if err != nil {
			log.Fatalf("%v", err)
		}
		AtExit(func() {
			// Profile all outstanding allocations.
			runtime.GC()
			// compilebench parses the memory profile to extract memstats,
			// which are only written in the legacy pprof format.
			// See golang.org/issue/18641 and runtime/pprof/pprof.go:writeHeap.
			const writeLegacyFormat = 1

				nbytes := int64(npages * pageSize)
				gcController.heapReleased.add(nbytes)
				gcController.heapFree.add(-nbytes)

				stats := memstats.heapStats.acquire()
				atomic.Xaddint64(&stats.committed, -nbytes)
				atomic.Xaddint64(&stats.released, nbytes)
				memstats.heapStats.release()
			}

			// Relock the heap, because now we need to make these pages

	}
	if stackDebug >= 1 {
		print("stackalloc ", n, "\n")
	}

	if debug.efence != 0 || stackFromSystem != 0 {
		n = uint32(alignUp(uintptr(n), physPageSize))
		v := sysAlloc(uintptr(n), &memstats.stacks_sys)
		if v == nil {
			throw("out of memory (stackalloc)")
		}
		return stack{uintptr(v), uintptr(v) + uintptr(n)}
	}

	// Small stacks are allocated with a fixed-size free-list allocator.

	})
}

func TestTimerGC(t *testing.T) {
	run := func(t *testing.T, what string, f func()) {
		t.Helper()
		t.Run(what, func(t *testing.T) {
			t.Helper()
			const N = 1e4
			var stats runtime.MemStats
			runtime.GC()
			runtime.GC()
			runtime.GC()
			runtime.ReadMemStats(&stats)
			before := int64(stats.Mallocs - stats.Frees)

			for j := 0; j < N; j++ {
				f()
			}

			runtime.GC()

		throw("queuefinalizer during GC")
	}

	lock(&finlock)
	if finq == nil || finq.cnt == uint32(len(finq.fin)) {
		if finc == nil {
			finc = (*finblock)(persistentalloc(_FinBlockSize, 0, &memstats.gcMiscSys))
			finc.alllink = allfin
			allfin = finc
			if finptrmask[0] == 0 {
				// Build pointer mask for Finalizer array in block.
				// Check assumptions made in finalizer1 array above.

		n := pollBlockSize / pdSize
		if n == 0 {
			n = 1
		}
		// Must be in non-GC memory because can be referenced
		// only from epoll/kqueue internals.
		mem := persistentalloc(n*pdSize, 0, &memstats.other_sys)
		for i := uintptr(0); i < n; i++ {
			pd := (*pollDesc)(add(mem, i*pdSize))
			lockInit(&pd.lock, lockRankPollDesc)
			pd.rt.init(nil, nil)
			pd.wt.init(nil, nil)
			pd.link = c.first

		unlock(&itabLock)
		goto finish
	}

	// Entry doesn't exist yet. Make a new entry & add it.
	m = (*itab)(persistentalloc(unsafe.Sizeof(itab{})+uintptr(len(inter.Methods)-1)*goarch.PtrSize, 0, &memstats.other_sys))
	m.Inter = inter
	m.Type = typ
	// The hash is used in type switches. However, compiler statically generates itab's
	// for all interface/type pairs used in switches (which are added to itabTable

	case memProfile:
		size += unsafe.Sizeof(memRecord{})
	case blockProfile, mutexProfile:
		size += unsafe.Sizeof(blockRecord{})
	}

	b := (*bucket)(persistentalloc(size, 0, &memstats.buckhash_sys))
	b.typ = typ
	b.nstk = uintptr(nstk)
	return b
}

// stk returns the slice in b holding the stack. The caller can asssume that the
// backing array is immutable.

				throw("trace: reading after shutdown")
			}
			// Free all the empty buffers.
			for trace.empty != nil {
				buf := trace.empty
				trace.empty = buf.link
				sysFree(unsafe.Pointer(buf), unsafe.Sizeof(*buf), &memstats.other_sys)
			}
			// Clear trace.shutdown and other flags.
			trace.headerWritten = false
			trace.shutdown.Store(false)
		}
		unlock(&trace.lock)
	})

	if stopTrace {

std::optional<AllocatorStats> XlaDeviceAllocator::GetStats() {
  std::optional<stream_executor::AllocatorStats> se_stats =
      stream_executor_->GetAllocatorStats();
  if (!se_stats) {
    return std::nullopt;
  }

  tensorflow::AllocatorStats tf_stats;
  tf_stats.num_allocs = se_stats->num_allocs;
  tf_stats.bytes_in_use = se_stats->bytes_in_use;
  tf_stats.peak_bytes_in_use = se_stats->peak_bytes_in_use;