diff --git a/src/init.c b/src/init.c
index 0a99c049..07a34f57 100644
--- a/src/init.c
+++ b/src/init.c
@@ -201,7 +201,7 @@ static bool _mi_heap_init(void) {
     tld->segments.stats = &tld->stats;
     tld->segments.os = &tld->os;
     tld->os.stats = &tld->stats;
-    _mi_heap_set_default_direct(heap);
+    _mi_heap_set_default_direct(heap);    
   }
   return false;
 }
@@ -235,9 +235,8 @@ static bool _mi_heap_done(mi_heap_t* heap) {
     _mi_heap_collect_abandon(heap);
   }
   
-
   // merge stats
-  _mi_stats_done(&heap->tld->stats);
+  _mi_stats_done(&heap->tld->stats);  
 
   // free if not the main thread
   if (heap != &_mi_heap_main) {
@@ -337,18 +336,13 @@ void mi_thread_init(void) mi_attr_noexcept
 {
   // ensure our process has started already
   mi_process_init();
-
+  
   // initialize the thread local default heap
   // (this will call `_mi_heap_set_default_direct` and thus set the
   //  fiber/pthread key to a non-zero value, ensuring `_mi_thread_done` is called)
   if (_mi_heap_init()) return;  // returns true if already initialized
 
-  // don't further initialize for the main thread
-  if (_mi_is_main_thread()) return;
-
-  mi_heap_t* const heap = mi_get_default_heap();
-  if (mi_heap_is_initialized(heap)) { _mi_stat_increase(&heap->tld->stats.threads, 1); }
-
+  _mi_stat_increase(&_mi_stats_main.threads, 1);
   //_mi_verbose_message("thread init: 0x%zx\n", _mi_thread_id());
 }
 
@@ -357,14 +351,11 @@ void mi_thread_done(void) mi_attr_noexcept {
 }
 
 static void _mi_thread_done(mi_heap_t* heap) {
+  _mi_stat_decrease(&_mi_stats_main.threads, 1);
+
   // check thread-id as on Windows shutdown with FLS the main (exit) thread may call this on thread-local heaps...
   if (heap->thread_id != _mi_thread_id()) return;
-
-  // stats
-  if (!_mi_is_main_thread() && mi_heap_is_initialized(heap))  {
-    _mi_stat_decrease(&heap->tld->stats.threads, 1);
-  }
-
+  
   // abandon the thread local heap
   if (_mi_heap_done(heap)) return;  // returns true if already ran
 }
diff --git a/src/os.c b/src/os.c
index f41844e2..93e022ac 100644
--- a/src/os.c
+++ b/src/os.c
@@ -100,14 +100,12 @@ static bool use_large_os_page(size_t size, size_t alignment) {
 
 // round to a good OS allocation size (bounded by max 12.5% waste)
 size_t _mi_os_good_alloc_size(size_t size) {
-  size_t align_size = _mi_os_page_size();
-  /*
+  size_t align_size;
   if (size < 512*KiB) align_size = _mi_os_page_size();
   else if (size < 2*MiB) align_size = 64*KiB;
   else if (size < 8*MiB) align_size = 256*KiB;
   else if (size < 32*MiB) align_size = 1*MiB;
   else align_size = 4*MiB;
-  */
   if (size >= (SIZE_MAX - align_size)) return size; // possible overflow?
   return _mi_align_up(size, align_size);
 }
diff --git a/test/test-stress.c b/test/test-stress.c
index 29a9d2e5..05dbd4b4 100644
--- a/test/test-stress.c
+++ b/test/test-stress.c
@@ -272,7 +272,7 @@ static void run_os_threads(size_t nthreads, void (*fun)(intptr_t)) {
   DWORD* tids = (DWORD*)custom_calloc(nthreads,sizeof(DWORD));
   HANDLE* thandles = (HANDLE*)custom_calloc(nthreads,sizeof(HANDLE));
   for (uintptr_t i = 0; i < nthreads; i++) {
-    thandles[i] = CreateThread(0, 4096, &thread_entry, (void*)(i), 0, &tids[i]);
+    thandles[i] = CreateThread(0, 8*1024, &thread_entry, (void*)(i), 0, &tids[i]);
   }
   for (size_t i = 0; i < nthreads; i++) {
     WaitForSingleObject(thandles[i], INFINITE);