/* -----------------------------------------------------------------------------
- * $Id: Profiling.c,v 1.17 2000/04/03 15:54:49 simonmar Exp $
*
* (c) The GHC Team, 1998-2000
*
#ifdef PROFILING
+#include "PosixSource.h"
#include "Rts.h"
#include "RtsUtils.h"
#include "RtsFlags.h"
#include "Profiling.h"
#include "Storage.h"
#include "Proftimer.h"
-#include "Itimer.h"
+#include "Timer.h"
#include "ProfHeap.h"
+#include "Arena.h"
+#include "RetainerProfile.h"
+#include "LdvProfile.h"
+
+#include <string.h>
+
+/*
+ * Profiling allocation arena.
+ */
+Arena *prof_arena;
/*
* Global variables used to assign unique IDs to cc's, ccs's, and
unsigned int CCS_ID;
unsigned int HP_ID;
-/* Table sizes from old profiling system. Not sure if we'll need
- * these.
- */
-nat time_intervals = 0;
-nat earlier_ticks = 0;
-nat max_cc_no = 0;
-nat max_mod_no = 0;
-nat max_grp_no = 0;
-nat max_descr_no = 0;
-nat max_type_no = 0;
-
-/* Are we time-profiling?
- */
-rtsBool time_profiling = rtsFalse;
-
/* figures for the profiling report.
*/
-static lnat total_alloc, total_prof_ticks;
+static ullong total_alloc;
+static lnat total_prof_ticks;
-/* Globals for opening the profiling log file
+/* Globals for opening the profiling log file(s)
*/
static char *prof_filename; /* prof report file name = <program>.prof */
FILE *prof_file;
+static char *hp_filename; /* heap profile (hp2ps style) log file */
+FILE *hp_file;
+
/* The Current Cost Centre Stack (for attributing costs)
*/
CostCentreStack *CCCS;
* Static Functions
*/
-static CostCentreStack * ActualPush_ ( CostCentreStack *ccs, CostCentre *cc,
- CostCentreStack *new_ccs );
-
-static rtsBool ccs_to_ignore ( CostCentreStack *ccs );
-static void count_ticks ( CostCentreStack *ccs );
-static void reportCCS ( CostCentreStack *ccs, nat indent );
-static void DecCCS ( CostCentreStack *ccs );
-static void DecBackEdge ( CostCentreStack *ccs, CostCentreStack *oldccs );
-static CostCentreStack *CheckLoop ( CostCentreStack *ccs, CostCentre *cc );
-static CostCentreStack *pruneCCSTree ( CostCentreStack *ccs );
-
-static CostCentreStack *ActualPush ( CostCentreStack *, CostCentre * );
-static CostCentreStack *IsInIndexTable ( IndexTable *, CostCentre * );
-static IndexTable *AddToIndexTable ( IndexTable *, CostCentreStack *,
- CostCentre *, unsigned int );
-
-#ifdef DEBUG
-static void printCCS ( CostCentreStack *ccs );
-#endif
-static void initTimeProfiling ( void );
-static void initProfilingLogFile( void );
-
-static void reportCCS_XML ( CostCentreStack *ccs );
+static CostCentreStack * ActualPush_ ( CostCentreStack *ccs, CostCentre *cc,
+ CostCentreStack *new_ccs );
+static rtsBool ccs_to_ignore ( CostCentreStack *ccs );
+static void count_ticks ( CostCentreStack *ccs );
+static void inherit_costs ( CostCentreStack *ccs );
+static void reportCCS ( CostCentreStack *ccs, nat indent );
+static void DecCCS ( CostCentreStack *ccs );
+static void DecBackEdge ( CostCentreStack *ccs,
+ CostCentreStack *oldccs );
+static CostCentreStack * CheckLoop ( CostCentreStack *ccs, CostCentre *cc );
+static CostCentreStack * pruneCCSTree ( CostCentreStack *ccs );
+static CostCentreStack * ActualPush ( CostCentreStack *, CostCentre * );
+static CostCentreStack * IsInIndexTable ( IndexTable *, CostCentre * );
+static IndexTable * AddToIndexTable ( IndexTable *, CostCentreStack *,
+ CostCentre *, unsigned int );
+static void ccsSetSelected ( CostCentreStack *ccs );
+
+static void initTimeProfiling ( void );
+static void initProfilingLogFile( void );
+
+static void reportCCS_XML ( CostCentreStack *ccs );
/* -----------------------------------------------------------------------------
Initialise the profiling environment
void
initProfiling1 (void)
{
+ // initialise our arena
+ prof_arena = newArena();
+
/* for the benefit of allocate()... */
CCCS = CCS_SYSTEM;
-
+
/* Initialize counters for IDs */
CC_ID = 1;
CCS_ID = 1;
*/
ASSERT(CCS_MAIN->prevStack == 0);
CCS_MAIN->root = CC_MAIN;
+ ccsSetSelected(CCS_MAIN);
DecCCS(CCS_MAIN);
+
for (ccs = CCS_LIST; ccs != CCS_MAIN; ) {
next = ccs->prevStack;
ccs->prevStack = 0;
initHeapProfiling();
}
}
-
-static void
-initProfilingLogFile(void)
+
+// Decide whether closures with this CCS should contribute to the heap
+// profile.
+static void
+ccsSetSelected( CostCentreStack *ccs )
{
- /* Initialise the log file name */
- prof_filename = stgMallocBytes(strlen(prog_argv[0]) + 6, "initProfiling");
- sprintf(prof_filename, "%s.prof", prog_argv[0]);
+ if (RtsFlags.ProfFlags.modSelector) {
+ if (! strMatchesSelector( ccs->cc->module,
+ RtsFlags.ProfFlags.modSelector ) ) {
+ ccs->selected = 0;
+ return;
+ }
+ }
+ if (RtsFlags.ProfFlags.ccSelector) {
+ if (! strMatchesSelector( ccs->cc->label,
+ RtsFlags.ProfFlags.ccSelector ) ) {
+ ccs->selected = 0;
+ return;
+ }
+ }
+ if (RtsFlags.ProfFlags.ccsSelector) {
+ CostCentreStack *c;
+ for (c = ccs; c != NULL; c = c->prevStack) {
+ if ( strMatchesSelector( c->cc->label,
+ RtsFlags.ProfFlags.ccsSelector )) {
+ break;
+ }
+ }
+ if (c == NULL) {
+ ccs->selected = 0;
+ return;
+ }
+ }
- /* open the log file */
- if ((prof_file = fopen(prof_filename, "w")) == NULL) {
- fprintf(stderr, "Can't open profiling report file %s\n", prof_filename);
- RtsFlags.CcFlags.doCostCentres = 0;
+ ccs->selected = 1;
return;
- }
+}
+
+
+static void
+initProfilingLogFile(void)
+{
+ /* Initialise the log file name */
+ prof_filename = arenaAlloc(prof_arena, strlen(prog_name) + 6);
+ sprintf(prof_filename, "%s.prof", prog_name);
+
+ /* open the log file */
+ if ((prof_file = fopen(prof_filename, "w")) == NULL) {
+ debugBelch("Can't open profiling report file %s\n", prof_filename);
+ RtsFlags.CcFlags.doCostCentres = 0;
+ // The following line was added by Sung; retainer/LDV profiling may need
+ // two output files, i.e., <program>.prof/hp.
+ if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER)
+ RtsFlags.ProfFlags.doHeapProfile = 0;
+ return;
+ }
- if (RtsFlags.CcFlags.doCostCentres == COST_CENTRES_XML) {
- /* dump the time, and the profiling interval */
- fprintf(prof_file, "\"%s\"\n", time_str());
- fprintf(prof_file, "\"%d ms\"\n", TICK_MILLISECS);
+ if (RtsFlags.CcFlags.doCostCentres == COST_CENTRES_XML) {
+ /* dump the time, and the profiling interval */
+ fprintf(prof_file, "\"%s\"\n", time_str());
+ fprintf(prof_file, "\"%d ms\"\n", TICK_MILLISECS);
+
+ /* declare all the cost centres */
+ {
+ CostCentre *cc;
+ for (cc = CC_LIST; cc != NULL; cc = cc->link) {
+ fprintf(prof_file, "%d %d \"%s\" \"%s\"\n",
+ CC_UQ, cc->ccID, cc->label, cc->module);
+ }
+ }
+ }
- /* declare all the cost centres */
- {
- CostCentre *cc;
- for (cc = CC_LIST; cc != NULL; cc = cc->link) {
- fprintf(prof_file, "%d %d \"%s\" \"%s\"\n",
- CC_UQ, cc->ccID, cc->label, cc->module);
- }
+ if (RtsFlags.ProfFlags.doHeapProfile) {
+ /* Initialise the log file name */
+ hp_filename = arenaAlloc(prof_arena, strlen(prog_name) + 6);
+ sprintf(hp_filename, "%s.hp", prog_name);
+
+ /* open the log file */
+ if ((hp_file = fopen(hp_filename, "w")) == NULL) {
+ debugBelch("Can't open profiling report file %s\n",
+ hp_filename);
+ RtsFlags.ProfFlags.doHeapProfile = 0;
+ return;
+ }
}
- }
}
void
initTimeProfiling(void)
{
- time_profiling = rtsTrue;
-
/* Start ticking */
startProfTimer();
};
-------------------------------------------------------------------------- */
rtsBool entering_PAP;
-CostCentreStack *
-EnterFunCCS ( CostCentreStack *cccs, CostCentreStack *ccsfn )
+void
+EnterFunCCS ( CostCentreStack *ccsfn )
{
/* PAP_entry has already set CCCS for us */
if (entering_PAP) {
entering_PAP = rtsFalse;
- return CCCS;
+ return;
}
if (ccsfn->root->is_caf == CC_IS_CAF) {
- return AppendCCS(cccs,ccsfn);
+ CCCS = AppendCCS(CCCS,ccsfn);
} else {
- return ccsfn;
+ CCCS = ccsfn;
}
}
#define PushCostCentre _PushCostCentre
{
IF_DEBUG(prof,
- fprintf(stderr,"Pushing %s on ", cc->label);
- printCCS(ccs);
- fprintf(stderr,"\n"));
+ debugBelch("Pushing %s on ", cc->label);
+ debugCCS(ccs);
+ debugBelch("\n"));
return PushCostCentre(ccs,cc);
}
#endif
{
IF_DEBUG(prof,
if (ccs1 != ccs2) {
- fprintf(stderr,"Appending ");
- printCCS(ccs1);
- fprintf(stderr," to ");
- printCCS(ccs2);
- fprintf(stderr,"\n");});
+ debugBelch("Appending ");
+ debugCCS(ccs1);
+ debugBelch(" to ");
+ debugCCS(ccs2);
+ debugBelch("\n");});
return AppendCCS(ccs1,ccs2);
}
#endif
CostCentreStack *new_ccs;
/* allocate space for a new CostCentreStack */
- new_ccs = (CostCentreStack *) stgMallocBytes(sizeof(CostCentreStack), "Error allocating space for CostCentreStack");
+ new_ccs = (CostCentreStack *) arenaAlloc(prof_arena, sizeof(CostCentreStack));
return ActualPush_(ccs, cc, new_ccs);
}
ActualPush_ ( CostCentreStack *ccs, CostCentre *cc, CostCentreStack *new_ccs )
{
/* assign values to each member of the structure */
- ASSIGN_CCS_ID(new_ccs->ccsID);
-
+ new_ccs->ccsID = CCS_ID++;
new_ccs->cc = cc;
new_ccs->prevStack = ccs;
/* Initialise the various _scc_ counters to zero
*/
new_ccs->scc_count = 0;
- new_ccs->sub_scc_count = 0;
- new_ccs->sub_cafcc_count = 0;
/* Initialize all other stats here. There should be a quick way
* that's easily used elsewhere too
*/
new_ccs->time_ticks = 0;
new_ccs->mem_alloc = 0;
+ new_ccs->inherited_ticks = 0;
+ new_ccs->inherited_alloc = 0;
new_ccs->root = ccs->root;
+ // Set the selected field.
+ ccsSetSelected(new_ccs);
+
/* update the memoization table for the parent stack */
if (ccs != EMPTY_STACK)
ccs->indexTable = AddToIndexTable(ccs->indexTable, new_ccs, cc,
{
IndexTable *new_it;
- new_it = stgMallocBytes(sizeof(IndexTable), "AddToIndexTable");
+ new_it = arenaAlloc(prof_arena, sizeof(IndexTable));
new_it->cc = cc;
new_it->ccs = new_ccs;
}
-void
-print_ccs (FILE *fp, CostCentreStack *ccs)
-{
- if (ccs == CCCS) {
- fprintf(fp, "Cost-Centre Stack: ");
- }
-
- if (ccs != CCS_MAIN)
- {
- print_ccs(fp, ccs->prevStack);
- fprintf(fp, "->[%s,%s]", ccs->cc->label, ccs->cc->module);
- } else {
- fprintf(fp, "[%s,%s]", ccs->cc->label, ccs->cc->module);
- }
-
- if (ccs == CCCS) {
- fprintf(fp, "\n");
- }
-}
-
-
static void
DecCCS(CostCentreStack *ccs)
{
Generating a time & allocation profiling report.
-------------------------------------------------------------------------- */
+/* We omit certain system-related CCs and CCSs from the default
+ * reports, so as not to cause confusion.
+ */
+static rtsBool
+cc_to_ignore (CostCentre *cc)
+{
+ if ( cc == CC_OVERHEAD
+ || cc == CC_DONT_CARE
+ || cc == CC_GC
+ || cc == CC_SYSTEM) {
+ return rtsTrue;
+ } else {
+ return rtsFalse;
+ }
+}
+
+static rtsBool
+ccs_to_ignore (CostCentreStack *ccs)
+{
+ if ( ccs == CCS_OVERHEAD
+ || ccs == CCS_DONT_CARE
+ || ccs == CCS_GC
+ || ccs == CCS_SYSTEM) {
+ return rtsTrue;
+ } else {
+ return rtsFalse;
+ }
+}
+
/* -----------------------------------------------------------------------------
Generating the aggregated per-cost-centre time/alloc report.
-------------------------------------------------------------------------- */
for (cc = CC_LIST; cc != NULL; cc = next) {
next = cc->link;
if (cc->time_ticks > total_prof_ticks/100
- || cc->mem_alloc > total_alloc/100) {
+ || cc->mem_alloc > total_alloc/100
+ || RtsFlags.CcFlags.doCostCentres >= COST_CENTRES_ALL) {
insert_cc_in_sorted_list(cc);
}
}
- fprintf(prof_file, "%-20s %-10s", "COST CENTRE", "MODULE");
+ fprintf(prof_file, "%-30s %-20s", "COST CENTRE", "MODULE");
fprintf(prof_file, "%6s %6s", "%time", "%alloc");
if (RtsFlags.CcFlags.doCostCentres >= COST_CENTRES_VERBOSE) {
fprintf(prof_file, " %5s %9s", "ticks", "bytes");
fprintf(prof_file, "\n\n");
for (cc = sorted_cc_list; cc != NULL; cc = cc->link) {
- fprintf(prof_file, "%-20s %-10s", cc->label, cc->module);
- fprintf(prof_file, "%6.1f %6.1f",
- total_prof_ticks == 0 ? 0.0 : (cc->time_ticks / (StgFloat) total_prof_ticks * 100),
- total_alloc == 0 ? 0.0 : (cc->mem_alloc / (StgFloat)
- total_alloc * 100)
- );
-
- if (RtsFlags.CcFlags.doCostCentres >= COST_CENTRES_VERBOSE) {
- fprintf(prof_file, " %5ld %9ld", cc->time_ticks, cc->mem_alloc);
- }
- fprintf(prof_file, "\n");
+ if (cc_to_ignore(cc)) {
+ continue;
+ }
+ fprintf(prof_file, "%-30s %-20s", cc->label, cc->module);
+ fprintf(prof_file, "%6.1f %6.1f",
+ total_prof_ticks == 0 ? 0.0 : (cc->time_ticks / (StgFloat) total_prof_ticks * 100),
+ total_alloc == 0 ? 0.0 : (cc->mem_alloc / (StgFloat)
+ total_alloc * 100)
+ );
+
+ if (RtsFlags.CcFlags.doCostCentres >= COST_CENTRES_VERBOSE) {
+ fprintf(prof_file, " %5llu %9llu", (StgWord64)(cc->time_ticks), cc->mem_alloc);
+ }
+ fprintf(prof_file, "\n");
}
fprintf(prof_file,"\n\n");
static void
fprint_header( void )
{
- fprintf(prof_file, "%-24s %-10s", "COST CENTRE", "MODULE");
+ fprintf(prof_file, "%-24s %-10s individual inherited\n", "", "");
- fprintf(prof_file, "%8s %5s %5s %8s %5s", "scc", "%time", "%alloc", "inner", "cafs");
+ fprintf(prof_file, "%-24s %-50s", "COST CENTRE", "MODULE");
+ fprintf(prof_file, "%6s %10s %5s %5s %5s %5s", "no.", "entries", "%time", "%alloc", "%time", "%alloc");
if (RtsFlags.CcFlags.doCostCentres >= COST_CENTRES_VERBOSE) {
fprintf(prof_file, " %5s %9s", "ticks", "bytes");
}
void
-report_ccs_profiling( void )
+reportCCSProfiling( void )
{
nat count;
char temp[128]; /* sigh: magic constant */
gen_XML_logfile();
return;
default:
+ break;
}
fprintf(prof_file, "\t%s Time and Allocation Profiling Report (%s)\n",
time_str(), "Final");
fprintf(prof_file, "\n\t ");
- fprintf(prof_file, " %s", prog_argv[0]);
+ fprintf(prof_file, " %s", prog_name);
fprintf(prof_file, " +RTS");
for (count = 0; rts_argv[count]; count++)
fprintf(prof_file, " %s", rts_argv[count]);
total_prof_ticks, TICK_MILLISECS);
fprintf(prof_file, "\ttotal alloc = %11s bytes",
- ullong_format_string((ullong) total_alloc * sizeof(W_),
+ ullong_format_string(total_alloc * sizeof(W_),
temp, rtsTrue/*commas*/));
- /* ToDo: 64-bit error! */
#if defined(PROFILING_DETAIL_COUNTS)
fprintf(prof_file, " (%lu closures)", total_allocs);
report_per_cc_costs();
+ inherit_costs(CCS_MAIN);
+
fprint_header();
reportCCS(pruneCCSTree(CCS_MAIN), 0);
-
- fclose(prof_file);
}
static void
/* force printing of *all* cost centres if -P -P */
{
- fprintf(prof_file, "%-*s%-*s %-10s",
+ fprintf(prof_file, "%-*s%-*s %-50s",
indent, "", 24-indent, cc->label, cc->module);
- fprintf(prof_file, "%8ld %5.1f %5.1f %8ld %5ld",
- ccs->scc_count,
- total_prof_ticks == 0 ? 0.0 : (ccs->time_ticks / (StgFloat) total_prof_ticks * 100),
- total_alloc == 0 ? 0.0 : (ccs->mem_alloc / (StgFloat) total_alloc * 100),
- ccs->sub_scc_count, ccs->sub_cafcc_count);
-
+ fprintf(prof_file, "%6d %11.0f %5.1f %5.1f %5.1f %5.1f",
+ ccs->ccsID, (double) ccs->scc_count,
+ total_prof_ticks == 0 ? 0.0 : ((double)ccs->time_ticks / (double)total_prof_ticks * 100.0),
+ total_alloc == 0 ? 0.0 : ((double)ccs->mem_alloc / (double)total_alloc * 100.0),
+ total_prof_ticks == 0 ? 0.0 : ((double)ccs->inherited_ticks / (double)total_prof_ticks * 100.0),
+ total_alloc == 0 ? 0.0 : ((double)ccs->inherited_alloc / (double)total_alloc * 100.0)
+ );
+
if (RtsFlags.CcFlags.doCostCentres >= COST_CENTRES_VERBOSE) {
- fprintf(prof_file, " %5ld %9ld", ccs->time_ticks, ccs->mem_alloc*sizeof(W_));
+ fprintf(prof_file, " %5llu %9llu", (StgWord64)(ccs->time_ticks), ccs->mem_alloc*sizeof(W_));
#if defined(PROFILING_DETAIL_COUNTS)
fprintf(prof_file, " %8ld %8ld %8ld %8ld %8ld %8ld %8ld",
ccs->mem_allocs, ccs->thunk_count,
}
}
+
/* Traverse the cost centre stack tree and accumulate
* ticks/allocations.
*/
}
}
-/* return rtsTrue if it is one of the ones that
- * should not be reported normally (because it confuses
- * the users)
+/* Traverse the cost centre stack tree and inherit ticks & allocs.
*/
-static rtsBool
-ccs_to_ignore (CostCentreStack *ccs)
+static void
+inherit_costs(CostCentreStack *ccs)
{
- if ( ccs == CCS_OVERHEAD
- || ccs == CCS_DONT_CARE
- || ccs == CCS_GC
- || ccs == CCS_SYSTEM) {
- return rtsTrue;
- } else {
- return rtsFalse;
- }
+ IndexTable *i;
+
+ if (ccs_to_ignore(ccs)) { return; }
+
+ ccs->inherited_ticks += ccs->time_ticks;
+ ccs->inherited_alloc += ccs->mem_alloc;
+
+ for (i = ccs->indexTable; i != NULL; i = i->next)
+ if (!i->back_edge) {
+ inherit_costs(i->ccs);
+ ccs->inherited_ticks += i->ccs->inherited_ticks;
+ ccs->inherited_alloc += i->ccs->inherited_alloc;
+ }
+
+ return;
}
static CostCentreStack *
/* force printing of *all* cost centres if -P -P */ )
|| ( ccs->indexTable != 0 )
- || ( (ccs->scc_count || ccs->sub_scc_count ||
- ccs->time_ticks || ccs->mem_alloc
- || (RtsFlags.CcFlags.doCostCentres >= COST_CENTRES_VERBOSE
- && (ccs->sub_cafcc_count
-#if defined(PROFILING_DETAIL_COUNTS)
- || cc->thunk_count || cc->function_count || cc->pap_count
-#endif
- ))))) {
- return ccs;
+ || ( ccs->scc_count || ccs->time_ticks || ccs->mem_alloc )
+ ) {
+ return ccs;
} else {
- return NULL;
+ return NULL;
}
}
CostCentre *cc;
IndexTable *i;
+ if (ccs_to_ignore(ccs)) { return; }
+
cc = ccs->cc;
- fprintf(prof_file, " 1 %d %lu %lu %lu",
- ccs->ccsID, ccs->scc_count, ccs->time_ticks, ccs->mem_alloc);
+ fprintf(prof_file, " 1 %d %llu %llu %llu",
+ ccs->ccsID, ccs->scc_count, (StgWord64)(ccs->time_ticks), ccs->mem_alloc);
for (i = ccs->indexTable; i != 0; i = i->next) {
if (!i->back_edge) {
}
}
+void
+fprintCCS( FILE *f, CostCentreStack *ccs )
+{
+ fprintf(f,"<");
+ for (; ccs && ccs != CCS_MAIN; ccs = ccs->prevStack ) {
+ fprintf(f,"%s.%s", ccs->cc->module, ccs->cc->label);
+ if (ccs->prevStack && ccs->prevStack != CCS_MAIN) {
+ fprintf(f,",");
+ }
+ }
+ fprintf(f,">");
+}
+
+/* For calling from .cmm code, where we can't reliably refer to stderr */
+void
+fprintCCS_stderr( CostCentreStack *ccs )
+{
+ fprintCCS(stderr, ccs);
+}
+
#ifdef DEBUG
-static void
-printCCS ( CostCentreStack *ccs )
+void
+debugCCS( CostCentreStack *ccs )
{
- fprintf(stderr,"<");
- for (; ccs; ccs = ccs->prevStack ) {
- fprintf(stderr,ccs->cc->label);
- if (ccs->prevStack) {
- fprintf(stderr,",");
- }
+ debugBelch("<");
+ for (; ccs && ccs != CCS_MAIN; ccs = ccs->prevStack ) {
+ debugBelch("%s.%s", ccs->cc->module, ccs->cc->label);
+ if (ccs->prevStack && ccs->prevStack != CCS_MAIN) {
+ debugBelch(",");
+ }
}
- fprintf(stderr,">");
+ debugBelch(">");
}
-#endif
+#endif /* DEBUG */
#endif /* PROFILING */