X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FStats.c;h=1fcb94d44b74f8475fba827dbbccf65fefb115c7;hb=c883f6969ad957637649f3af1a2b6977555bdd32;hp=9dd0612aa092ab11bb04844fd9a74cdb92d243bd;hpb=7f309f1c021e7583f724cce599ce2dd3c439361b;p=ghc-hetmet.git diff --git a/ghc/rts/Stats.c b/ghc/rts/Stats.c index 9dd0612..1fcb94d 100644 --- a/ghc/rts/Stats.c +++ b/ghc/rts/Stats.c @@ -1,70 +1,49 @@ /* ----------------------------------------------------------------------------- - * $Id: Stats.c,v 1.6 1999/02/05 16:02:56 simonm Exp $ * - * (c) The GHC Team, 1998-1999 + * (c) The GHC Team, 1998-2004 * * Statistics and timing-related functions. * * ---------------------------------------------------------------------------*/ -#define NON_POSIX_SOURCE +/* Alas, no. This source is non-posix. + #include "PosixSource.h" +*/ #include "Rts.h" #include "RtsFlags.h" #include "RtsUtils.h" -#include "StoragePriv.h" #include "MBlock.h" - -/** - * Ian: For the moment we just want to ignore - * these on Nemesis - **/ -#ifdef _NEMESIS_OS_ -#ifdef HAVE_SYS_TIMES_H -#undef HAVE_SYS_TIMES_H /* */ -#endif -#ifdef HAVE_SYS_RESOURCE_H /* */ -#undef HAVE_SYS_RESOURCE_H -#endif -#ifdef HAVE_SYS_TIME_H /* */ -#undef HAVE_SYS_TIME_H -#endif -#ifdef HAVE_SYS_TIMEB_H -#undef HAVE_SYS_TIMEB_H /* */ -#endif -#ifdef HAVE_UNISTD_H -#undef HAVE_UNISTD_H /* */ -#endif -#ifdef HAVE_TIMES -#undef HAVE_TIMES -#endif -#ifdef HAVE_FTIME -#undef HAVE_FTIME -#endif -#ifdef HAVE_GETRUSAGE -#undef HAVE_GETRUSAGE -#endif -#ifdef HAVE_SYSCONF -#undef HAVE_SYSCONF -#endif -#endif /* _NEMESIS_OS_ */ - +#include "Schedule.h" #include "Stats.h" +#include "ParTicky.h" /* ToDo: move into Rts.h */ +#include "Profiling.h" +#include "Storage.h" #ifdef HAVE_UNISTD_H #include #endif -#ifdef HAVE_SYS_TIMES_H -#include +#ifndef mingw32_HOST_OS +# ifdef HAVE_SYS_TIMES_H +# include +# endif #endif #ifdef HAVE_SYS_TIME_H #include #endif -#if defined(HAVE_SYS_RESOURCE_H) && ! irix_TARGET_OS -#include +#ifdef __CYGWIN32__ +# ifdef HAVE_TIME_H +# include +# endif +#endif + +#if ! irix_HOST_OS && ! defined(mingw32_HOST_OS) +# if defined(HAVE_SYS_RESOURCE_H) +# include +# endif #endif #ifdef HAVE_SYS_TIMEB_H @@ -75,65 +54,217 @@ #include #endif +#if HAVE_WINDOWS_H +#include +#endif + +#if defined(PAR) || !(!defined(HAVE_GETRUSAGE) || irix_HOST_OS || defined(mingw32_HOST_OS) || defined(cygwin32_HOST_OS)) +#include +#endif + /* huh? */ #define BIG_STRING_LEN 512 -static StgDouble ElapsedTimeStart = 0.0; -static StgDouble TicksPerSecond = 0.0; +/* We're not trying to be terribly accurate here, using the + * basic times() function to get a resolution of about 100ths of a + * second, depending on the OS. A long int will do fine for holding + * these values. + */ +#define TICK_TYPE long int +#define TICK_TO_DBL(t) ((double)(t) / TicksPerSecond) + +static int TicksPerSecond = 0; + +static TICK_TYPE ElapsedTimeStart = 0; +static TICK_TYPE CurrentElapsedTime = 0; +static TICK_TYPE CurrentUserTime = 0; -static StgDouble InitUserTime = 0.0; -static StgDouble InitElapsedTime = 0.0; +static TICK_TYPE InitUserTime = 0; +static TICK_TYPE InitElapsedTime = 0; +static TICK_TYPE InitElapsedStamp = 0; -static ullong GC_tot_alloc = 0; +static TICK_TYPE MutUserTime = 0; +static TICK_TYPE MutElapsedTime = 0; +static TICK_TYPE MutElapsedStamp = 0; -static StgDouble GC_start_time, GC_tot_time = 0; /* User GC Time */ -static StgDouble GCe_start_time, GCe_tot_time = 0; /* Elapsed GC time */ +static TICK_TYPE ExitUserTime = 0; +static TICK_TYPE ExitElapsedTime = 0; -lnat MaxResidency = 0; /* in words; for stats only */ -lnat ResidencySamples = 0; /* for stats only */ +static ullong GC_tot_alloc = 0; +static ullong GC_tot_copied = 0; + +static TICK_TYPE GC_start_time = 0, GC_tot_time = 0; /* User GC Time */ +static TICK_TYPE GCe_start_time = 0, GCe_tot_time = 0; /* Elapsed GC time */ + +#ifdef PROFILING +static TICK_TYPE RP_start_time = 0, RP_tot_time = 0; /* retainer prof user time */ +static TICK_TYPE RPe_start_time = 0, RPe_tot_time = 0; /* retainer prof elap time */ + +static TICK_TYPE HC_start_time, HC_tot_time = 0; // heap census prof user time +static TICK_TYPE HCe_start_time, HCe_tot_time = 0; // heap census prof elap time +#endif + +#ifdef PROFILING +#define PROF_VAL(x) (x) +#else +#define PROF_VAL(x) 0 +#endif + +static lnat MaxResidency = 0; // in words; for stats only +static lnat AvgResidency = 0; +static lnat ResidencySamples = 0; // for stats only static lnat GC_start_faults = 0, GC_end_faults = 0; -/* ToDo: convert this to use integers? --SDM */ +static TICK_TYPE *GC_coll_times; + +static void getTimes(void); +static nat pageFaults(void); + +static void statsPrintf( char *s, ... ); +static void statsFlush( void ); +static void statsClose( void ); /* elapsedtime() -- The current elapsed time in seconds */ -StgDouble -elapsedtime(void) +#if defined(mingw32_HOST_OS) || defined(cygwin32_HOST_OS) +#define HNS_PER_SEC 10000000LL /* FILETIMES are in units of 100ns */ +/* Convert FILETIMEs into secs */ +#define FT2longlong(ll,ft) \ + (ll)=(ft).dwHighDateTime; \ + (ll) <<= 32; \ + (ll) |= (ft).dwLowDateTime; \ + (ll) /= (unsigned long long) (HNS_PER_SEC / CLOCKS_PER_SEC) +#endif + +#if defined(mingw32_HOST_OS) || defined(cygwin32_HOST_OS) +/* cygwin32 or mingw32 version */ +static void +getTimes(void) +{ + static int is_win9x = -1; + + FILETIME creationTime, exitTime, userTime, kernelTime = {0,0}; + long long int kT, uT; + + if (is_win9x < 0) { + /* figure out whether we're on a Win9x box or not. */ + OSVERSIONINFO oi; + BOOL b; + + /* Need to init the size field first.*/ + oi.dwOSVersionInfoSize = sizeof(OSVERSIONINFO); + b = GetVersionEx(&oi); + + is_win9x = ( (b && (oi.dwPlatformId & VER_PLATFORM_WIN32_WINDOWS)) ? 1 : 0); + } + + if (is_win9x) { + /* On Win9x, just attribute all running time to the user. */ + SYSTEMTIME st; + + GetSystemTime(&st); + SystemTimeToFileTime(&st,&userTime); + } else { + /* ToDo: pin down elapsed times to just the OS thread(s) that + are evaluating/managing Haskell code. + */ + if (!GetProcessTimes (GetCurrentProcess(), &creationTime, + &exitTime, &kernelTime, &userTime)) { + /* Probably on a Win95 box..*/ + CurrentElapsedTime = 0; + CurrentUserTime = 0; + return; + } + } + + FT2longlong(kT,kernelTime); + FT2longlong(uT,userTime); + CurrentElapsedTime = uT + kT; + CurrentUserTime = uT; + + if (is_win9x) { + /* Adjust for the fact that we're using system time & not + process time on Win9x. */ + CurrentUserTime -= ElapsedTimeStart; + CurrentElapsedTime -= ElapsedTimeStart; + } +} + +#else /* !win32 */ + +static void +getTimes(void) { -#if ! (defined(HAVE_TIMES) || defined(HAVE_FTIME)) + +#ifndef HAVE_TIMES /* We will #ifdef around the fprintf for machines we *know* are unsupported. (WDP 94/05) */ - fprintf(stderr, "NOTE: `elapsedtime' does nothing!\n"); + debugBelch("NOTE: `getTimes' does nothing!\n"); return 0.0; #else /* not stumped */ - -/* "ftime" may be nicer, but "times" is more standard; - but, on a Sun, if you do not get the SysV one, you are *hosed*... - */ - -# if defined(HAVE_TIMES) && ! sunos4_TARGET_OS struct tms t; clock_t r = times(&t); - return (((StgDouble)r)/TicksPerSecond - ElapsedTimeStart); + CurrentElapsedTime = r; + CurrentUserTime = t.tms_utime; +#endif + +} +#endif /* !win32 */ + +/* mut_user_time_during_GC() and mut_user_time() + * + * The former function can be used to get the current mutator time + * *during* a GC, i.e. between stat_startGC and stat_endGC. This is + * used in the heap profiler for accurately time stamping the heap + * sample. + * + * ATTENTION: mut_user_time_during_GC() relies on GC_start_time being + * defined in stat_startGC() - to minimise system calls, + * GC_start_time is, however, only defined when really needed (check + * stat_startGC() for details) + */ +double +mut_user_time_during_GC( void ) +{ + return TICK_TO_DBL(GC_start_time - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time)); +} -# else /* HAVE_FTIME */ - struct timeb t; +double +mut_user_time( void ) +{ + getTimes(); + return TICK_TO_DBL(CurrentUserTime - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time)); +} - ftime(&t); - return (fabs(t.time + 1e-3*t.millitm - ElapsedTimeStart)); +#ifdef PROFILING +/* + mut_user_time_during_RP() is similar to mut_user_time_during_GC(); + it returns the MUT time during retainer profiling. + The same is for mut_user_time_during_HC(); + */ +double +mut_user_time_during_RP( void ) +{ + return TICK_TO_DBL(RP_start_time - GC_tot_time - RP_tot_time - HC_tot_time); +} -# endif /* HAVE_FTIME */ -#endif /* not stumped */ +double +mut_user_time_during_heap_census( void ) +{ + return TICK_TO_DBL(HC_start_time - GC_tot_time - RP_tot_time - HC_tot_time); } +#endif /* PROFILING */ static nat -pagefaults(void) +pageFaults(void) { -# if !defined(HAVE_GETRUSAGE) || irix_TARGET_OS + /* ToDo (on NT): better, get this via the performance data + that's stored in the registry. */ +# if !defined(HAVE_GETRUSAGE) || irix_HOST_OS || defined(mingw32_HOST_OS) || defined(cygwin32_HOST_OS) return 0; # else struct rusage t; @@ -143,86 +274,122 @@ pagefaults(void) # endif } -/* ToDo: use gettimeofday on systems that support it (u-sec accuracy) */ +void +initStats(void) +{ + nat i; + + if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) { + statsPrintf(" Alloc Collect Live GC GC TOT TOT Page Flts\n"); + statsPrintf(" bytes bytes bytes user elap user elap\n"); + } + GC_coll_times = + (TICK_TYPE *)stgMallocBytes( + sizeof(TICK_TYPE)*RtsFlags.GcFlags.generations, + "initStats"); + for (i = 0; i < RtsFlags.GcFlags.generations; i++) { + GC_coll_times[i] = 0; + } +} + +/* ----------------------------------------------------------------------------- + Initialisation time... + -------------------------------------------------------------------------- */ void -start_time(void) +stat_startInit(void) { - long ticks; /* Determine TicksPerSecond ... */ -#ifdef HAVE_SYSCONF +#if defined(CLK_TCK) /* defined by POSIX */ + TicksPerSecond = CLK_TCK; + +#elif defined(HAVE_SYSCONF) + long ticks; + ticks = sysconf(_SC_CLK_TCK); if ( ticks == -1 ) { - fprintf(stderr, "stat_init: bad call to 'sysconf'!\n"); + debugBelch("stat_init: bad call to 'sysconf'!\n"); stg_exit(EXIT_FAILURE); } - TicksPerSecond = (StgDouble) ticks; + TicksPerSecond = ticks; + +/* no "sysconf" or CLK_TCK; had better guess */ +#elif defined(HZ) + TicksPerSecond = HZ; -#else /* no "sysconf"; had better guess */ -# ifdef HZ - TicksPerSecond = (StgDouble) (HZ); +#elif defined(CLOCKS_PER_SEC) + TicksPerSecond = CLOCKS_PER_SEC; -# else /* had better guess wildly */ +#else /* had better guess wildly */ /* We will #ifdef around the fprintf for machines we *know* are unsupported. (WDP 94/05) */ - fprintf(stderr, "NOTE: Guessing `TicksPerSecond = 60'!\n"); - TicksPerSecond = 60.0; -# endif + debugBelch("NOTE: Guessing `TicksPerSecond = 60'!\n"); + TicksPerSecond = 60; #endif - ElapsedTimeStart = elapsedtime(); + getTimes(); + ElapsedTimeStart = CurrentElapsedTime; } - -void -initStats(void) +void +stat_endInit(void) { - FILE *sf = RtsFlags.GcFlags.statsFile; - - if (RtsFlags.GcFlags.giveStats) { - fprintf(sf, " Alloc Collect Live GC GC TOT TOT Page Flts\n"); - fprintf(sf, " bytes bytes bytes user elap user elap\n"); - } -} + getTimes(); + InitUserTime = CurrentUserTime; + InitElapsedStamp = CurrentElapsedTime; + if (ElapsedTimeStart > CurrentElapsedTime) { + InitElapsedTime = 0; + } else { + InitElapsedTime = CurrentElapsedTime - ElapsedTimeStart; + } +} +/* ----------------------------------------------------------------------------- + stat_startExit and stat_endExit + + These two measure the time taken in shutdownHaskell(). + -------------------------------------------------------------------------- */ -StgDouble -usertime(void) +void +stat_startExit(void) { -#if ! (defined(HAVE_GETRUSAGE) || defined(HAVE_TIMES)) - /* We will #ifdef around the fprintf for machines - we *know* are unsupported. (WDP 94/05) - */ - fprintf(stderr, "NOTE: `usertime' does nothing!\n"); - return 0.0; - -#else /* not stumped */ - -# if defined(HAVE_TIMES) - struct tms t; - - times(&t); - return(((StgDouble)(t.tms_utime))/TicksPerSecond); - -#else /* HAVE_GETRUSAGE */ - struct rusage t; - - getrusage(RUSAGE_SELF, &t); - return(t.ru_utime.tv_sec + 1e-6*t.ru_utime.tv_usec); - -# endif /* HAVE_GETRUSAGE */ -#endif /* not stumped */ + getTimes(); + MutElapsedStamp = CurrentElapsedTime; + MutElapsedTime = CurrentElapsedTime - GCe_tot_time - + PROF_VAL(RPe_tot_time + HCe_tot_time) - InitElapsedStamp; + if (MutElapsedTime < 0) { MutElapsedTime = 0; } /* sometimes -0.00 */ + + /* for SMP, we don't know the mutator time yet, we have to inspect + * all the running threads to find out, and they haven't stopped + * yet. So we just timestamp MutUserTime at this point so we can + * calculate the EXIT time. The real MutUserTime is calculated + * in stat_exit below. + */ +#ifdef SMP + MutUserTime = CurrentUserTime; +#else + MutUserTime = CurrentUserTime - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime; + if (MutUserTime < 0) { MutUserTime = 0; } +#endif } -void -end_init(void) +void +stat_endExit(void) { - InitUserTime = usertime(); - InitElapsedTime = elapsedtime(); - if (InitElapsedTime < 0.0) { - InitElapsedTime = 0.0; - } + getTimes(); +#ifdef SMP + ExitUserTime = CurrentUserTime - MutUserTime; +#else + ExitUserTime = CurrentUserTime - MutUserTime - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime; +#endif + ExitElapsedTime = CurrentElapsedTime - MutElapsedStamp; + if (ExitUserTime < 0) { + ExitUserTime = 0; + } + if (ExitElapsedTime < 0) { + ExitElapsedTime = 0; + } } /* ----------------------------------------------------------------------------- @@ -231,27 +398,38 @@ end_init(void) static nat rub_bell = 0; +/* initialise global variables needed during GC + * + * * GC_start_time is read in mut_user_time_during_GC(), which in turn is + * needed if either PROFILING or DEBUGing is enabled + */ void stat_startGC(void) { - FILE *sf = RtsFlags.GcFlags.statsFile; - nat bell = RtsFlags.GcFlags.ringBell; if (bell) { if (bell > 1) { - fprintf(stderr, " GC "); + debugBelch(" GC "); rub_bell = 1; } else { - fprintf(stderr, "\007"); + debugBelch("\007"); } } - if (sf != NULL) { - GC_start_time = usertime(); - GCe_start_time = elapsedtime(); +#if defined(PROFILING) || defined(DEBUG) + getTimes(); + GC_start_time = CurrentUserTime; /* needed in mut_user_time_during_GC() */ +#endif + + if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) { +#if !defined(PROFILING) && !defined(DEBUG) + getTimes(); + GC_start_time = CurrentUserTime; +#endif + GCe_start_time = CurrentElapsedTime; if (RtsFlags.GcFlags.giveStats) { - GC_start_faults = pagefaults(); + GC_start_faults = pageFaults(); } } } @@ -261,51 +439,177 @@ stat_startGC(void) -------------------------------------------------------------------------- */ void -stat_endGC(lnat alloc, lnat collect, lnat live, lnat gen) +stat_endGC(lnat alloc, lnat collect, lnat live, lnat copied, lnat gen) { - FILE *sf = RtsFlags.GcFlags.statsFile; - - if (sf != NULL) { - StgDouble time = usertime(); - StgDouble etime = elapsedtime(); - - if (RtsFlags.GcFlags.giveStats) { - nat faults = pagefaults(); - - fprintf(sf, "%9ld %9ld %9ld", + if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) { + TICK_TYPE time, etime, gc_time, gc_etime; + + getTimes(); + time = CurrentUserTime; + etime = CurrentElapsedTime; + gc_time = time - GC_start_time; + gc_etime = etime - GCe_start_time; + + if (RtsFlags.GcFlags.giveStats == VERBOSE_GC_STATS) { + nat faults = pageFaults(); + + statsPrintf("%9ld %9ld %9ld", alloc*sizeof(W_), collect*sizeof(W_), live*sizeof(W_)); - fprintf(sf, " %5.2f %5.2f %7.2f %7.2f %4ld %4ld (Gen: %2ld)\n", - (time-GC_start_time), - (etime-GCe_start_time), - time, - etime, + statsPrintf(" %5.2f %5.2f %7.2f %7.2f %4ld %4ld (Gen: %2ld)\n", + TICK_TO_DBL(gc_time), + TICK_TO_DBL(gc_etime), + TICK_TO_DBL(time), + TICK_TO_DBL(etime - ElapsedTimeStart), faults - GC_start_faults, GC_start_faults - GC_end_faults, gen); GC_end_faults = faults; - fflush(sf); + statsFlush(); } - GC_tot_alloc += (ullong) alloc; - GC_tot_time += time-GC_start_time; - GCe_tot_time += etime-GCe_start_time; + GC_coll_times[gen] += gc_time; + + GC_tot_copied += (ullong) copied; + GC_tot_alloc += (ullong) alloc; + GC_tot_time += gc_time; + GCe_tot_time += gc_etime; + +#ifdef SMP + { + nat i; + pthread_t me = pthread_self(); + + for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) { + if (me == task_ids[i].id) { + task_ids[i].gc_time += gc_time; + task_ids[i].gc_etime += gc_etime; + break; + } + } + } +#endif if (gen == RtsFlags.GcFlags.generations-1) { /* major GC? */ - if (live > MaxResidency) { - MaxResidency = live; - } - ResidencySamples++; + if (live > MaxResidency) { + MaxResidency = live; + } + ResidencySamples++; + AvgResidency += live; } } if (rub_bell) { - fprintf(stderr, "\b\b\b \b\b\b"); + debugBelch("\b\b\b \b\b\b"); rub_bell = 0; } } /* ----------------------------------------------------------------------------- + Called at the beginning of each Retainer Profiliing + -------------------------------------------------------------------------- */ +#ifdef PROFILING +void +stat_startRP(void) +{ + getTimes(); + RP_start_time = CurrentUserTime; + RPe_start_time = CurrentElapsedTime; +} +#endif /* PROFILING */ + +/* ----------------------------------------------------------------------------- + Called at the end of each Retainer Profiliing + -------------------------------------------------------------------------- */ + +#ifdef PROFILING +void +stat_endRP( + nat retainerGeneration, +#ifdef DEBUG_RETAINER + nat maxCStackSize, + int maxStackSize, +#endif + double averageNumVisit) +{ + getTimes(); + RP_tot_time += CurrentUserTime - RP_start_time; + RPe_tot_time += CurrentElapsedTime - RPe_start_time; + + fprintf(prof_file, "Retainer Profiling: %d, at %f seconds\n", + retainerGeneration, mut_user_time_during_RP()); +#ifdef DEBUG_RETAINER + fprintf(prof_file, "\tMax C stack size = %u\n", maxCStackSize); + fprintf(prof_file, "\tMax auxiliary stack size = %u\n", maxStackSize); +#endif + fprintf(prof_file, "\tAverage number of visits per object = %f\n", averageNumVisit); +} +#endif /* PROFILING */ + +/* ----------------------------------------------------------------------------- + Called at the beginning of each heap census + -------------------------------------------------------------------------- */ +#ifdef PROFILING +void +stat_startHeapCensus(void) +{ + getTimes(); + HC_start_time = CurrentUserTime; + HCe_start_time = CurrentElapsedTime; +} +#endif /* PROFILING */ + +/* ----------------------------------------------------------------------------- + Called at the end of each heap census + -------------------------------------------------------------------------- */ +#ifdef PROFILING +void +stat_endHeapCensus(void) +{ + getTimes(); + HC_tot_time += CurrentUserTime - HC_start_time; + HCe_tot_time += CurrentElapsedTime - HCe_start_time; +} +#endif /* PROFILING */ + +/* ----------------------------------------------------------------------------- + stat_workerStop + + Called under SMP when a worker thread finishes. We drop the timing + stats for this thread into the task_ids struct for that thread. + -------------------------------------------------------------------------- */ + +#if defined(SMP) +void +stat_workerStop(void) +{ + nat i; + pthread_t me = pthread_self(); + + getTimes(); + + for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) { + if (task_ids[i].id == me) { + task_ids[i].mut_time = CurrentUserTime - task_ids[i].gc_time; + task_ids[i].mut_etime = CurrentElapsedTime + - GCe_tot_time + - task_ids[i].elapsedtimestart; + if (task_ids[i].mut_time < 0.0) { task_ids[i].mut_time = 0.0; } + if (task_ids[i].mut_etime < 0.0) { task_ids[i].mut_etime = 0.0; } + } + } +} +#endif + +#if defined(SMP) +long int stat_getElapsedTime () +{ + getTimes(); + return CurrentElapsedTime; +} +#endif + +/* ----------------------------------------------------------------------------- Called at the end of execution NOTE: number of allocations is not entirely accurate: it doesn't @@ -316,77 +620,145 @@ stat_endGC(lnat alloc, lnat collect, lnat live, lnat gen) void stat_exit(int alloc) { - FILE *sf = RtsFlags.GcFlags.statsFile; + if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) { - if (sf != NULL){ char temp[BIG_STRING_LEN]; - StgDouble time = usertime(); - StgDouble etime = elapsedtime(); - StgDouble MutTime, MutElapsedTime; + TICK_TYPE time; + TICK_TYPE etime; + nat g, total_collections = 0; + + getTimes(); + time = CurrentUserTime; + etime = CurrentElapsedTime - ElapsedTimeStart; + + GC_tot_alloc += alloc; /* avoid divide by zero if time is measured as 0.00 seconds -- SDM */ - if (time == 0.0) time = 0.0001; - if (etime == 0.0) etime = 0.0001; + if (time == 0.0) time = 1; + if (etime == 0.0) etime = 1; + /* Count total garbage collections */ + for (g = 0; g < RtsFlags.GcFlags.generations; g++) + total_collections += generations[g].collections; + + /* For SMP, we have to get the user time from each thread + * and try to work out the total time. + */ +#ifdef SMP + { nat i; + MutUserTime = 0.0; + for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) { + MutUserTime += task_ids[i].mut_time; + } + } + time = MutUserTime + GC_tot_time + InitUserTime + ExitUserTime; + if (MutUserTime < 0) { MutUserTime = 0; } +#endif - fprintf(sf, "%9ld %9.9s %9.9s", - (lnat)alloc*sizeof(W_), "", ""); - fprintf(sf, " %5.2f %5.2f\n\n", 0.0, 0.0); + if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) { + statsPrintf("%9ld %9.9s %9.9s", (lnat)alloc*sizeof(W_), "", ""); + statsPrintf(" %5.2f %5.2f\n\n", 0.0, 0.0); + } - GC_tot_alloc += alloc; + if (RtsFlags.GcFlags.giveStats >= SUMMARY_GC_STATS) { + ullong_format_string(GC_tot_alloc*sizeof(W_), + temp, rtsTrue/*commas*/); + statsPrintf("%11s bytes allocated in the heap\n", temp); - ullong_format_string(GC_tot_alloc*sizeof(W_), temp, rtsTrue/*commas*/); - fprintf(sf, "%11s bytes allocated in the heap\n", temp); + ullong_format_string(GC_tot_copied*sizeof(W_), + temp, rtsTrue/*commas*/); + statsPrintf("%11s bytes copied during GC\n", temp); - if ( ResidencySamples > 0 ) { - ullong_format_string(MaxResidency*sizeof(W_), temp, rtsTrue/*commas*/); - fprintf(sf, "%11s bytes maximum residency (%ld sample(s))\n", - temp, - ResidencySamples); - } - fprintf(sf,"\n"); - - { /* Count garbage collections */ - nat g; - for (g = 0; g < RtsFlags.GcFlags.generations; g++) { - fprintf(sf, "%11d collections in generation %d\n", - generations[g].collections, g); - } - } - fprintf(sf,"\n%11ld Mb total memory in use\n\n", - mblocks_allocated * MBLOCK_SIZE / (1024 * 1024)); - - MutTime = time - GC_tot_time - InitUserTime; - if (MutTime < 0) { MutTime = 0; } - MutElapsedTime = etime - GCe_tot_time - InitElapsedTime; - if (MutElapsedTime < 0) { MutElapsedTime = 0; } /* sometimes -0.00 */ - - fprintf(sf, " INIT time %6.2fs (%6.2fs elapsed)\n", - InitUserTime, InitElapsedTime); - fprintf(sf, " MUT time %6.2fs (%6.2fs elapsed)\n", - MutTime, MutElapsedTime); - fprintf(sf, " GC time %6.2fs (%6.2fs elapsed)\n", - GC_tot_time, GCe_tot_time); - fprintf(sf, " Total time %6.2fs (%6.2fs elapsed)\n\n", - time, etime); - - fprintf(sf, " %%GC time %5.1f%% (%.1f%% elapsed)\n\n", - GC_tot_time*100./time, GCe_tot_time*100./etime); - - if (time - GC_tot_time == 0.0) - ullong_format_string(0, temp, rtsTrue/*commas*/); - else - ullong_format_string((ullong)(GC_tot_alloc*sizeof(W_)/ - (time - GC_tot_time)), + if ( ResidencySamples > 0 ) { + ullong_format_string(MaxResidency*sizeof(W_), temp, rtsTrue/*commas*/); + statsPrintf("%11s bytes maximum residency (%ld sample(s))\n", + temp, ResidencySamples); + } + statsPrintf("\n"); + + /* Print garbage collections in each gen */ + for (g = 0; g < RtsFlags.GcFlags.generations; g++) { + statsPrintf("%11d collections in generation %d (%6.2fs)\n", + generations[g].collections, g, + TICK_TO_DBL(GC_coll_times[g])); + } + + statsPrintf("\n%11ld Mb total memory in use\n\n", + mblocks_allocated * MBLOCK_SIZE / (1024 * 1024)); + +#ifdef SMP + { + nat i; + for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) { + statsPrintf(" Task %2d: MUT time: %6.2fs (%6.2fs elapsed)\n" + " GC time: %6.2fs (%6.2fs elapsed)\n\n", + i, + TICK_TO_DBL(task_ids[i].mut_time), + TICK_TO_DBL(task_ids[i].mut_etime), + TICK_TO_DBL(task_ids[i].gc_time), + TICK_TO_DBL(task_ids[i].gc_etime)); + } + } +#endif + + statsPrintf(" INIT time %6.2fs (%6.2fs elapsed)\n", + TICK_TO_DBL(InitUserTime), TICK_TO_DBL(InitElapsedTime)); + statsPrintf(" MUT time %6.2fs (%6.2fs elapsed)\n", + TICK_TO_DBL(MutUserTime), TICK_TO_DBL(MutElapsedTime)); + statsPrintf(" GC time %6.2fs (%6.2fs elapsed)\n", + TICK_TO_DBL(GC_tot_time), TICK_TO_DBL(GCe_tot_time)); +#ifdef PROFILING + statsPrintf(" RP time %6.2fs (%6.2fs elapsed)\n", + TICK_TO_DBL(RP_tot_time), TICK_TO_DBL(RPe_tot_time)); + statsPrintf(" PROF time %6.2fs (%6.2fs elapsed)\n", + TICK_TO_DBL(HC_tot_time), TICK_TO_DBL(HCe_tot_time)); +#endif + statsPrintf(" EXIT time %6.2fs (%6.2fs elapsed)\n", + TICK_TO_DBL(ExitUserTime), TICK_TO_DBL(ExitElapsedTime)); + statsPrintf(" Total time %6.2fs (%6.2fs elapsed)\n\n", + TICK_TO_DBL(time), TICK_TO_DBL(etime)); + statsPrintf(" %%GC time %5.1f%% (%.1f%% elapsed)\n\n", + TICK_TO_DBL(GC_tot_time)*100/TICK_TO_DBL(time), + TICK_TO_DBL(GCe_tot_time)*100/TICK_TO_DBL(etime)); + + if (time - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time) == 0) + ullong_format_string(0, temp, rtsTrue/*commas*/); + else + ullong_format_string( + (ullong)((GC_tot_alloc*sizeof(W_))/ + TICK_TO_DBL(time - GC_tot_time - + PROF_VAL(RP_tot_time + HC_tot_time))), + temp, rtsTrue/*commas*/); + + statsPrintf(" Alloc rate %s bytes per MUT second\n\n", temp); + + statsPrintf(" Productivity %5.1f%% of total user, %.1f%% of total elapsed\n\n", + TICK_TO_DBL(time - GC_tot_time - + PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime) * 100 + / TICK_TO_DBL(time), + TICK_TO_DBL(time - GC_tot_time - + PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime) * 100 + / TICK_TO_DBL(etime)); + } - fprintf(sf, " Alloc rate %s bytes per MUT second\n\n", temp); + if (RtsFlags.GcFlags.giveStats == ONELINE_GC_STATS) { + /* print the long long separately to avoid bugginess on mingwin (2001-07-02, mingw-0.5) */ + statsPrintf("<>\n", + total_collections, + ResidencySamples == 0 ? 0 : + AvgResidency*sizeof(W_)/ResidencySamples, + MaxResidency*sizeof(W_), + ResidencySamples, + (unsigned long)(mblocks_allocated * MBLOCK_SIZE / (1024L * 1024L)), + TICK_TO_DBL(InitUserTime), TICK_TO_DBL(InitElapsedTime), + TICK_TO_DBL(MutUserTime), TICK_TO_DBL(MutElapsedTime), + TICK_TO_DBL(GC_tot_time), TICK_TO_DBL(GCe_tot_time)); + } - fprintf(sf, " Productivity %5.1f%% of total user, %.1f%% of total elapsed\n\n", - (time - GC_tot_time - InitUserTime) * 100. / time, - (time - GC_tot_time - InitUserTime) * 100. / etime); - fflush(sf); - fclose(sf); + statsFlush(); + statsClose(); } } @@ -395,40 +767,90 @@ stat_exit(int alloc) Produce some detailed info on the state of the generational GC. -------------------------------------------------------------------------- */ +#ifdef DEBUG void -stat_describe_gens(void) +statDescribeGens(void) { - nat g, s, mut, mut_once, lge, live; - StgMutClosure *m; + nat g, s, mut, lge, live; bdescr *bd; step *step; - fprintf(stderr, " Gen Steps Max Mutable Mut-Once Step Blocks Live Large\n Blocks Closures Closures Objects\n"); + debugBelch(" Gen Steps Max Mutable Step Blocks Live Large\n Blocks Closures Closures Objects\n"); + mut = 0; for (g = 0; g < RtsFlags.GcFlags.generations; g++) { - for (m = generations[g].mut_list, mut = 0; m != END_MUT_LIST; - m = m->mut_link) - mut++; - for (m = generations[g].mut_once_list, mut_once = 0; m != END_MUT_LIST; - m = m->mut_link) - mut_once++; - fprintf(stderr, "%8d %8d %8d %9d %9d", g, generations[g].n_steps, - generations[g].max_blocks, mut, mut_once); + for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) { + mut += bd->free - bd->start; + } + + debugBelch("%8d %8d %8d %9d", g, generations[g].n_steps, + generations[g].max_blocks, mut); for (s = 0; s < generations[g].n_steps; s++) { step = &generations[g].steps[s]; for (bd = step->large_objects, lge = 0; bd; bd = bd->link) lge++; live = 0; - for (bd = step->blocks; bd; bd = bd->link) { + if (RtsFlags.GcFlags.generations == 1) { + bd = step->to_blocks; + } else { + bd = step->blocks; + } + for (; bd; bd = bd->link) { live += (bd->free - bd->start) * sizeof(W_); } if (s != 0) { - fprintf(stderr,"%46s",""); + debugBelch("%46s",""); } - fprintf(stderr,"%6d %8d %8d %8d\n", s, step->n_blocks, + debugBelch("%6d %8d %8d %8d\n", s, step->n_blocks, live, lge); } } - fprintf(stderr,"\n"); + debugBelch("\n"); +} +#endif + +/* ----------------------------------------------------------------------------- + Stats available via a programmatic interface, so eg. GHCi can time + each compilation and expression evaluation. + -------------------------------------------------------------------------- */ + +extern HsInt64 getAllocations( void ) +{ return (HsInt64)total_allocated * sizeof(W_); } + +/* ----------------------------------------------------------------------------- + Dumping stuff in the stats file, or via the debug message interface + -------------------------------------------------------------------------- */ + +static void +statsPrintf( char *s, ... ) +{ + FILE *sf = RtsFlags.GcFlags.statsFile; + va_list ap; + + va_start(ap,s); + if (sf == NULL) { + vdebugBelch(s,ap); + } else { + vfprintf(sf, s, ap); + } + va_end(ap); +} + +static void +statsFlush( void ) +{ + FILE *sf = RtsFlags.GcFlags.statsFile; + if (sf != NULL) { + fflush(sf); + } +} + +static void +statsClose( void ) +{ + FILE *sf = RtsFlags.GcFlags.statsFile; + if (sf != NULL) { + fclose(sf); + } }