14510Speter #ifndef lint 2*4563Speter static char *sccsid = "@(#)gprof.c 1.2 (Berkeley) 10/20/81"; 34510Speter #endif lint 44510Speter 5*4563Speter #include "gprof.h" 64510Speter 74510Speter main(argc, argv) 84510Speter int argc; 94510Speter char **argv; 104510Speter { 114510Speter 124510Speter --argc; 134510Speter argv++; 144510Speter debug = 0; 154510Speter while ( *argv != 0 && **argv == '-' ) { 164510Speter (*argv)++; 174510Speter if ( **argv == 'd' ) { 184510Speter (*argv)++; 194510Speter debug |= atoi( *argv ); 204510Speter debug |= ANYDEBUG; 214510Speter # ifdef DEBUG 224510Speter printf( "[main] debug = %d\n" , debug ); 234510Speter # endif DEBUG 244510Speter } else if ( **argv == 'z' ) { 254510Speter zflg++; 264510Speter } 274510Speter argv++; 284510Speter } 294510Speter if ( *argv != 0 ) { 304510Speter a_outname = *argv; 314510Speter argv++; 324510Speter } else { 334510Speter a_outname = A_OUTNAME; 344510Speter } 354510Speter if ( *argv != 0 ) { 36*4563Speter gmonname = *argv; 374510Speter argv++; 384510Speter } else { 39*4563Speter gmonname = GMONNAME; 404510Speter } 414510Speter /* 424510Speter * get information about a.out file. 434510Speter */ 444510Speter getnfile(); 454510Speter /* 464510Speter * get information about mon.out file(s). 474510Speter */ 48*4563Speter getpfile( gmonname ); 494510Speter /* 504510Speter * assign samples to procedures 514510Speter */ 524510Speter asgnsamples(); 534510Speter /* 544510Speter * print the usual profile 554510Speter */ 564510Speter printprof(); 574510Speter /* 584510Speter * assemble and print the dynamic profile 594510Speter */ 604510Speter doarcs(); 614510Speter done(); 624510Speter } 634510Speter 644510Speter printprof() 654510Speter { 664510Speter register nltype *np; 674510Speter nltype **sortednlp; 684510Speter int index; 694510Speter 704510Speter actime = 0.0; 714510Speter putprofheader(); 724510Speter /* 734510Speter * Sort the symbol table in by time 744510Speter */ 754510Speter sortednlp = (nltype **) calloc( nname , sizeof(nltype *) ); 764510Speter if ( sortednlp == (nltype **) 0 ) { 774510Speter fprintf( stderr , "[printprof] ran out of memory for time sorting\n" ); 784510Speter } 794510Speter for ( index = 0 ; index < nname ; index += 1 ) { 804510Speter sortednlp[ index ] = &nl[ index ]; 814510Speter } 824510Speter qsort( sortednlp , nname , sizeof(nltype *) , timecmp ); 834510Speter for ( index = 0 ; index < nname ; index += 1 ) { 844510Speter np = sortednlp[ index ]; 854510Speter putprofline( np , 1 ); 864510Speter } 874510Speter actime = 0.0; 884510Speter printf( "\ngranularity: each sample hit covers %.1f bytes" , scale ); 894510Speter printf( " for %.2f%% of %.2f seconds\n" , 100.0/totime , totime / HZ ); 904510Speter } 914510Speter 924510Speter putprofline( np , cumflag ) 934510Speter register nltype *np; 944510Speter int cumflag; 954510Speter { 964510Speter double time; 974510Speter long calls = np -> ncall + np -> selfcalls; 984510Speter 994510Speter if ( zflg == 0 && calls == 0 && np -> time == 0 && np -> childtime == 0 ) { 1004510Speter return; 1014510Speter } 1024510Speter if ( cumflag ) { 1034510Speter time = (np->time + np->childtime) / totime; 1044510Speter actime += np->time; 1054510Speter if ( np -> index != 0 ) { 1064510Speter printf( "[%d]" , np -> index ); 1074510Speter } 1084510Speter printf( "\t%5.1f %7.1f" , 100 * time , actime / HZ ); 1094510Speter } else { 1104510Speter printf( "\t%5.5s %7.7s" , "" , "" ); 1114510Speter } 1124510Speter printf( " %7.1f", np -> time / HZ ); 1134510Speter if ( np -> childtime != 0.0 ) { 1144510Speter printf( " %7.1f" , np -> childtime / HZ ); 1154510Speter } else { 1164510Speter printf( " %7.7s" , "" ); 1174510Speter } 1184510Speter if ( calls != 0 ) { 1194510Speter printf( " %7d" , np -> ncall ); 1204510Speter if ( np -> selfcalls != 0 ) { 1214510Speter printf( "+%-7d " , np -> selfcalls ); 1224510Speter } else { 1234510Speter printf( " %7.7s " , "" ); 1244510Speter } 1254510Speter } else { 1264510Speter printf( " %7.7s %7.7s " , "" , "" ); 1274510Speter } 1284510Speter if ( ! cumflag ) { 1294510Speter printf( " " ); 1304510Speter } 1314510Speter printname( np ); 1324510Speter printf( "\n" ); 1334510Speter } 1344510Speter 1354510Speter /* 1364510Speter * header for putprofline 1374510Speter */ 1384510Speter putprofheader() 1394510Speter { 1404510Speter 1414510Speter printf( "\n\t%5.5s %7.7s %-7.7s %-7.7s %7.7s %7.7s %5.5s\n" , 1424510Speter "%time" , "cumsecs" , " self" , " child" , "ncall" , "" , "name" ); 1434510Speter } 1444510Speter 1454510Speter /* 1464510Speter * Set up string and symbol tables from a.out. 1474510Speter * On return symbol table is sorted by value. 1484510Speter */ 1494510Speter getnfile() 1504510Speter { 1514510Speter FILE *nfile; 1524510Speter 1534510Speter nfile = fopen( a_outname ,"r"); 1544510Speter if (nfile == NULL) { 1554510Speter perror( a_outname ); 1564510Speter done(); 1574510Speter } 1584510Speter fread(&xbuf, 1, sizeof(xbuf), nfile); 1594510Speter if (N_BADMAG(xbuf)) { 1604510Speter fprintf(stderr, "%s: bad format\n", a_outname ); 1614510Speter done(); 1624510Speter } 1634510Speter getstrtab(nfile); 1644510Speter getsymtab(nfile); 1654510Speter qsort(nl, nname, sizeof(nltype), valcmp); 1664510Speter fclose(nfile); 1674510Speter # ifdef DEBUG 1684510Speter if ( debug & AOUTDEBUG ) { 1694510Speter register int j; 1704510Speter 1714510Speter for (j = 0; j < nname; j++){ 1724510Speter printf("[getnfile] 0X%08x\t%s\n", nl[j].value, nl[j].name); 1734510Speter } 1744510Speter } 1754510Speter # endif DEBUG 1764510Speter } 1774510Speter 1784510Speter getstrtab(nfile) 1794510Speter FILE *nfile; 1804510Speter { 1814510Speter 1824510Speter fseek(nfile, (long)(N_SYMOFF(xbuf) + xbuf.a_syms), 0); 1834510Speter if (fread(&ssiz, sizeof (ssiz), 1, nfile) == 0) { 1844510Speter fprintf(stderr, "%s: no string table (old format?)\n", a_outname ); 1854510Speter done(); 1864510Speter } 1874510Speter strtab = (char *)calloc(ssiz, 1); 1884510Speter if (strtab == NULL) { 1894510Speter fprintf(stderr, "%s: no room for %d bytes of string table", 1904510Speter a_outname , ssiz); 1914510Speter done(); 1924510Speter } 1934510Speter if (fread(strtab+sizeof(ssiz), ssiz-sizeof(ssiz), 1, nfile) != 1) { 1944510Speter fprintf(stderr, "%s: error reading string table\n", a_outname ); 1954510Speter done(); 1964510Speter } 1974510Speter } 1984510Speter 1994510Speter /* 2004510Speter * Read in symbol table 2014510Speter */ 2024510Speter getsymtab(nfile) 2034510Speter FILE *nfile; 2044510Speter { 2054510Speter register long i; 2064510Speter int askfor; 2074510Speter struct nlist nbuf; 2084510Speter 2094510Speter /* pass1 - count symbols */ 2104510Speter fseek(nfile, (long)N_SYMOFF(xbuf), 0); 2114510Speter nname = 0; 2124510Speter for (i = xbuf.a_syms; i > 0; i -= sizeof(struct nlist)) { 2134510Speter fread(&nbuf, sizeof(nbuf), 1, nfile); 2144510Speter if ( nbuf.n_type != N_TEXT+N_EXT ) { 2154510Speter continue; 2164510Speter } 2174510Speter nname++; 2184510Speter } 2194510Speter if (nname == 0) { 2204510Speter fprintf(stderr, "%s: no symbols\n", a_outname ); 2214510Speter done(); 2224510Speter } 2234510Speter /* 2244510Speter * ask also for CYCLEFRACTION extra namelist entries for 2254510Speter * cycle entries. these hide out at the end of the namelist 2264510Speter * and aren't accessed unless the whole namelist (nname+ncycles) 2274510Speter * is sorted and searched. 2284510Speter */ 2294510Speter ncycles = nname * CYCLEFRACTION; 2304510Speter askfor = nname + 1 + ncycles; 2314510Speter nl = (nltype *) calloc( askfor , sizeof(nltype) ); 2324510Speter if (nl == 0) { 2334510Speter fprintf(stderr, "prof: No room for %d bytes of symbol table\n", 2344510Speter askfor * sizeof(nltype) ); 2354510Speter done(); 2364510Speter } 2374510Speter 2384510Speter /* pass2 - read symbols */ 2394510Speter fseek(nfile, (long)N_SYMOFF(xbuf), 0); 2404510Speter npe = nl; 2414510Speter nname = 0; 2424510Speter for (i = xbuf.a_syms; i > 0; i -= sizeof(struct nlist)) { 2434510Speter fread(&nbuf, sizeof(nbuf), 1, nfile); 2444510Speter if ( nbuf.n_type != N_TEXT+N_EXT ) { 2454510Speter continue; 2464510Speter } 2474510Speter npe->value = nbuf.n_value; 2484510Speter npe->name = strtab+nbuf.n_un.n_strx; 2494510Speter # ifdef DEBUG 2504510Speter if ( debug & AOUTDEBUG ) { 2514510Speter printf( "[getsymtab] %d %s 0x%08x\n" , 2524510Speter nname , npe -> name , npe -> value ); 2534510Speter } 2544510Speter # endif DEBUG 2554510Speter npe++; 2564510Speter nname++; 2574510Speter } 2584510Speter npe->value = -1; 2594510Speter npe++; 2604510Speter } 2614510Speter 2624510Speter /* 263*4563Speter * information from a gmon.out file is in two parts: 2644510Speter * an array of sampling hits within pc ranges, 2654510Speter * and the arcs. 2664510Speter */ 2674510Speter getpfile(filename) 2684510Speter char *filename; 2694510Speter { 2704510Speter FILE *pfile; 2714510Speter FILE *openpfile(); 2724510Speter struct rawarc arc; 2734510Speter 2744510Speter pfile = openpfile(filename); 2754510Speter readsamples(pfile); 2764510Speter /* 2774510Speter * the rest of the file consists of 2784510Speter * a bunch of <from,self,count> tuples. 2794510Speter */ 2804510Speter while ( fread( &arc , sizeof arc , 1 , pfile ) == 1 ) { 2814510Speter # ifdef DEBUG 2824510Speter if ( debug & SAMPLEDEBUG ) { 2834510Speter printf( "[getpfile] frompc %d selfpc %d count %d\n" , 2844510Speter arc.raw_frompc , arc.raw_selfpc , arc.raw_count ); 2854510Speter } 2864510Speter # endif DEBUG 2874510Speter /* 2884510Speter * add this arc 2894510Speter */ 2904510Speter tally( &arc ); 2914510Speter } 2924510Speter fclose(pfile); 2934510Speter } 2944510Speter 2954510Speter FILE *openpfile(filename) 2964510Speter char *filename; 2974510Speter { 2984510Speter FILE *pfile; 2994510Speter 3004510Speter if((pfile = fopen(filename, "r")) == NULL) { 3014510Speter perror(filename); 3024510Speter done(); 3034510Speter } 3044510Speter fread(&h, sizeof(struct hdr), 1, pfile); 3054510Speter lowpc = h.lowpc - (UNIT *)0; 3064510Speter highpc = h.highpc - (UNIT *)0; 3074510Speter sampbytes = h.ncnt - sizeof(struct hdr); 3084510Speter nsamples = sampbytes / sizeof (unsigned UNIT); 3094510Speter return(pfile); 3104510Speter } 3114510Speter 3124510Speter tally( rawp ) 3134510Speter struct rawarc *rawp; 3144510Speter { 3154510Speter nltype *parentp; 3164510Speter nltype *childp; 3174510Speter arctype *arcp; 3184510Speter arctype *malloc(); 3194510Speter 3204510Speter parentp = nllookup( rawp -> raw_frompc ); 3214510Speter childp = nllookup( rawp -> raw_selfpc ); 3224510Speter childp -> ncall += rawp -> raw_count; 3234510Speter # ifdef DEBUG 3244510Speter if ( debug & TALLYDEBUG ) { 3254510Speter printf( "[tally] arc from %s to %s traversed %d times\n" , 3264510Speter parentp -> name , childp -> name , rawp -> raw_count ); 3274510Speter } 3284510Speter # endif DEBUG 3294510Speter arcp = arclookup( parentp , childp ); 3304510Speter if ( arcp != 0 ) { 3314510Speter /* 3324510Speter * a hit: just increment the count. 3334510Speter */ 3344510Speter # ifdef DEBUG 3354510Speter if ( debug & TALLYDEBUG ) { 3364510Speter printf( "[tally] hit %d += %d\n" , 3374510Speter arcp -> arc_count , rawp -> raw_count ); 3384510Speter } 3394510Speter # endif DEBUG 3404510Speter arcp -> arc_count += rawp -> raw_count; 3414510Speter return; 3424510Speter } 3434510Speter arcp = malloc( sizeof *arcp ); 3444510Speter arcp -> arc_parentp = parentp; 3454510Speter arcp -> arc_childp = childp; 3464510Speter arcp -> arc_count = rawp -> raw_count; 3474510Speter /* 3484510Speter * prepend this child to the children of this parent 3494510Speter */ 3504510Speter arcp -> arc_childlist = parentp -> children; 3514510Speter parentp -> children = arcp; 3524510Speter /* 3534510Speter * prepend this parent to the parents of this child 3544510Speter */ 3554510Speter arcp -> arc_parentlist = childp -> parents; 3564510Speter childp -> parents = arcp; 3574510Speter } 3584510Speter 3594510Speter valcmp(p1, p2) 3604510Speter nltype *p1, *p2; 3614510Speter { 3624510Speter if ( p1 -> value < p2 -> value ) { 3634510Speter return LESSTHAN; 3644510Speter } 3654510Speter if ( p1 -> value > p2 -> value ) { 3664510Speter return GREATERTHAN; 3674510Speter } 3684510Speter return EQUALTO; 3694510Speter } 3704510Speter 3714510Speter readsamples(pfile) 3724510Speter FILE *pfile; 3734510Speter { 3744510Speter register i; 3754510Speter unsigned UNIT sample; 3764510Speter 3774510Speter if (samples == 0) { 3784510Speter samples = (unsigned UNIT *) calloc(sampbytes, sizeof (unsigned UNIT)); 3794510Speter if (samples == 0) { 3804510Speter fprintf( stderr , "prof: No room for %d sample pc's\n", 3814510Speter sampbytes / sizeof (unsigned UNIT)); 3824510Speter done(); 3834510Speter } 3844510Speter } 3854510Speter for (i = 0; i < nsamples; i++) { 3864510Speter fread(&sample, sizeof (unsigned UNIT), 1, pfile); 3874510Speter if (feof(pfile)) 3884510Speter break; 3894510Speter samples[i] += sample; 3904510Speter } 3914510Speter if (i != nsamples) { 3924510Speter fprintf(stderr, 3934510Speter "prof: unexpected EOF after reading %d/%d samples\n", 3944510Speter --i, nsamples); 3954510Speter done(); 3964510Speter } 3974510Speter } 3984510Speter 3994510Speter /* 4004510Speter * Assign samples to the procedures to which they belong. 4014510Speter */ 4024510Speter asgnsamples() 4034510Speter { 4044510Speter register int j; 4054510Speter unsigned UNIT ccnt; 4064510Speter double time; 4074510Speter unsigned long pcl, pch; 4084510Speter register int i; 4094510Speter int overlap; 4104510Speter unsigned long svalue0, svalue1; 4114510Speter 4124510Speter /* read samples and assign to namelist symbols */ 4134510Speter scale = highpc - lowpc; 4144510Speter scale /= nsamples; 4154510Speter for (i=0; i < nsamples; i++) { 4164510Speter ccnt = samples[i]; 4174510Speter if (ccnt == 0) 4184510Speter continue; 4194510Speter pcl = lowpc + scale*i; 4204510Speter pch = lowpc + scale*(i+1); 4214510Speter time = ccnt; 4224510Speter # ifdef DEBUG 4234510Speter if ( debug & SAMPLEDEBUG ) { 4244510Speter printf( "[asgnsamples] ccnt %d time %f totime %f\n" , 4254510Speter ccnt , time , totime ); 4264510Speter } 4274510Speter # endif DEBUG 4284510Speter totime += time; 4294510Speter for (j=0; j<nname; j++) { 4304510Speter svalue0 = nl[j].value / sizeof(UNIT); 4314510Speter svalue1 = nl[j+1].value / sizeof(UNIT); 4324510Speter if (pch < svalue0) 4334510Speter break; 4344510Speter if (pcl >= svalue1) 4354510Speter continue; 4364510Speter overlap=min(pch,svalue1) - max(pcl,svalue0); 4374510Speter if (overlap>0) { 4384510Speter # ifdef DEBUG 4394510Speter if ( debug & SAMPLEDEBUG ) { 4404510Speter printf( "[asgnsamples] %s gets %f ticks\n" , 4414510Speter nl[j].name , overlap*time/scale ); 4424510Speter } 4434510Speter # endif DEBUG 4444510Speter nl[j].time += overlap*time/scale; 4454510Speter } 4464510Speter } 4474510Speter } 4484510Speter # ifdef DEBUG 4494510Speter if ( debug & SAMPLEDEBUG ) { 4504510Speter printf( "[asgnsamples] totime %f\n" , totime ); 4514510Speter } 4524510Speter # endif DEBUG 4534510Speter if (totime==0.0) { 4544510Speter fprintf( stderr , "No time accumulated\n" ); 4554510Speter totime=1.0; 4564510Speter } 4574510Speter } 4584510Speter 4594510Speter 4604510Speter min(a, b) 4614510Speter unsigned a,b; 4624510Speter { 4634510Speter if (a<b) 4644510Speter return(a); 4654510Speter return(b); 4664510Speter } 4674510Speter 4684510Speter max(a, b) 4694510Speter unsigned a,b; 4704510Speter { 4714510Speter if (a>b) 4724510Speter return(a); 4734510Speter return(b); 4744510Speter } 4754510Speter 4764510Speter timecmp( npp1 , npp2 ) 4774510Speter nltype **npp1, **npp2; 4784510Speter { 4794510Speter double d; 4804510Speter 4814510Speter d = (*npp2)->time - (*npp1)->time; 4824510Speter if (d > 0.0) 4834510Speter return(1); 4844510Speter if (d < 0.0) 4854510Speter return(-1); 4864510Speter return(strcmp((*npp1)->name,(*npp2)->name)); 4874510Speter } 4884510Speter 4894510Speter done() 4904510Speter { 4914510Speter 4924510Speter exit(0); 4934510Speter } 494