Add peak memory usage measurement capability

Add (currently disabled) faciility to try to filter out pool allocation overhead from memory usage stats


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@4753 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 121 insertions, 8 deletions

diff --git a/support/lib/Support/Timer.cpp b/support/lib/Support/Timer.cpp
index 30fa9be..17fea8b 100644
--- a/support/lib/Support/Timer.cpp
+++ b/support/lib/Support/Timer.cpp
@@ -15,6 +15,12 @@
 #include <algorithm>
 #include <functional>
 
+// getNumBytesToNotCount - This function is supposed to return the number of
+// bytes that are to be considered not allocated, even though malloc thinks they
+// are allocated.
+//
+static unsigned getNumBytesToNotCount();
+
 static TimerGroup *DefaultTimerGroup = 0;
 static TimerGroup *getDefaultTimerGroup() {
   if (DefaultTimerGroup) return DefaultTimerGroup;
@@ -22,13 +28,13 @@ static TimerGroup *getDefaultTimerGroup() {
 }
 
 Timer::Timer(const std::string &N)
-  : Elapsed(0), UserTime(0), SystemTime(0), MemUsed(0), Name(N),
+  : Elapsed(0), UserTime(0), SystemTime(0), MemUsed(0), PeakMem(0), Name(N),
     Started(false), TG(getDefaultTimerGroup()) {
   TG->addTimer();
 }
 
 Timer::Timer(const std::string &N, TimerGroup &tg)
-  : Elapsed(0), UserTime(0), SystemTime(0), MemUsed(0), Name(N),
+  : Elapsed(0), UserTime(0), SystemTime(0), MemUsed(0), PeakMem(0), Name(N),
     Started(false), TG(&tg) {
   TG->addTimer();
 }
@@ -58,42 +64,70 @@ Timer::~Timer() {
   }
 }
 
+static long getMemUsage() {
+  struct mallinfo MI = mallinfo();
+  return MI.uordblks/*+MI.hblkhd-getNumBytesToNotCount()*/;
+}
+
 struct TimeRecord {
   double Elapsed, UserTime, SystemTime;
   long MemUsed;
 };
 
-static TimeRecord getTimeRecord() {
+static TimeRecord getTimeRecord(bool Start) {
   struct rusage RU;
   struct timeval T;
+  long MemUsed;
+  if (Start) {
+    MemUsed = getMemUsage();
+    if (getrusage(RUSAGE_SELF, &RU))
+      perror("getrusage call failed: -time-passes info incorrect!");
+  }
   gettimeofday(&T, 0);
-  if (getrusage(RUSAGE_SELF, &RU)) {
-    perror("getrusage call failed: -time-passes info incorrect!");
+
+  if (!Start) {
+    MemUsed = getMemUsage();
+    if (getrusage(RUSAGE_SELF, &RU))
+      perror("getrusage call failed: -time-passes info incorrect!");
   }
 
   TimeRecord Result;
   Result.Elapsed    =           T.tv_sec +           T.tv_usec/1000000.0;
   Result.UserTime   = RU.ru_utime.tv_sec + RU.ru_utime.tv_usec/1000000.0;
   Result.SystemTime = RU.ru_stime.tv_sec + RU.ru_stime.tv_usec/1000000.0;
-  Result.MemUsed    = mallinfo().uordblks;
+  Result.MemUsed = MemUsed;
+
   return Result;
 }
 
+static std::vector<Timer*> ActiveTimers;
+
 void Timer::startTimer() {
   Started = true;
-  TimeRecord TR = getTimeRecord();
+  TimeRecord TR = getTimeRecord(true);
   Elapsed    -= TR.Elapsed;
   UserTime   -= TR.UserTime;
   SystemTime -= TR.SystemTime;
   MemUsed    -= TR.MemUsed;
+  PeakMemBase = TR.MemUsed;
+  ActiveTimers.push_back(this);
 }
 
 void Timer::stopTimer() {
-  TimeRecord TR = getTimeRecord();
+  TimeRecord TR = getTimeRecord(false);
   Elapsed    += TR.Elapsed;
   UserTime   += TR.UserTime;
   SystemTime += TR.SystemTime;
   MemUsed    += TR.MemUsed;
+
+  if (ActiveTimers.back() == this) {
+    ActiveTimers.pop_back();
+  } else {
+    std::vector<Timer*>::iterator I =
+      std::find(ActiveTimers.begin(), ActiveTimers.end(), this);
+    assert(I != ActiveTimers.end() && "stop but no startTimer?");
+    ActiveTimers.erase(I);
+  }
 }
 
 void Timer::sum(const Timer &T) {
@@ -101,8 +135,22 @@ void Timer::sum(const Timer &T) {
   UserTime   += T.UserTime;
   SystemTime += T.SystemTime;
   MemUsed    += T.MemUsed;
+  PeakMem    += T.PeakMem;
 }
 
+/// addPeakMemoryMeasurement - This method should be called whenever memory
+/// usage needs to be checked.  It adds a peak memory measurement to the
+/// currently active timers, which will be printed when the timer group prints
+///
+void Timer::addPeakMemoryMeasurement() {
+  long MemUsed = getMemUsage();
+
+  for (std::vector<Timer*>::iterator I = ActiveTimers.begin(),
+         E = ActiveTimers.end(); I != E; ++I)
+    (*I)->PeakMem = std::max((*I)->PeakMem, MemUsed-(*I)->PeakMemBase);
+}
+
+
 //===----------------------------------------------------------------------===//
 //   TimerGroup Implementation
 //===----------------------------------------------------------------------===//
@@ -127,6 +175,12 @@ void Timer::print(const Timer &Total) {
 
   if (Total.MemUsed)
     fprintf(stderr, " %8ld  ", MemUsed);
+  if (Total.PeakMem) {
+    if (PeakMem)
+      fprintf(stderr, " %8ld  ", PeakMem);
+    else
+      fprintf(stderr, "           ");
+  }
   std::cerr << Name << "\n";
 
   Started = false;  // Once printed, don't print again
@@ -168,6 +222,8 @@ void TimerGroup::removeTimer() {
       std::cerr << "   ---Wall Time---";
       if (Total.getMemUsed())
         std::cerr << "  ---Mem---";
+      if (Total.getPeakMem())
+        std::cerr << "  -PeakMem-";
       std::cerr << "  --- Name ---\n";
       
       // Loop through all of the timing data, printing it out...
@@ -188,3 +244,60 @@ void TimerGroup::removeTimer() {
     DefaultTimerGroup = 0;
   }
 }
+
+
+
+#if (__GNUC__ == 3) && (__GNUC_MINOR__ == 2) && (__GNUC_PATCHLEVEL__ == 0)
+// If we have GCC 3.2.0, we can calculate pool allocation bookkeeping info
+#define HAVE_POOL
+extern "C" {
+  // Taken from GCC 3.2's stl_alloc.h file:
+  enum {_ALIGN = 8, _MAX_BYTES = 128, NFREE = _MAX_BYTES / _ALIGN};
+  struct FreeList { FreeList *Next; };
+
+  FreeList *_ZNSt24__default_alloc_templateILb1ELi0EE12_S_free_listE[NFREE];
+  char *_ZNSt24__default_alloc_templateILb1ELi0EE13_S_start_freeE;
+  char *_ZNSt24__default_alloc_templateILb1ELi0EE11_S_end_freeE;
+  size_t _ZNSt24__default_alloc_templateILb1ELi0EE12_S_heap_sizeE;
+  
+  // Make the symbols possible to use...
+  FreeList* (&TheFreeList)[NFREE] = _ZNSt24__default_alloc_templateILb1ELi0EE12_S_free_listE;
+  char * &StartFree = _ZNSt24__default_alloc_templateILb1ELi0EE13_S_start_freeE;
+  char * &EndFree   = _ZNSt24__default_alloc_templateILb1ELi0EE11_S_end_freeE;
+  size_t &HeapSize  = _ZNSt24__default_alloc_templateILb1ELi0EE12_S_heap_sizeE;
+}
+#endif
+
+// getNumBytesToNotCount - This function is supposed to return the number of
+// bytes that are to be considered not allocated, even though malloc thinks they
+// are allocated.
+//
+static unsigned getNumBytesToNotCount() {
+#ifdef HAVE_POOL
+  // If we have GCC 3.2.0, we can subtract off pool allocation bookkeeping info
+
+  // Size of the free slab section... 
+  unsigned FreePoolMem = (unsigned)(EndFree-StartFree);
+
+  // Walk all of the free lists, adding memory to the free counter whenever we
+  // have a free bucket.
+  for (unsigned i = 0; i != NFREE; ++i) {
+    unsigned NumEntries = 0;
+    for (FreeList *FL = TheFreeList[i]; FL; ++NumEntries, FL = FL->Next)
+      /*empty*/ ;
+    
+#if 0
+    if (NumEntries)
+      std::cerr << "  For Size[" << (i+1)*_ALIGN << "]: " << NumEntries
+                << " Free entries\n";
+#endif
+    FreePoolMem += NumEntries*(i+1)*_ALIGN;
+  }
+  return FreePoolMem;
+  
+#else
+#warning "Don't know how to avoid pool allocation accounting overhead for this"
+#warning " compiler: Space usage numbers (with -time-passes) may be off!"
+  return 0;
+#endif
+}