 590                 bt = bt->next;
 591         }
 592         bt = malloc (sizeof (BackTrace) + ((count - 1) * sizeof (void*)));
 593         bt->next = backtrace_hash [slot];
 594         backtrace_hash [slot] = bt;
 595         if (next_backtrace == num_backtraces) {
 596                 num_backtraces *= 2;
 597                 if (!num_backtraces)
 598                         num_backtraces = 16;
 599                 backtraces = realloc (backtraces, sizeof (void*) * num_backtraces);
 600         }
 601         bt->id = next_backtrace++;
 602         backtraces [bt->id] = bt;
 603         bt->count = count;
 604         bt->hash = hash;
 605         for (slot = 0; slot < count; ++slot)
 606                 bt->methods [slot] = methods [slot];
 607
 608         return bt;
 609 }
 610
 611 typedef struct _MonitorDesc MonitorDesc;
 612 typedef struct _ThreadContext ThreadContext;
 613
 614 typedef struct {
 615         FILE *file;
 616 #if defined (HAVE_SYS_ZLIB)
 617         gzFile *gzfile;
 618 #endif
 619         unsigned char *buf;
 620         int size;
 621         int data_version;
 622         int version_major;
 623         int version_minor;
 624         int timer_overhead;
 625         int pid;
 626         uint64_t startup_time;
 627         ThreadContext *threads;
 628         ThreadContext *current;
 629 } ProfContext;
 630
 631 struct _ThreadContext {
 632         ThreadContext *next;
 633         intptr_t thread_id;
 634         char *name;
 635         /* emulated stack */
 636         MethodDesc **stack;
 637         uint64_t *time_stack;
 638         uint64_t *callee_time_stack;
 639         uint64_t last_time;
 640         uint64_t contention_start;
 641         MonitorDesc *monitor;
 642         int stack_size;
 643         int stack_id;
 644         HeapShot *current_heap_shot;
 645 };
 646
 647 static void
 648 ensure_buffer (ProfContext *ctx, int size)
 649 {
 650         if (ctx->size < size) {
 651                 ctx->buf = realloc (ctx->buf, size);
 652                 ctx->size = size;
 653         }
 654 }
 655
 656 static int
 657 load_data (ProfContext *ctx, int size)
 658 {
 659         ensure_buffer (ctx, size);
 660 #if defined (HAVE_SYS_ZLIB)
 661         if (ctx->gzfile) {
 662                 int r = gzread (ctx->gzfile, ctx->buf, size);
 663                 if (r == 0)
 664                         return size == 0? 1: 0;
 665                 return r == size;
 666         } else {
 667 #endif
 668                 int r = fread (ctx->buf, size, 1, ctx->file);
 669                 if (r == 0)
 670                         return size == 0? 1: 0;
 671                 return r;
 672         }
 673 }
 674
 675 static ThreadContext*
 676 get_thread (ProfContext *ctx, intptr_t thread_id)
 677 {
 678         ThreadContext *thread;
 679         if (ctx->current && ctx->current->thread_id == thread_id)
 680                 return ctx->current;
 681         thread = ctx->threads;
 682         while (thread) {
 683                 if (thread->thread_id == thread_id) {
 684                         return thread;
 685                 }
 686                 thread = thread->next;
 687         }
 688         thread = calloc (sizeof (ThreadContext), 1);
 689         thread->next = ctx->threads;
 690         ctx->threads = thread;
 691         thread->thread_id = thread_id;
 692         thread->last_time = 0;
 693         thread->stack_id = 0;
 694         thread->stack_size = 32;
 695         thread->stack = malloc (thread->stack_size * sizeof (void*));
 696         thread->time_stack = malloc (thread->stack_size * sizeof (uint64_t));
 697         thread->callee_time_stack = malloc (thread->stack_size * sizeof (uint64_t));
 698         return thread;
 699 }
 700
 701 static ThreadContext*
 702 load_thread (ProfContext *ctx, intptr_t thread_id)
 703 {
 704         ThreadContext *thread = get_thread (ctx, thread_id);
 705         ctx->current = thread;
 706         return thread;
 707 }
 708
 709 static void
 710 ensure_thread_stack (ThreadContext *thread)
 711 {
 712         if (thread->stack_id == thread->stack_size) {
 713                 thread->stack_size *= 2;
 714                 thread->stack = realloc (thread->stack, thread->stack_size * sizeof (void*));
 715                 thread->time_stack = realloc (thread->time_stack, thread->stack_size * sizeof (uint64_t));
 716                 thread->callee_time_stack = realloc (thread->callee_time_stack, thread->stack_size * sizeof (uint64_t));
 717         }
 718 }
 719
 720 static int
 721 add_trace_hashed (CallContext *traces, int size, BackTrace *bt, uint64_t value)
 722 {
 723         int i;
 724         unsigned int start_pos;
 725         start_pos = bt->hash % size;
 726         i = start_pos;
 727         do {
 728                 if (traces [i].bt == bt) {
 729                         traces [i].count += value;
 730                         return 0;
 731                 } else if (!traces [i].bt) {
 732                         traces [i].bt = bt;
 733                         traces [i].count += value;
 734                         return 1;
 735                 }
 736                 /* wrap around */
 737                 if (++i == size)
 738                         i = 0;
 739         } while (i != start_pos);
 740         /* should not happen */
 741         printf ("failed trace store\n");
 742         return 0;
 743 }
 744
 745 static void
 746 add_trace_bt (BackTrace *bt, TraceDesc *trace, uint64_t value)
 747 {
 748         int i;
 749         if (!collect_traces)
 750                 return;
 751         if (trace->count * 2 >= trace->size) {
 752                 CallContext *n;
 753                 int old_size = trace->size;
 754                 trace->size *= 2;
 755                 if (trace->size == 0)
 756                         trace->size = 4;
 757                 n = calloc (sizeof (CallContext) * trace->size, 1);
 758                 for (i = 0; i < old_size; ++i) {
 759                         if (trace->traces [i].bt)
 760                                 add_trace_hashed (n, trace->size, trace->traces [i].bt, trace->traces [i].count);
 761                 }
 762                 if (trace->traces)
 763                         free (trace->traces);
 764                 trace->traces = n;
 765         }
 766         trace->count += add_trace_hashed (trace->traces, trace->size, bt, value);
 767 }
 768
 769 static BackTrace*
 770 add_trace_thread (ThreadContext *thread, TraceDesc *trace, uint64_t value)
 771 {
 772         BackTrace *bt;
 773         int count = thread->stack_id;
 774         if (!collect_traces)
 775                 return NULL;
 776         if (count > trace_max)
 777                 count = trace_max;
 778         bt = add_backtrace (count, thread->stack + thread->stack_id - count);
 779         add_trace_bt (bt, trace, value);
 780         return bt;
 781 }
 782
 783 static BackTrace*
 784 add_trace_methods (MethodDesc **methods, int count, TraceDesc *trace, uint64_t value)
 785 {
 786         BackTrace *bt;
 787         if (!collect_traces)
 788                 return NULL;
 789         if (count > trace_max)
 790                 count = trace_max;
 791         bt = add_backtrace (count, methods);
 792         add_trace_bt (bt, trace, value);
 793         return bt;
 794 }
 795
 796 static int
 797 compare_callc (const void *a, const void *b)
 798 {
 799         const CallContext *A = a;
 800         const CallContext *B = b;
 801         if (B->count == A->count)
 802                 return 0;
 803         if (B->count < A->count)
 804                 return -1;
 805         return 1;
 806 }
 807
 808 static void
 809 sort_context_array (TraceDesc* traces)
 810 {
 811         int i, j;
 812         for (i = 0, j = 0; i < traces->size; ++i) {
 813                 if (traces->traces [i].bt) {
 814                         traces->traces [j].bt = traces->traces [i].bt;
 815                         traces->traces [j].count = traces->traces [i].count;
 816                         j++;
 817                 }
 818         }
 819         qsort (traces->traces, traces->count, sizeof (CallContext), compare_callc);
 820 }
 821
 822 static void
 823 push_method (ThreadContext *thread, MethodDesc *method, uint64_t timestamp)
 824 {
 825         ensure_thread_stack (thread);
 826         thread->time_stack [thread->stack_id] = timestamp;
 827         thread->callee_time_stack [thread->stack_id] = 0;
 828         thread->stack [thread->stack_id++] = method;
 829         method->recurse_count++;
 830 }
 831
 832 static void
 833 pop_method (ThreadContext *thread, MethodDesc *method, uint64_t timestamp)
 834 {
 835         method->recurse_count--;
 836         if (thread->stack_id > 0 && thread->stack [thread->stack_id - 1] == method) {
 837                 uint64_t tdiff;
 838                 thread->stack_id--;
 839                 method->calls++;
 840                 if (timestamp < thread->time_stack [thread->stack_id])
 841                         fprintf (outfile, "time went backwards for %s\n", method->name);
 842                 tdiff = timestamp - thread->time_stack [thread->stack_id];
 843                 if (thread->callee_time_stack [thread->stack_id] > tdiff)
 844                         fprintf (outfile, "callee time bigger for %s\n", method->name);
 845                 method->self_time += tdiff - thread->callee_time_stack [thread->stack_id];
 846                 method->callee_time += thread->callee_time_stack [thread->stack_id];
 847                 if (thread->stack_id)
 848                         thread->callee_time_stack [thread->stack_id - 1] += tdiff;
 849                 //fprintf (outfile, "method %s took %d\n", method->name, (int)(tdiff/1000));
 850         } else {
 851                 fprintf (outfile, "unmatched leave at stack pos: %d for method %s\n", thread->stack_id, method->name);
 852         }
 853 }
 854
 855 typedef struct {
 856         uint64_t start_time; /* move this per-thread? */
 857         uint64_t total_time;
 858         uint64_t max_time;
 859         int count;
 860 } GCDesc;
 861 static GCDesc gc_info [3];
 862 static uint64_t max_heap_size;
 863 static uint64_t gc_object_moves;
 864 static int gc_resizes;
 865 typedef struct {
 866         uint64_t created;
 867         uint64_t destroyed;
 868         uint64_t max_live;
 869         TraceDesc traces;
 870 } HandleInfo;
 871 static HandleInfo handle_info [4];
 872
 873 static const char*
 874 gc_event_name (int ev)
 875 {
 876         switch (ev) {
 877         case MONO_GC_EVENT_START: return "start";
 878         case MONO_GC_EVENT_MARK_START: return "mark start";
 879         case MONO_GC_EVENT_MARK_END: return "mark end";
 880         case MONO_GC_EVENT_RECLAIM_START: return "reclaim start";
 881         case MONO_GC_EVENT_RECLAIM_END: return "reclaim end";
 882         case MONO_GC_EVENT_END: return "end";
 883         case MONO_GC_EVENT_PRE_STOP_WORLD: return "pre stop";
 884         case MONO_GC_EVENT_POST_STOP_WORLD: return "post stop";
 885         case MONO_GC_EVENT_PRE_START_WORLD: return "pre start";
 886         case MONO_GC_EVENT_POST_START_WORLD: return "post start";
 887         default:
 888                 return "unknown";
 889         }
 890 }
 891
 892 static uint64_t clause_summary [MONO_EXCEPTION_CLAUSE_FAULT + 1];
 893 static uint64_t throw_count = 0;
 894 static TraceDesc exc_traces;
 895
 896 static const char*
 897 clause_name (int type)
 898 {
 899         switch (type) {
 900         case MONO_EXCEPTION_CLAUSE_NONE: return "catch";
 901         case MONO_EXCEPTION_CLAUSE_FILTER: return "filter";
 902         case MONO_EXCEPTION_CLAUSE_FINALLY: return "finally";
 903         case MONO_EXCEPTION_CLAUSE_FAULT: return "fault";
 904         default: return "invalid";
 905         }
 906 }
 907
 908 static uint64_t monitor_contention;
 909 static uint64_t monitor_failed;
 910 static uint64_t monitor_acquired;
 911
 912 struct _MonitorDesc {
 913         MonitorDesc *next;
 914         uintptr_t objid;
 915         uintptr_t contentions;
 916         uint64_t wait_time;
 917         uint64_t max_wait_time;
 918         TraceDesc traces;
 919 };
 920
 921 static MonitorDesc* monitor_hash [SMALL_HASH_SIZE] = {0};
 922 static int num_monitors = 0;
 923
 924 static MonitorDesc*
 925 lookup_monitor (uintptr_t objid)
 926 {
 927         int slot = ((objid >> 3) & 0xffff) % SMALL_HASH_SIZE;
 928         MonitorDesc *cd = monitor_hash [slot];
 929         while (cd && cd->objid != objid)
 930                 cd = cd->next;
 931         if (!cd) {
 932                 cd = calloc (sizeof (MonitorDesc), 1);
 933                 cd->objid = objid;
 934                 cd->next = monitor_hash [slot];
 935                 monitor_hash [slot] = cd;
 936                 num_monitors++;
 937         }
 938         return cd;
 939 }
 940
 941 static const char*
 942 monitor_ev_name (int ev)
 943 {
 944         switch (ev) {
 945         case MONO_PROFILER_MONITOR_CONTENTION: return "contended";
 946         case MONO_PROFILER_MONITOR_DONE: return "acquired";
 947         case MONO_PROFILER_MONITOR_FAIL: return "not taken";
 948         default: return "invalid";
 949         }
 950 }
 951
 952 static const char*
 953 get_handle_name (int htype)
 954 {
 955         switch (htype) {
 956         case 0: return "weak";
 957         case 1: return "weaktrack";
 958         case 2: return "normal";
 959         case 3: return "pinned";
 960         default: return "unknown";
 961         }
 962 }
 963
 964 static MethodDesc**
 965 decode_bt (MethodDesc** sframes, int *size, unsigned char *p, unsigned char **endp, intptr_t ptr_base)
 966 {
 967         MethodDesc **frames;
 968         int i;
 969         int flags = decode_uleb128 (p, &p);
 970         int count = decode_uleb128 (p, &p);
 971         if (flags != 0)
 972                 return NULL;
 973         if (count > *size)
 974                 frames = malloc (count * sizeof (void*));
 975         else
 976                 frames = sframes;
 977         for (i = 0; i < count; ++i) {
 978                 intptr_t ptrdiff = decode_sleb128 (p, &p);
 979                 frames [i] = lookup_method (ptr_base + ptrdiff);
 980         }
 981         *size = count;
 982         *endp = p;
 983         return frames;
 984 }
 985
 986 static void
 987 tracked_creation (uintptr_t obj, ClassDesc *cd, uint64_t size, BackTrace *bt, uint64_t timestamp)
 988 {
 989         int i;
 990         for (i = 0; i < num_tracked_objects; ++i) {
 991                 if (tracked_objects [i] != obj)
 992                         continue;
 993                 fprintf (outfile, "Object %p created (%s, %llu bytes) at %.3f secs.\n", (void*)obj, cd->name, size, (timestamp - startup_time)/1000000000.0);
 994                 if (bt && bt->count) {
 995                         int k;
 996                         for (k = 0; k < bt->count; ++k)
 997                                 fprintf (outfile, "\t%s\n", bt->methods [k]->name);
 998                 }
 999         }
1000 }
1001
1002 static void
1003 track_handle (uintptr_t obj, int htype, uint32_t handle)
1004 {
1005         int i;
1006         for (i = 0; i < num_tracked_objects; ++i) {
1007                 if (tracked_objects [i] == obj)
1008                         fprintf (outfile, "Object %p referenced from handle %u\n", (void*)obj, handle);
1009         }
1010 }
1011
1012 static void
1013 track_move (uintptr_t src, uintptr_t dst)
1014 {
1015         int i;
1016         for (i = 0; i < num_tracked_objects; ++i) {
1017                 if (tracked_objects [i] == src)
1018                         fprintf (outfile, "Object %p moved to %p\n", (void*)src, (void*)dst);
1019                 else if (tracked_objects [i] == dst)
1020                         fprintf (outfile, "Object %p moved from %p\n", (void*)dst, (void*)src);
1021         }
1022 }
1023
1024 static void
1025 track_obj_reference (uintptr_t obj, uintptr_t parent, ClassDesc *cd)
1026 {
1027         int i;
1028         for (i = 0; i < num_tracked_objects; ++i) {
1029                 if (tracked_objects [i] == obj)
1030                         fprintf (outfile, "Object %p referenced from %p (%s).\n", (void*)obj, (void*)parent, cd->name);
1031         }
1032 }
1033
1034 static void
1035 found_object (uintptr_t obj)
1036 {
1037         num_tracked_objects ++;
1038         tracked_objects = realloc (tracked_objects, num_tracked_objects * sizeof (tracked_objects [0]));
1039         tracked_objects [num_tracked_objects - 1] = obj;
1040 }
1041
1042 #define OBJ_ADDR(diff) ((obj_base + diff) << 3)
1043 #define LOG_TIME(base,diff) /*fprintf("outfile, time %llu + %llu near offset %d\n", base, diff, p - ctx->buf)*/
1044
1045 static int
1046 decode_buffer (ProfContext *ctx)
1047 {
1048         unsigned char *p;
1049         unsigned char *end;
1050         intptr_t thread_id;
1051         intptr_t ptr_base;
1052         intptr_t obj_base;
1053         intptr_t method_base;
1054         uint64_t time_base;
1055         uint64_t file_offset;
1056         int len, i;
1057         ThreadContext *thread;
1058
1059 #ifdef HAVE_SYS_ZLIB
1060         if (ctx->gzfile)
1061                 file_offset = gztell (ctx->gzfile);
1062         else
1063 #endif
1064                 file_offset = ftell (ctx->file);
1065         if (!load_data (ctx, 48))
1066                 return 0;
1067         p = ctx->buf;
1068         if (read_int32 (p) != BUF_ID) {
1069                 fprintf (outfile, "Incorrect buffer id: 0x%x\n", read_int32 (p));
1070                 for (i = 0; i < 48; ++i) {
1071                         fprintf (outfile, "0x%x%s", p [i], i % 8?" ":"\n");
1072                 }
1073                 return 0;
1074         }
1075         len = read_int32 (p + 4);
1076         time_base = read_int64 (p + 8);
1077         ptr_base = read_int64 (p + 16);
1078         obj_base = read_int64 (p + 24);
1079         thread_id = read_int64 (p + 32);
1080         method_base = read_int64 (p + 40);
1081         if (debug)
1082                 fprintf (outfile, "buf: thread:%x, len: %d, time: %llu, file offset: %llu\n", thread_id, len, time_base, file_offset);
1083         thread = load_thread (ctx, thread_id);
1084         if (!load_data (ctx, len))
1085                 return 0;
1086         if (!startup_time) {
1087                 startup_time = time_base;
1088                 if (time_from) {
1089                         time_from += startup_time;
1090                         time_to += startup_time;
1091                 }
1092                 if (!thread->name)
1093                         thread->name = pstrdup ("Main");
1094         }
1095         for (i = 0; i < thread->stack_id; ++i)
1096                 thread->stack [i]->recurse_count++;
1097         p = ctx->buf;
1098         end = p + len;
1099         while (p < end) {
1100                 switch (*p & 0xf) {
1101                 case TYPE_GC: {
1102                         int subtype = *p & 0xf0;
1103                         uint64_t tdiff = decode_uleb128 (p + 1, &p);
1104                         LOG_TIME (time_base, tdiff);
1105                         time_base += tdiff;
1106                         if (subtype == TYPE_GC_RESIZE) {
1107                                 uint64_t new_size = decode_uleb128 (p, &p);
1108                                 if (debug)
1109                                         fprintf (outfile, "gc heap resized to %llu\n", new_size);
1110                                 gc_resizes++;
1111                                 if (new_size > max_heap_size)
1112                                         max_heap_size = new_size;
1113                         } else if (subtype == TYPE_GC_EVENT) {
1114                                 uint64_t ev = decode_uleb128 (p, &p);
1115                                 int gen = decode_uleb128 (p, &p);
1116                                 if (debug)
1117                                         fprintf (outfile, "gc event for gen%d: %s at %llu\n", gen - 1, gc_event_name (ev), time_base);
1118                                 if (gen > 2) {
1119                                         fprintf (outfile, "incorrect gc gen: %d\n", gen);
1120                                         break;
1121                                 }
1122                                 if (ev == MONO_GC_EVENT_START) {
1123                                         gc_info [gen].start_time = time_base;
1124                                         gc_info [gen].count++;
1125                                 } else if (ev == MONO_GC_EVENT_END) {
1126                                         tdiff = time_base - gc_info [gen].start_time;
1127                                         gc_info [gen].total_time += tdiff;
1128                                         if (tdiff > gc_info [gen].max_time)
1129                                                 gc_info [gen].max_time = tdiff;
1130                                 }
1131                         } else if (subtype == TYPE_GC_MOVE) {
1132                                 int j, num = decode_uleb128 (p, &p);
1133                                 gc_object_moves += num / 2;
1134                                 for (j = 0; j < num; j += 2) {
1135                                         intptr_t obj1diff = decode_sleb128 (p, &p);
1136                                         intptr_t obj2diff = decode_sleb128 (p, &p);
1137                                         if (num_tracked_objects)
1138                                                 track_move (OBJ_ADDR (obj1diff), OBJ_ADDR (obj2diff));
1139                                         if (debug) {
1140                                                 fprintf (outfile, "moved obj %p to %p\n", (void*)OBJ_ADDR (obj1diff), (void*)OBJ_ADDR (obj2diff));
1141                                         }
1142                                 }
1143                         } else if (subtype == TYPE_GC_HANDLE_CREATED) {
1144                                 int htype = decode_uleb128 (p, &p);
1145                                 uint32_t handle = decode_uleb128 (p, &p);
1146                                 intptr_t objdiff = decode_sleb128 (p, &p);
1147                                 if (htype > 3)
1148                                         return 0;
1149                                 handle_info [htype].created++;
1150                                 add_trace_thread (thread, &handle_info [htype].traces, 1);
1151                                 /* FIXME: we don't take into account timing here */
1152                                 if (handle_info [htype].created > handle_info [htype].max_live)
1153                                         handle_info [htype].max_live = handle_info [htype].created;
1154                                 if (num_tracked_objects)
1155                                         track_handle (OBJ_ADDR (objdiff), htype, handle);
1156                                 if (debug)
1157                                         fprintf (outfile, "handle (%s) %u created for object %p\n", get_handle_name (htype), handle, (void*)OBJ_ADDR (objdiff));
1158                         } else if (subtype == TYPE_GC_HANDLE_DESTROYED) {
1159                                 int htype = decode_uleb128 (p, &p);
1160                                 uint32_t handle = decode_uleb128 (p, &p);
1161                                 if (htype > 3)
1162                                         return 0;
1163                                 handle_info [htype].created--;
1164                                 if (debug)
1165                                         fprintf (outfile, "handle (%s) %u destroyed\n", get_handle_name (htype), handle);
1166                         }
1167                         break;
1168                 }
1169                 case TYPE_METADATA: {
1170                         int error = *p & TYPE_LOAD_ERR;
1171                         uint64_t tdiff = decode_uleb128 (p + 1, &p);
1172                         int mtype = *p++;
1173                         intptr_t ptrdiff = decode_sleb128 (p, &p);
1174                         LOG_TIME (time_base, tdiff);
1175                         time_base += tdiff;
1176                         if (mtype == TYPE_CLASS) {
1177                                 intptr_t imptrdiff = decode_sleb128 (p, &p);
1178                                 uint64_t flags = decode_uleb128 (p, &p);
1179                                 if (flags) {
1180                                         fprintf (outfile, "non-zero flags in class\n");
1181                                         return 0;
1182                                 }
1183                                 if (debug)
1184                                         fprintf (outfile, "loaded class %p (%s in %p) at %llu\n", (void*)(ptr_base + ptrdiff), p, (void*)(ptr_base + imptrdiff), time_base);
1185                                 if (!error)
1186                                         add_class (ptr_base + ptrdiff, (char*)p);
1187                                 while (*p) p++;
1188                                 p++;
1189                         } else if (mtype == TYPE_IMAGE) {
1190                                 uint64_t flags = decode_uleb128 (p, &p);
1191                                 if (flags) {
1192                                         fprintf (outfile, "non-zero flags in image\n");
1193                                         return 0;
1194                                 }
1195                                 if (debug)
1196                                         fprintf (outfile, "loaded image %p (%s) at %llu\n", (void*)(ptr_base + ptrdiff), p, time_base);
1197                                 if (!error)
1198                                         add_image (ptr_base + ptrdiff, (char*)p);
1199                                 while (*p) p++;
1200                                 p++;
1201                         } else if (mtype == TYPE_THREAD) {
1202                                 ThreadContext *nt;
1203                                 uint64_t flags = decode_uleb128 (p, &p);
1204                                 if (flags) {
1205                                         fprintf (outfile, "non-zero flags in thread\n");
1206                                         return 0;
1207                                 }
1208                                 nt = get_thread (ctx, ptr_base * ptrdiff);
1209                                 nt->name = pstrdup ((char*)p);
1210                                 if (debug)
1211                                         fprintf (outfile, "thread %p named: %s\n", (void*)(ptr_base + ptrdiff), p);
1212                                 while (*p) p++;
1213                                 p++;
1214                         }
1215                         break;
1216                 }
1217                 case TYPE_ALLOC: {
1218                         int has_bt = *p & TYPE_ALLOC_BT;
1219                         uint64_t tdiff = decode_uleb128 (p + 1, &p);
1220                         intptr_t ptrdiff = decode_sleb128 (p, &p);
1221                         intptr_t objdiff = decode_sleb128 (p, &p);
1222                         uint64_t len;
1223                         int num_bt = 0;
1224                         MethodDesc* sframes [8];
1225                         MethodDesc** frames = sframes;
1226                         ClassDesc *cd = lookup_class (ptr_base + ptrdiff);
1227                         len = decode_uleb128 (p, &p);
1228                         LOG_TIME (time_base, tdiff);
1229                         time_base += tdiff;
1230                         if (debug)
1231                                 fprintf (outfile, "alloced object %p, size %llu (%s) at %llu\n", (void*)OBJ_ADDR (objdiff), len, lookup_class (ptr_base + ptrdiff)->name, time_base);
1232                         if (has_bt) {
1233                                 num_bt = 8;
1234                                 frames = decode_bt (sframes, &num_bt, p, &p, ptr_base);
1235                                 if (!frames) {
1236                                         fprintf (outfile, "Cannot load backtrace\n");
1237                                         return 0;
1238                                 }
1239                         }
1240                         if ((thread_filter && thread_filter == thread->thread_id) || (time_base >= time_from && time_base < time_to)) {
1241                                 BackTrace *bt;
1242                                 cd->allocs++;
1243                                 cd->alloc_size += len;
1244                                 if (has_bt)
1245                                         bt = add_trace_methods (frames, num_bt, &cd->traces, len);
1246                                 else
1247                                         bt = add_trace_thread (thread, &cd->traces, len);
1248                                 if (find_size && len >= find_size) {
1249                                         if (!find_name || strstr (cd->name, find_name))
1250                                                 found_object (OBJ_ADDR (objdiff));
1251                                 } else if (!find_size && find_name && strstr (cd->name, find_name)) {
1252                                         found_object (OBJ_ADDR (objdiff));
1253                                 }
1254                                 if (num_tracked_objects)
1255                                         tracked_creation (OBJ_ADDR (objdiff), cd, len, bt, time_base);
1256                         }
1257                         if (frames != sframes)
1258                                 free (frames);
1259                         break;
1260                 }
1261                 case TYPE_METHOD: {
1262                         int subtype = *p & 0xf0;
1263                         uint64_t tdiff = decode_uleb128 (p + 1, &p);
1264                         int64_t ptrdiff = decode_sleb128 (p, &p);
1265                         LOG_TIME (time_base, tdiff);
1266                         time_base += tdiff;
1267                         method_base += ptrdiff;
1268                         if (subtype == TYPE_JIT) {
1269                                 intptr_t codediff = decode_sleb128 (p, &p);
1270                                 int codelen = decode_uleb128 (p, &p);
1271                                 if (debug)
1272                                         fprintf (outfile, "jitted method %p (%s), size: %d\n", (void*)(method_base), p, codelen);
1273                                 add_method (method_base, (char*)p, ptr_base + codediff, codelen);
1274                                 while (*p) p++;
1275                                 p++;
1276                         } else {
1277                                 MethodDesc *method;
1278                                 if ((thread_filter && thread_filter != thread->thread_id))
1279                                         break;
1280                                 method = lookup_method (method_base);
1281                                 if (subtype == TYPE_ENTER) {
1282                                         add_trace_thread (thread, &method->traces, 1);
1283                                         push_method (thread, method, time_base);
1284                                 } else {
1285                                         pop_method (thread, method, time_base);
1286                                 }
1287                                 if (debug)
1288                                         fprintf (outfile, "%s method %s\n", subtype == TYPE_ENTER? "enter": subtype == TYPE_EXC_LEAVE? "exleave": "leave", method->name);
1289                         }
1290                         break;
1291                 }
1292                 case TYPE_HEAP: {
1293                         int subtype = *p & 0xf0;
1294                         if (subtype == TYPE_HEAP_OBJECT) {
1295                                 HeapObjectDesc *ho;
1296                                 int i;
1297                                 intptr_t objdiff = decode_sleb128 (p + 1, &p);
1298                                 intptr_t ptrdiff = decode_sleb128 (p, &p);
1299                                 uint64_t size = decode_uleb128 (p, &p);
1300                                 uintptr_t num = decode_uleb128 (p, &p);
1301                                 uintptr_t ref_offset;
1302                                 uintptr_t last_obj_offset = 0;
1303                                 ClassDesc *cd = lookup_class (ptr_base + ptrdiff);
1304                                 if (size) {
1305                                         HeapClassDesc *hcd = add_heap_shot_class (thread->current_heap_shot, cd, size);
1306                                         if (collect_traces) {
1307                                                 ho = alloc_heap_obj (OBJ_ADDR (objdiff), hcd, num);
1308                                                 add_heap_shot_obj (thread->current_heap_shot, ho);
1309                                                 ref_offset = 0;
1310                                         }
1311                                 } else {
1312                                         if (collect_traces)
1313                                                 ho = heap_shot_obj_add_refs (thread->current_heap_shot, OBJ_ADDR (objdiff), num, &ref_offset);
1314                                 }
1315                                 for (i = 0; i < num; ++i) {
1316                                         /* FIXME: use object distance to measure how good
1317                                          * the GC is at keeping related objects close
1318                                          */
1319                                         uintptr_t offset = ctx->data_version > 1? last_obj_offset + decode_uleb128 (p, &p): -1;
1320                                         intptr_t obj1diff = decode_sleb128 (p, &p);
1321                                         last_obj_offset = offset;
1322                                         if (collect_traces)
1323                                                 ho->refs [ref_offset + i] = OBJ_ADDR (obj1diff);
1324                                         if (num_tracked_objects)
1325                                                 track_obj_reference (OBJ_ADDR (obj1diff), OBJ_ADDR (objdiff), cd);
1326                                 }
1327                                 if (debug && size)
1328                                         fprintf (outfile, "traced object %p, size %llu (%s), refs: %d\n", (void*)OBJ_ADDR (objdiff), size, cd->name, num);
1329                         } else if (subtype == TYPE_HEAP_END) {
1330                                 uint64_t tdiff = decode_uleb128 (p + 1, &p);
1331                                 LOG_TIME (time_base, tdiff);
1332                                 time_base += tdiff;
1333                                 if (debug)
1334                                         fprintf (outfile, "heap shot end\n");
1335                                 if (collect_traces) {
1336                                         heap_shot_resolve_reverse_refs (thread->current_heap_shot);
1337                                         heap_shot_free_objects (thread->current_heap_shot);
1338                                 }
1339                                 thread->current_heap_shot = NULL;
1340                         } else if (subtype == TYPE_HEAP_START) {
1341                                 uint64_t tdiff = decode_uleb128 (p + 1, &p);
1342                                 LOG_TIME (time_base, tdiff);
1343                                 time_base += tdiff;
1344                                 if (debug)
1345                                         fprintf (outfile, "heap shot start\n");
1346                                 thread->current_heap_shot = new_heap_shot (time_base);
1347                         }
1348                         break;
1349                 }
1350                 case TYPE_MONITOR: {
1351                         int event = (*p >> 4) & 0x3;
1352                         int has_bt = *p & TYPE_MONITOR_BT;
1353                         uint64_t tdiff = decode_uleb128 (p + 1, &p);
1354                         intptr_t objdiff = decode_sleb128 (p, &p);
1355                         MethodDesc* sframes [8];
1356                         MethodDesc** frames = sframes;
1357                         int record;
1358                         int num_bt = 0;
1359                         LOG_TIME (time_base, tdiff);
1360                         time_base += tdiff;
1361                         record = (!thread_filter || thread_filter == thread->thread_id);
1362                         if (event == MONO_PROFILER_MONITOR_CONTENTION) {
1363                                 MonitorDesc *mdesc = lookup_monitor (OBJ_ADDR (objdiff));
1364                                 if (record) {
1365                                         monitor_contention++;
1366                                         mdesc->contentions++;
1367                                         thread->monitor = mdesc;
1368                                         thread->contention_start = time_base;
1369                                 }
1370                                 if (has_bt) {
1371                                         num_bt = 8;
1372                                         frames = decode_bt (sframes, &num_bt, p, &p, ptr_base);
1373                                         if (!frames) {
1374                                                 fprintf (outfile, "Cannot load backtrace\n");
1375                                                 return 0;
1376                                         }
1377                                         if (record)
1378                                                 add_trace_methods (frames, num_bt, &mdesc->traces, 1);
1379                                 } else {
1380                                         if (record)
1381                                                 add_trace_thread (thread, &mdesc->traces, 1);
1382                                 }
1383                         } else if (event == MONO_PROFILER_MONITOR_FAIL) {
1384                                 if (record) {
1385                                         monitor_failed++;
1386                                         if (thread->monitor && thread->contention_start) {
1387                                                 uint64_t wait_time = time_base - thread->contention_start;
1388                                                 if (wait_time > thread->monitor->max_wait_time)
1389                                                         thread->monitor->max_wait_time = wait_time;
1390                                                 thread->monitor->wait_time += wait_time;
1391                                                 thread->monitor = NULL;
1392                                                 thread->contention_start = 0;
1393                                         }
1394                                 }
1395                         } else if (event == MONO_PROFILER_MONITOR_DONE) {
1396                                 if (record) {
1397                                         monitor_acquired++;
1398                                         if (thread->monitor && thread->contention_start) {
1399                                                 uint64_t wait_time = time_base - thread->contention_start;
1400                                                 if (wait_time > thread->monitor->max_wait_time)
1401                                                         thread->monitor->max_wait_time = wait_time;
1402                                                 thread->monitor->wait_time += wait_time;
1403                                                 thread->monitor = NULL;
1404                                                 thread->contention_start = 0;
1405                                         }
1406                                 }
1407                         }
1408                         if (debug)
1409                                 fprintf (outfile, "monitor %s for object %p\n", monitor_ev_name (event), (void*)OBJ_ADDR (objdiff));
1410                         if (frames != sframes)
1411                                 free (frames);
1412                         break;
1413                 }
1414                 case TYPE_EXCEPTION: {
1415                         int subtype = *p & 0x70;
1416                         int has_bt = *p & TYPE_EXCEPTION_BT;
1417                         uint64_t tdiff = decode_uleb128 (p + 1, &p);
1418                         MethodDesc* sframes [8];
1419                         MethodDesc** frames = sframes;
1420                         int record;
1421                         LOG_TIME (time_base, tdiff);
1422                         time_base += tdiff;
1423                         record = (!thread_filter || thread_filter == thread->thread_id);
1424                         if (subtype == TYPE_CLAUSE) {
1425                                 int clause_type = decode_uleb128 (p, &p);
1426                                 int clause_num = decode_uleb128 (p, &p);
1427                                 int64_t ptrdiff = decode_sleb128 (p, &p);
1428                                 method_base += ptrdiff;
1429                                 if (record)
1430                                         clause_summary [clause_type]++;
1431                                 if (debug)
1432                                         fprintf (outfile, "clause %s (%d) in method %s\n", clause_name (clause_type), clause_num, lookup_method (method_base)->name);
1433                         } else {
1434                                 intptr_t objdiff = decode_sleb128 (p, &p);
1435                                 if (record)
1436                                         throw_count++;
1437                                 if (has_bt) {
1438                                         has_bt = 8;
1439                                         frames = decode_bt (sframes, &has_bt, p, &p, ptr_base);
1440                                         if (!frames) {
1441                                                 fprintf (outfile, "Cannot load backtrace\n");
1442                                                 return 0;
1443                                         }
1444                                         if (record)
1445                                                 add_trace_methods (frames, has_bt, &exc_traces, 1);
1446                                 } else {
1447                                         if (record)
1448                                                 add_trace_thread (thread, &exc_traces, 1);
1449                                 }
1450                                 if (frames != sframes)
1451                                         free (frames);
1452                                 if (debug)
1453                                         fprintf (outfile, "throw %p\n", (void*)OBJ_ADDR (objdiff));
1454                         }
1455                         break;
1456                 }
1457                 default:
1458                         fprintf (outfile, "unhandled profiler event: 0x%x at file offset: %llu + %d (len: %d\n)\n", *p, file_offset, p - ctx->buf, len);
1459                         exit (1);
1460                 }
1461         }
1462         thread->last_time = time_base;
1463         for (i = 0; i < thread->stack_id; ++i)
1464                 thread->stack [i]->recurse_count = 0;
1465         return 1;
1466 }
1467
1468 static ProfContext*
1469 load_file (char *name)
1470 {
1471         unsigned char *p;
1472         ProfContext *ctx = calloc (sizeof (ProfContext), 1);
1473         if (strcmp (name, "-") == 0)
1474                 ctx->file = stdin;
1475         else
1476                 ctx->file = fopen (name, "rb");
1477         if (!ctx->file) {
1478                 printf ("Cannot open file: %s\n", name);
1479                 exit (1);
1480         }
1481 #if defined (HAVE_SYS_ZLIB)
1482         if (ctx->file != stdin)
1483                 ctx->gzfile = gzdopen (fileno (ctx->file), "rb");
1484 #endif
1485         if (!load_data (ctx, 32))
1486                 return NULL;
1487         p = ctx->buf;
1488         if (read_int32 (p) != LOG_HEADER_ID || p [6] > LOG_DATA_VERSION)
1489                 return NULL;
1490         ctx->version_major = p [4];
1491         ctx->version_minor = p [5];
1492         ctx->data_version = p [6];
1493         /* reading 64 bit files on 32 bit systems not supported yet */
1494         if (p [7] > sizeof (void*))
1495                 return NULL;
1496         if (read_int32 (p + 20)) /* flags must be 0 */
1497                 return NULL;
1498         ctx->startup_time = read_int64 (p + 8);
1499         ctx->timer_overhead = read_int32 (p + 16);
1500         ctx->pid = read_int32 (p + 24);
1501         return ctx;
1502 }
1503
1504 enum {
1505         ALLOC_SORT_BYTES,
1506         ALLOC_SORT_COUNT
1507 };
1508 static int alloc_sort_mode = ALLOC_SORT_BYTES;
1509
1510 static int
1511 compare_class (const void *a, const void *b)
1512 {
1513         ClassDesc *const*A = a;
1514         ClassDesc *const*B = b;
1515         uint64_t vala, valb;
1516         if (alloc_sort_mode == ALLOC_SORT_BYTES) {
1517                 vala = (*A)->alloc_size;
1518                 valb = (*B)->alloc_size;
1519         } else {
1520                 vala = (*A)->allocs;
1521                 valb = (*B)->allocs;
1522         }
1523         if (valb == vala)
1524                 return 0;
1525         if (valb < vala)
1526                 return -1;
1527         return 1;
1528 }
1529
1530 static void
1531 dump_header (ProfContext *ctx)
1532 {
1533         time_t st = ctx->startup_time / 1000;
1534         char *t = ctime (&st);
1535         fprintf (outfile, "\nMono log profiler data\n");
1536         fprintf (outfile, "\tProfiler version: %d.%d\n", ctx->version_major, ctx->version_minor);
1537         fprintf (outfile, "\tData version: %d\n", ctx->data_version);
1538         fprintf (outfile, "\tMean timer overhead: %d nanoseconds\n", ctx->timer_overhead);
1539         fprintf (outfile, "\tProgram startup: %s", t);
1540         if (ctx->pid)
1541                 fprintf (outfile, "\tProgram ID: %d\n", ctx->pid);
1542 }
1543
1544 static void
1545 dump_traces (TraceDesc *traces, const char *desc)
1546 {
1547         int j;
1548         if (!show_traces)
1549                 return;
1550         if (!traces->count)
1551                 return;
1552         sort_context_array (traces);
1553         for (j = 0; j < traces->count; ++j) {
1554                 int k;
1555                 BackTrace *bt;
1556                 bt = traces->traces [j].bt;
1557                 if (!bt->count)
1558                         continue;
1559                 fprintf (outfile, "\t%llu %s from:\n", traces->traces [j].count, desc);
1560                 for (k = 0; k < bt->count; ++k)
1561                         fprintf (outfile, "\t\t%s\n", bt->methods [k]->name);
1562         }
1563 }
1564
1565 static void
1566 dump_threads (ProfContext *ctx)
1567 {
1568         ThreadContext *thread;
1569         fprintf (outfile, "\nThread summary\n");
1570         for (thread = ctx->threads; thread; thread = thread->next) {
1571                 fprintf (outfile, "\tThread: %p, name: \"%s\"\n", (void*)thread->thread_id, thread->name? thread->name: "");
1572         }
1573 }
1574
1575 static void
1576 dump_exceptions (void)
1577 {
1578         int i;
1579         fprintf (outfile, "\nException summary\n");
1580         fprintf (outfile, "\tThrows: %llu\n", throw_count);
1581         dump_traces (&exc_traces, "throws");
1582         for (i = 0; i <= MONO_EXCEPTION_CLAUSE_FAULT; ++i) {
1583                 if (!clause_summary [i])
1584                         continue;
1585                 fprintf (outfile, "\tExecuted %s clauses: %llu\n", clause_name (i), clause_summary [i]);
1586         }
1587 }
1588
1589 static int
1590 compare_monitor (const void *a, const void *b)
1591 {
1592         MonitorDesc *const*A = a;
1593         MonitorDesc *const*B = b;
1594         if ((*B)->wait_time == (*A)->wait_time)
1595                 return 0;
1596         if ((*B)->wait_time < (*A)->wait_time)
1597                 return -1;
1598         return 1;
1599 }
1600
1601 static void
1602 dump_monitors (void)
1603 {
1604         MonitorDesc **monitors;
1605         int i, j;
1606         if (!num_monitors)
1607                 return;
1608         monitors = malloc (sizeof (void*) * num_monitors);
1609         for (i = 0, j = 0; i < SMALL_HASH_SIZE; ++i) {
1610                 MonitorDesc *mdesc = monitor_hash [i];
1611                 while (mdesc) {
1612                         monitors [j++] = mdesc;
1613                         mdesc = mdesc->next;
1614                 }
1615         }
1616         qsort (monitors, num_monitors, sizeof (void*), compare_monitor);
1617         fprintf (outfile, "\nMonitor lock summary\n");
1618         for (i = 0; i < num_monitors; ++i) {
1619                 MonitorDesc *mdesc = monitors [i];
1620                 fprintf (outfile, "\tLock object %p: %d contentions\n", (void*)mdesc->objid, (int)mdesc->contentions);
1621                 fprintf (outfile, "\t\t%.6f secs total wait time, %.6f max, %.6f average\n",
1622                         mdesc->wait_time/1000000000.0, mdesc->max_wait_time/1000000000.0, mdesc->wait_time/1000000000.0/mdesc->contentions);
1623                 dump_traces (&mdesc->traces, "contentions");
1624         }
1625         fprintf (outfile, "\tLock contentions: %llu\n", monitor_contention);
1626         fprintf (outfile, "\tLock acquired: %llu\n", monitor_acquired);
1627         fprintf (outfile, "\tLock failures: %llu\n", monitor_failed);
1628 }
1629
1630 static void
1631 dump_gcs (void)
1632 {
1633         int i;
1634         fprintf (outfile, "\nGC summary\n");
1635         fprintf (outfile, "\tGC resizes: %d\n", gc_resizes);
1636         fprintf (outfile, "\tMax heap size: %llu\n", max_heap_size);
1637         fprintf (outfile, "\tObject moves: %llu\n", gc_object_moves);
1638         for (i = 0; i < 3; ++i) {
1639                 if (!gc_info [i].count)
1640                         continue;
1641                 fprintf (outfile, "\tGen%d collections: %d, max time: %lluus, total time: %lluus, average: %lluus\n",
1642                         i, gc_info [i].count, gc_info [i].max_time / 1000, gc_info [i].total_time / 1000,
1643                         gc_info [i].total_time / gc_info [i].count / 1000);
1644         }
1645         for (i = 0; i < 3; ++i) {
1646                 if (!handle_info [i].max_live)
1647                         continue;
1648                 fprintf (outfile, "\tGC handles %s: created: %llu, destroyed: %llu, max: %llu\n",
1649                         get_handle_name (i), handle_info [i].created, handle_info [i].destroyed, handle_info [i].max_live);
1650                 dump_traces (&handle_info [i].traces, "created");
1651         }
1652 }
1653
1654 static void
1655 dump_allocations (void)
1656 {
1657         int i, c;
1658         intptr_t allocs = 0;
1659         uint64_t size = 0;
1660         int header_done = 0;
1661         ClassDesc **classes = malloc (num_classes * sizeof (void*));
1662         ClassDesc *cd;
1663         c = 0;
1664         for (i = 0; i < HASH_SIZE; ++i) {
1665                 cd = class_hash [i];
1666                 while (cd) {
1667                         classes [c++] = cd;
1668                         cd = cd->next;
1669                 }
1670         }
1671         qsort (classes, num_classes, sizeof (void*), compare_class);
1672         for (i = 0; i < num_classes; ++i) {
1673                 cd = classes [i];
1674                 if (!cd->allocs)
1675                         continue;
1676                 allocs += cd->allocs;
1677                 size += cd->alloc_size;
1678                 if (!header_done++) {
1679                         fprintf (outfile, "\nAllocation summary\n");
1680                         fprintf (outfile, "%10s %10s %8s Type name\n", "Bytes", "Count", "Average");
1681                 }
1682                 fprintf (outfile, "%10llu %10d %8llu %s\n", cd->alloc_size, cd->allocs, cd->alloc_size / cd->allocs, cd->name);
1683                 dump_traces (&cd->traces, "bytes");
1684         }
1685         if (allocs)
1686                 fprintf (outfile, "Total memory allocated: %llu bytes in %d objects\n", size, allocs);
1687 }
1688
1689 enum {
1690         METHOD_SORT_TOTAL,
1691         METHOD_SORT_SELF,
1692         METHOD_SORT_CALLS
1693 };
1694
1695 static int method_sort_mode = METHOD_SORT_TOTAL;
1696
1697 static int
1698 compare_method (const void *a, const void *b)
1699 {
1700         MethodDesc *const*A = a;
1701         MethodDesc *const*B = b;
1702         uint64_t vala, valb;
1703         if (method_sort_mode == METHOD_SORT_SELF) {
1704                 vala = (*A)->self_time;
1705                 valb = (*B)->self_time;
1706         } else if (method_sort_mode == METHOD_SORT_CALLS) {
1707                 vala = (*A)->calls;
1708                 valb = (*B)->calls;
1709         } else {
1710                 vala = (*A)->total_time;
1711                 valb = (*B)->total_time;
1712         }
1713         if (vala == valb)
1714                 return 0;
1715         if (valb < vala)
1716                 return -1;
1717         return 1;
1718 }
1719
1720 static void
1721 dump_metadata (void)
1722 {
1723         fprintf (outfile, "\nMetadata summary\n");
1724         fprintf (outfile, "\tLoaded images: %d\n", num_images);
1725         if (verbose) {
1726                 ImageDesc *image;
1727                 int i;
1728                 for (i = 0; i < SMALL_HASH_SIZE; ++i) {
1729                         image = image_hash [i];
1730                         while (image) {
1731                                 fprintf (outfile, "\t\t%s\n", image->filename);
1732                                 image = image->next;
1733                         }
1734                 }
1735         }
1736
1737 }
1738
1739 static void
1740 dump_methods (void)
1741 {
1742         int i, c;
1743         uint64_t calls = 0;
1744         int header_done = 0;
1745         MethodDesc **methods = malloc (num_methods * sizeof (void*));
1746         MethodDesc *cd;
1747         c = 0;
1748         for (i = 0; i < HASH_SIZE; ++i) {
1749                 cd = method_hash [i];
1750                 while (cd) {
1751                         cd->total_time = cd->self_time + cd->callee_time;
1752                         methods [c++] = cd;
1753                         cd = cd->next;
1754                 }
1755         }
1756         qsort (methods, num_methods, sizeof (void*), compare_method);
1757         for (i = 0; i < num_methods; ++i) {
1758                 uint64_t msecs;
1759                 uint64_t smsecs;
1760                 cd = methods [i];
1761                 if (!cd->calls)
1762                         continue;
1763                 calls += cd->calls;
1764                 msecs = cd->total_time / 1000000;
1765                 smsecs = (cd->total_time - cd->callee_time) / 1000000;
1766                 if (!msecs && !verbose)
1767                         continue;
1768                 if (!header_done++) {
1769                         fprintf (outfile, "\nMethod call summary\n");
1770                         fprintf (outfile, "%8s %8s %10s Method name\n", "Total(ms)", "Self(ms)", "Calls");
1771                 }
1772                 fprintf (outfile, "%8llu %8llu %10llu %s\n", msecs, smsecs, cd->calls, cd->name);
1773                 dump_traces (&cd->traces, "calls");
1774         }
1775         if (calls)
1776                 fprintf (outfile, "Total calls: %llu\n", calls);
1777 }
1778
1779 static int
1780 compare_heap_class (const void *a, const void *b)
1781 {
1782         HeapClassDesc *const*A = a;
1783         HeapClassDesc *const*B = b;
1784         uint64_t vala, valb;
1785         if (alloc_sort_mode == ALLOC_SORT_BYTES) {
1786                 vala = (*A)->total_size;
1787                 valb = (*B)->total_size;
1788         } else {
1789                 vala = (*A)->count;
1790                 valb = (*B)->count;
1791         }
1792         if (valb == vala)
1793                 return 0;
1794         if (valb < vala)
1795                 return -1;
1796         return 1;
1797 }
1798
1799 static int
1800 compare_rev_class (const void *a, const void *b)
1801 {
1802         const HeapClassRevRef *A = a;
1803         const HeapClassRevRef *B = b;
1804         if (B->count == A->count)
1805                 return 0;
1806         if (B->count < A->count)
1807                 return -1;
1808         return 1;
1809 }
1810
1811 static void
1812 dump_rev_claases (HeapClassRevRef *revs, int count)
1813 {
1814         int j;
1815         if (!show_traces)
1816                 return;
1817         if (!count)
1818                 return;
1819         for (j = 0; j < count; ++j) {
1820                 HeapClassDesc *cd = revs [j].klass;
1821                 fprintf (outfile, "\t\t%llu references from: %s\n", revs [j].count, cd->klass->name);
1822         }
1823 }
1824
1825 static void
1826 heap_shot_summary (HeapShot *hs, int hs_num, HeapShot *last_hs)
1827 {
1828         uint64_t size = 0;
1829         uint64_t count = 0;
1830         int ccount = 0;
1831         int i;
1832         HeapClassDesc *cd;
1833         HeapClassDesc **sorted;
1834         sorted = malloc (sizeof (void*) * hs->class_count);
1835         for (i = 0; i < hs->hash_size; ++i) {
1836                 cd = hs->class_hash [i];
1837                 if (!cd)
1838                         continue;
1839                 count += cd->count;
1840                 size += cd->total_size;
1841                 sorted [ccount++] = cd;
1842         }
1843         hs->sorted = sorted;
1844         qsort (sorted, ccount, sizeof (void*), compare_heap_class);
1845         fprintf (outfile, "\n\tHeap shot %d at %.3f secs: size: %llu, object count: %llu, class count: %d\n",
1846                 hs_num, (hs->timestamp - startup_time)/1000000000.0, size, count, ccount);
1847         if (!verbose && ccount > 30)
1848                 ccount = 30;
1849         fprintf (outfile, "\t%10s %10s %8s Class name\n", "Bytes", "Count", "Average");
1850         for (i = 0; i < ccount; ++i) {
1851                 HeapClassRevRef *rev_sorted;
1852                 int j, k;
1853                 HeapClassDesc *ocd = NULL;
1854                 cd = sorted [i];
1855                 if (last_hs)
1856                         ocd = heap_class_lookup (last_hs, cd->klass);
1857                 fprintf (outfile, "\t%10llu %10llu %8llu %s", cd->total_size, cd->count, cd->total_size / cd->count, cd->klass->name);
1858                 if (ocd) {
1859                         int64_t bdiff = cd->total_size - ocd->total_size;
1860                         int64_t cdiff = cd->count - ocd->count;
1861                         fprintf (outfile, " (bytes: %+lld, count: %+lld)\n", bdiff, cdiff);
1862                 } else {
1863                         fprintf (outfile, "\n");
1864                 }
1865                 if (!collect_traces)
1866                         continue;
1867                 rev_sorted = malloc (cd->rev_count * sizeof (HeapClassRevRef));
1868                 k = 0;
1869                 for (j = 0; j < cd->rev_hash_size; ++j) {
1870                         if (cd->rev_hash [j].klass)
1871                                 rev_sorted [k++] = cd->rev_hash [j];
1872                 }
1873                 assert (cd->rev_count == k);
1874                 qsort (rev_sorted, cd->rev_count, sizeof (HeapClassRevRef), compare_rev_class);
1875                 dump_rev_claases (rev_sorted, cd->rev_count);
1876                 free (rev_sorted);
1877         }
1878         free (sorted);
1879 }
1880
1881 static int
1882 compare_heap_shots (const void *a, const void *b)
1883 {
1884         HeapShot *const*A = a;
1885         HeapShot *const*B = b;
1886         if ((*B)->timestamp == (*A)->timestamp)
1887                 return 0;
1888         if ((*B)->timestamp > (*A)->timestamp)
1889                 return -1;
1890         return 1;
1891 }
1892
1893 static void
1894 dump_heap_shots (void)
1895 {
1896         HeapShot **hs_sorted;
1897         HeapShot *hs;
1898         HeapShot *last_hs = NULL;
1899         int i;
1900         if (!heap_shots)
1901                 return;
1902         hs_sorted = malloc (num_heap_shots * sizeof (void*));
1903         fprintf (outfile, "\nHeap shot summary\n");
1904         i = 0;
1905         for (hs = heap_shots; hs; hs = hs->next)
1906                 hs_sorted [i++] = hs;
1907         qsort (hs_sorted, num_heap_shots, sizeof (void*), compare_heap_shots);
1908         for (i = 0; i < num_heap_shots; ++i) {
1909                 hs = hs_sorted [i];
1910                 heap_shot_summary (hs, i, last_hs);
1911                 last_hs = hs;
1912         }
1913 }
1914
1915 static void
1916 flush_context (ProfContext *ctx)
1917 {
1918         ThreadContext *thread;
1919         /* FIXME: sometimes there are leftovers: indagate */
1920         for (thread = ctx->threads; thread; thread = thread->next) {
1921                 while (thread->stack_id) {
1922                         if (debug)
1923                                 fprintf (outfile, "thread %p has %d items on stack\n", (void*)thread->thread_id, thread->stack_id);
1924                         pop_method (thread, thread->stack [thread->stack_id - 1], thread->last_time);
1925                 }
1926         }
1927 }
1928
1929 static const char *reports = "header,gc,alloc,call,metadata,exception,monitor,thread,heapshot";
1930
1931 static const char*
1932 match_option (const char *p, const char *opt)
1933 {
1934         int len = strlen (opt);
1935         if (strncmp (p, opt, len) == 0) {
1936                 if (p [len] == ',')
1937                         len++;
1938                 return p + len;
1939         }
1940         return p;
1941 }
1942
1943 static int
1944 print_reports (ProfContext *ctx, const char *reps, int parse_only)
1945 {
1946         const char *opt;
1947         const char *p;
1948         for (p = reps; *p; p = opt) {
1949                 if ((opt = match_option (p, "header")) != p) {
1950                         if (!parse_only)
1951                                 dump_header (ctx);
1952                         continue;
1953                 }
1954                 if ((opt = match_option (p, "thread")) != p) {
1955                         if (!parse_only)
1956                                 dump_threads (ctx);
1957                         continue;
1958                 }
1959                 if ((opt = match_option (p, "gc")) != p) {
1960                         if (!parse_only)
1961                                 dump_gcs ();
1962                         continue;
1963                 }
1964                 if ((opt = match_option (p, "alloc")) != p) {
1965                         if (!parse_only)
1966                                 dump_allocations ();
1967                         continue;
1968                 }
1969                 if ((opt = match_option (p, "call")) != p) {
1970                         if (!parse_only)
1971                                 dump_methods ();
1972                         continue;
1973                 }
1974                 if ((opt = match_option (p, "metadata")) != p) {
1975                         if (!parse_only)
1976                                 dump_metadata ();
1977                         continue;
1978                 }
1979                 if ((opt = match_option (p, "exception")) != p) {
1980                         if (!parse_only)
1981                                 dump_exceptions ();
1982                         continue;
1983                 }
1984                 if ((opt = match_option (p, "monitor")) != p) {
1985                         if (!parse_only)
1986                                 dump_monitors ();
1987                         continue;
1988                 }
1989                 if ((opt = match_option (p, "heapshot")) != p) {
1990                         if (!parse_only)
1991                                 dump_heap_shots ();
1992                         continue;
1993                 }
1994                 return 0;
1995         }
1996         return 1;
1997 }
1998
1999 static int
2000 add_find_spec (const char *p)
2001 {
2002         if (p [0] == 'S' && p [1] == ':') {
2003                 char *vale;
2004                 find_size = strtoul (p + 2, &vale, 10);
2005                 return 1;
2006         } else if (p [0] == 'T' && p [1] == ':') {
2007                 find_name = p + 2;
2008                 return 1;
2009         }
2010         return 0;
2011 }
2012
2013 static void
2014 usage (void)
2015 {
2016         printf ("Mono log profiler report version %d.%d\n", LOG_VERSION_MAJOR, LOG_VERSION_MINOR);
2017         printf ("Usage: mprof-report [OPTIONS] FILENAME\n");
2018         printf ("FILENAME can be '-' to read from standard input.\n");
2019         printf ("Options:\n");
2020         printf ("\t--help               display this help\n");
2021         printf ("\t--out=FILE           write to FILE instead of stdout\n");
2022         printf ("\t--traces             collect and show backtraces\n");
2023         printf ("\t--maxframes=NUM      limit backtraces to NUM entries\n");
2024         printf ("\t--reports=R1[,R2...] print the specified reports. Defaults are:\n");
2025         printf ("\t                     %s\n", reports);
2026         printf ("\t--method-sort=MODE   sort methods according to MODE: total, self, calls\n");
2027         printf ("\t--alloc-sort=MODE    sort allocations according to MODE: bytes, count\n");
2028         printf ("\t--track=OB1[,OB2...] track what happens to objects OBJ1, O2 etc.\n");
2029         printf ("\t--find=FINDSPEC      find and track objects matching FINFSPEC, where FINDSPEC is:\n");
2030         printf ("\t                     S:minimum_size or T:partial_name\n");
2031         printf ("\t--thread=THREADID    consider just the data for thread THREADID\n");
2032         printf ("\t--time=FROM-TO       consider data FROM seconds from startup up to TO seconds\n");
2033         printf ("\t--verbose            increase verbosity level\n");
2034         printf ("\t--debug              display decoding debug info for mprof-report devs\n");
2035 }
2036
2037 int
2038 main (int argc, char *argv[])
2039 {
2040         ProfContext *ctx;
2041         int i;
2042         outfile = stdout;
2043         for (i = 1; i < argc; ++i) {
2044                 if (strcmp ("--debug", argv [i]) == 0) {
2045                         debug++;
2046                 } else if (strcmp ("--help", argv [i]) == 0) {
2047                         usage ();
2048                         return 0;
2049                 } else if (strncmp ("--alloc-sort=", argv [i], 13) == 0) {
2050                         const char *val = argv [i] + 13;
2051                         if (strcmp (val, "bytes") == 0) {
2052                                 alloc_sort_mode = ALLOC_SORT_BYTES;
2053                         } else if (strcmp (val, "count") == 0) {
2054                                 alloc_sort_mode = ALLOC_SORT_COUNT;
2055                         } else {
2056                                 usage ();
2057                                 return 1;
2058                         }
2059                 } else if (strncmp ("--method-sort=", argv [i], 14) == 0) {
2060                         const char *val = argv [i] + 14;
2061                         if (strcmp (val, "total") == 0) {
2062                                 method_sort_mode = METHOD_SORT_TOTAL;
2063                         } else if (strcmp (val, "self") == 0) {
2064                                 method_sort_mode = METHOD_SORT_SELF;
2065                         } else if (strcmp (val, "calls") == 0) {
2066                                 method_sort_mode = METHOD_SORT_CALLS;
2067                         } else {
2068                                 usage ();
2069                                 return 1;
2070                         }
2071                 } else if (strncmp ("--reports=", argv [i], 10) == 0) {
2072                         const char *val = argv [i] + 10;
2073                         if (!print_reports (NULL, val, 1)) {
2074                                 usage ();
2075                                 return 1;
2076                         }
2077                         reports = val;
2078                 } else if (strncmp ("--out=", argv [i], 6) == 0) {
2079                         const char *val = argv [i] + 6;
2080                         outfile = fopen (val, "w");
2081                         if (!outfile) {
2082                                 printf ("Cannot open output file: %s\n", val);
2083                                 return 1;
2084                         }
2085                 } else if (strncmp ("--maxframes=", argv [i], 12) == 0) {
2086                         const char *val = argv [i] + 12;
2087                         char *vale;
2088                         trace_max = strtoul (val, &vale, 10);
2089                 } else if (strncmp ("--find=", argv [i], 7) == 0) {
2090                         const char *val = argv [i] + 7;
2091                         if (!add_find_spec (val)) {
2092                                 usage ();
2093                                 return 1;
2094                         }
2095                 } else if (strncmp ("--track=", argv [i], 8) == 0) {
2096                         const char *val = argv [i] + 8;
2097                         char *vale;
2098                         while (*val) {
2099                                 uintptr_t tracked_obj;
2100                                 if (*val == ',') {
2101                                         val++;
2102                                         continue;
2103                                 }
2104                                 tracked_obj = strtoul (val, &vale, 0);
2105                                 found_object (tracked_obj);
2106                                 val = vale;
2107                         }
2108                 } else if (strncmp ("--thread=", argv [i], 9) == 0) {
2109                         const char *val = argv [i] + 9;
2110                         char *vale;
2111                         thread_filter = strtoul (val, &vale, 0);
2112                 } else if (strncmp ("--time=", argv [i], 7) == 0) {
2113                         char *val = pstrdup (argv [i] + 7);
2114                         double from_secs, to_secs;
2115                         char *top = strchr (val, '-');
2116                         if (!top) {
2117                                 usage ();
2118                                 return 1;
2119                         }
2120                         *top++ = 0;
2121                         from_secs = atof (val);
2122                         to_secs = atof (top);
2123                         free (val);
2124                         if (from_secs > to_secs) {
2125                                 usage ();
2126                                 return 1;
2127                         }
2128                         time_from = from_secs * 1000000000;
2129                         time_to = to_secs * 1000000000;
2130                 } else if (strcmp ("--verbose", argv [i]) == 0) {
2131                         verbose++;
2132                 } else if (strcmp ("--traces", argv [i]) == 0) {
2133                         show_traces = 1;
2134                         collect_traces = 1;
2135                 } else {
2136                         break;
2137                 }
2138         }
2139         if (i >= argc) {
2140                 usage ();
2141                 return 2;
2142         }
2143         ctx = load_file (argv [i]);
2144         if (!ctx) {
2145                 printf ("Not a log profiler data file (or unsupported version).\n");
2146                 return 1;
2147         }
2148         while (decode_buffer (ctx));
2149         flush_context (ctx);
2150         if (num_tracked_objects)
2151                 return 0;
2152         print_reports (ctx, reports, 0);
2153         return 0;
2154 }
2155