rts: Do not traverse nursery for dead closures in LDV profile
authorMatthew Pickering <matthewtpickering@gmail.com>
Thu, 6 Jun 2019 13:07:06 +0000 (14:07 +0100)
committerMarge Bot <ben+marge-bot@smart-cactus.org>
Thu, 27 Jun 2019 14:42:29 +0000 (10:42 -0400)
It is important that `heapCensus` and `LdvCensusForDead` traverse the
same areas.

`heapCensus` increases the `not_used` counter which tracks how many
closures are live but haven't been used yet.

`LdvCensusForDead` increases the `void_total` counter which tracks how
many dead closures there are.

The `LAG` is then calculated by substracting the `void_total` from
`not_used` and so it is essential that `not_used >= void_total`. This
fact is checked by quite a few assertions.

However, if a program has low maximum residency but allocates a lot in
the nursery then these assertions were failing (see #16753 and #15903)
because `LdvCensusForDead` was observing dead closures from the nursery
which totalled more than the `not_used`. The same closures were not
counted by `heapCensus`.

Therefore, it seems that the correct fix is to make `LdvCensusForDead`
agree with `heapCensus` and not traverse the nursery for dead closures.

Fixes #16100 #16753 #15903 #8982

rts/LdvProfile.c

index 608961e..c918be5 100644 (file)
@@ -173,28 +173,6 @@ processHeapForDead( bdescr *bd )
 }
 
 /* --------------------------------------------------------------------------
- * Calls processHeapClosureForDead() on every *dead* closures in the nursery.
- * ----------------------------------------------------------------------- */
-static void
-processNurseryForDead( void )
-{
-    StgPtr p;
-    bdescr *bd;
-
-    if (MainCapability.r.rNursery == NULL)
-        return;
-
-    for (bd = MainCapability.r.rNursery->blocks; bd != NULL; bd = bd->link) {
-        p = bd->start;
-        while (p < bd->free) {
-            while (p < bd->free && !*p) p++; // skip slop
-            if (p >= bd->free) break;
-            p += processHeapClosureForDead((StgClosure *)p);
-        }
-    }
-}
-
-/* --------------------------------------------------------------------------
  * Calls processHeapClosureForDead() on every *dead* closures in the closure
  * chain.
  * ----------------------------------------------------------------------- */
@@ -233,7 +211,6 @@ LdvCensusForDead( uint32_t N )
         //
         barf("Lag/Drag/Void profiling not supported with -G1");
     } else {
-        processNurseryForDead();
         for (g = 0; g <= N; g++) {
             processHeapForDead(generations[g].old_blocks);
             processChainForDead(generations[g].large_objects);