diff --git a/include/mimalloc-internal.h b/include/mimalloc-internal.h
index e996dfd0..a7f2f66d 100644
--- a/include/mimalloc-internal.h
+++ b/include/mimalloc-internal.h
@@ -90,7 +90,7 @@ void*       _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t size) mi_at
 void*       _mi_heap_malloc_zero(mi_heap_t* heap, size_t size, bool zero);
 void*       _mi_heap_realloc_zero(mi_heap_t* heap, void* p, size_t newsize, bool zero);
 mi_block_t* _mi_page_ptr_unalign(const mi_segment_t* segment, const mi_page_t* page, const void* p);
-void        _mi_free_delayed_block(mi_block_t* block);
+void        _mi_free_delayed_block(mi_page_t* page, mi_block_t* block);
 
 #if MI_DEBUG>1
 bool        _mi_page_is_valid(mi_page_t* page);
@@ -229,6 +229,23 @@ static inline mi_page_t* _mi_ptr_page(void* p) {
   return _mi_segment_page_of(_mi_ptr_segment(p), p);
 }
 
+// Thread free access
+static inline mi_block_t* mi_tf_block(mi_thread_free_t tf) {
+  return (mi_block_t*)(tf & ~0x03);
+}
+static inline mi_delayed_t mi_tf_delayed(mi_thread_free_t tf) {
+  return (mi_delayed_t)(tf & 0x03);
+}
+static inline mi_thread_free_t mi_tf_make(mi_block_t* block, mi_delayed_t delayed) {
+  return (mi_thread_free_t)((uintptr_t)block | (uintptr_t)delayed);
+}
+static inline mi_thread_free_t mi_tf_set_delayed(mi_thread_free_t tf, mi_delayed_t delayed) {
+  return mi_tf_make(mi_tf_block(tf),delayed);
+}
+static inline mi_thread_free_t mi_tf_set_block(mi_thread_free_t tf, mi_block_t* block) {
+  return mi_tf_make(block, mi_tf_delayed(tf));
+}
+
 // are all blocks in a page freed?
 static inline bool mi_page_all_free(const mi_page_t* page) {
   mi_assert_internal(page != NULL);
@@ -243,7 +260,7 @@ static inline bool mi_page_immediate_available(const mi_page_t* page) {
 // are there free blocks in this page?
 static inline bool mi_page_has_free(mi_page_t* page) {
   mi_assert_internal(page != NULL);
-  bool hasfree = (mi_page_immediate_available(page) || page->local_free != NULL || (page->thread_free.head != 0));
+  bool hasfree = (mi_page_immediate_available(page) || page->local_free != NULL || (mi_tf_block(page->thread_free) != NULL));
   mi_assert_internal(hasfree || page->used - page->thread_freed == page->capacity);
   return hasfree;
 }
diff --git a/include/mimalloc-types.h b/include/mimalloc-types.h
index 966de36a..2f16020f 100644
--- a/include/mimalloc-types.h
+++ b/include/mimalloc-types.h
@@ -134,20 +134,9 @@ typedef union mi_page_flags_u {
 } mi_page_flags_t;
 
 // Thread free list.
-// We use 2 bits of the pointer for the `use_delayed_free` and `delayed_freeing` flags.
-typedef union mi_thread_free_u {
-  volatile uintptr_t value;
-  struct {
-    uintptr_t delayed:2;
-#if MI_INTPTR_SIZE==8
-    uintptr_t head:62;    // head free block in the list (right-shifted by 2)
-#elif MI_INTPTR_SIZE==4
-    uintptr_t head:30;
-#endif
-  };
-} mi_thread_free_t;
+// We use bottom 2 bits of the pointer for mi_delayed_t flags
+typedef uintptr_t mi_thread_free_t;
 
-#define MI_TF_PTR_SHIFT (2)
 
 // A page contains blocks of one specific size (`block_size`).
 // Each page has three list of free blocks:
diff --git a/src/alloc.c b/src/alloc.c
index 6a2a263f..6a5395c7 100644
--- a/src/alloc.c
+++ b/src/alloc.c
@@ -106,25 +106,25 @@ void* mi_zalloc(size_t size) mi_attr_noexcept {
 // multi-threaded free
 static mi_decl_noinline void _mi_free_block_mt(mi_page_t* page, mi_block_t* block)
 {
-  mi_thread_free_t tfree = {0};
-  mi_thread_free_t tfreex = {0};
+  mi_thread_free_t tfree;
+  mi_thread_free_t tfreex;
   bool use_delayed;
 
   do {
-    tfreex.value = tfree.value = page->thread_free.value;
-    use_delayed = (tfree.delayed == MI_USE_DELAYED_FREE ||
-                   (tfree.delayed == MI_NO_DELAYED_FREE && page->used == page->thread_freed+1)
+    tfree = page->thread_free;
+    use_delayed = (mi_tf_delayed(tfree) == MI_USE_DELAYED_FREE ||
+                   (mi_tf_delayed(tfree) == MI_NO_DELAYED_FREE && page->used == page->thread_freed+1)
                   );
     if (mi_unlikely(use_delayed)) {
       // unlikely: this only happens on the first concurrent free in a page that is in the full list
-      tfreex.delayed = MI_DELAYED_FREEING;
+      tfreex = mi_tf_set_delayed(tfree,MI_DELAYED_FREEING);
     }
     else {
       // usual: directly add to page thread_free list
-      mi_block_set_next(page, block, (mi_block_t*)((uintptr_t)tfree.head << MI_TF_PTR_SHIFT));
-      tfreex.head = (uintptr_t)block >> MI_TF_PTR_SHIFT;
+      mi_block_set_next(page, block, mi_tf_block(tfree));
+      tfreex = mi_tf_set_block(tfree,block);
     }
-  } while (!mi_atomic_compare_exchange((volatile uintptr_t*)&page->thread_free, tfreex.value, tfree.value));
+  } while (!mi_atomic_compare_exchange((volatile uintptr_t*)&page->thread_free, tfreex, tfree));
 
   if (mi_likely(!use_delayed)) {
     // increment the thread free count and return
@@ -145,10 +145,10 @@ static mi_decl_noinline void _mi_free_block_mt(mi_page_t* page, mi_block_t* bloc
 
     // and reset the MI_DELAYED_FREEING flag
     do {
-      tfreex.value = tfree.value = page->thread_free.value;
-      mi_assert_internal(tfree.delayed == MI_NEVER_DELAYED_FREE || tfree.delayed == MI_DELAYED_FREEING);
-      if (tfree.delayed != MI_NEVER_DELAYED_FREE) tfreex.delayed = MI_NO_DELAYED_FREE;
-    } while (!mi_atomic_compare_exchange((volatile uintptr_t*)&page->thread_free, tfreex.value, tfree.value));
+      tfreex = tfree = page->thread_free;
+      mi_assert_internal(mi_tf_delayed(tfree) == MI_NEVER_DELAYED_FREE || mi_tf_delayed(tfree) == MI_DELAYED_FREEING);
+      if (mi_tf_delayed(tfree) != MI_NEVER_DELAYED_FREE) tfreex = mi_tf_set_delayed(tfree,MI_NO_DELAYED_FREE);
+    } while (!mi_atomic_compare_exchange((volatile uintptr_t*)&page->thread_free, tfreex, tfree));
   }
 }
 
@@ -249,12 +249,7 @@ void mi_free(void* p) mi_attr_noexcept
   }
 }
 
-void _mi_free_delayed_block(mi_block_t* block) {
-  mi_assert_internal(block != NULL);
-  const mi_segment_t* segment = _mi_ptr_segment(block);
-  mi_assert_internal(_mi_ptr_cookie(segment) == segment->cookie);
-  mi_assert_internal(_mi_thread_id() == segment->thread_id);
-  mi_page_t* page = _mi_segment_page_of(segment,block);
+void _mi_free_delayed_block(mi_page_t* page, mi_block_t* block) {
   _mi_free_block(page,true,block);
 }
 
diff --git a/src/init.c b/src/init.c
index f6f36c37..5b2d3c8e 100644
--- a/src/init.c
+++ b/src/init.c
@@ -14,7 +14,7 @@ const mi_page_t _mi_page_empty = {
   0, false, false, {0},
   0, 0,
   NULL, 0, 0,   // free, used, cookie
-  NULL, 0, {0},
+  NULL, 0, 0,
   0, NULL, NULL, NULL
   #if (MI_INTPTR_SIZE==4)
   , { NULL }
diff --git a/src/page.c b/src/page.c
index 9f012256..9bc1ab87 100644
--- a/src/page.c
+++ b/src/page.c
@@ -82,7 +82,7 @@ static bool mi_page_is_valid_init(mi_page_t* page) {
   mi_assert_internal(mi_page_list_is_valid(page,page->free));
   mi_assert_internal(mi_page_list_is_valid(page,page->local_free));
 
-  mi_block_t* tfree = (mi_block_t*)((uintptr_t)page->thread_free.head << MI_TF_PTR_SHIFT);
+  mi_block_t* tfree = mi_tf_block(page->thread_free);
   mi_assert_internal(mi_page_list_is_valid(page, tfree));
   size_t tfree_count = mi_page_list_count(page, tfree);
   mi_assert_internal(tfree_count <= page->thread_freed + 1);
@@ -114,17 +114,17 @@ void _mi_page_use_delayed_free(mi_page_t* page, mi_delayed_t delay  ) {
   mi_thread_free_t tfreex;
 
   do {
-    tfreex.value = tfree.value = page->thread_free.value;
-    if (mi_unlikely(tfree.delayed < MI_DELAYED_FREEING)) {
-      tfreex.delayed = delay;
+    tfreex = tfree = page->thread_free;
+    if (mi_unlikely(mi_tf_delayed(tfree) < MI_DELAYED_FREEING)) {
+      tfreex = mi_tf_set_delayed(tfree,delay);
     }
-    else if (mi_unlikely(tfree.delayed == MI_DELAYED_FREEING)) {
+    else if (mi_unlikely(mi_tf_delayed(tfree) == MI_DELAYED_FREEING)) {
       mi_atomic_yield(); // delay until outstanding MI_DELAYED_FREEING are done.
       continue;          // and try again
     }    
   }
-  while(tfreex.delayed != tfree.delayed && // avoid atomic operation if already equal
-        !mi_atomic_compare_exchange((volatile uintptr_t*)&page->thread_free, tfreex.value, tfree.value));
+  while((mi_tf_delayed(tfreex) !=  mi_tf_delayed(tfree)) && // avoid atomic operation if already equal
+        !mi_atomic_compare_exchange((volatile uintptr_t*)&page->thread_free, tfreex, tfree));
 }
 
 
@@ -139,13 +139,13 @@ void _mi_page_use_delayed_free(mi_page_t* page, mi_delayed_t delay  ) {
 static void mi_page_thread_free_collect(mi_page_t* page)
 {
   mi_block_t* head;
-  mi_thread_free_t tfree = {0};
-  mi_thread_free_t tfreex = {0};
+  mi_thread_free_t tfree = 0;
+  mi_thread_free_t tfreex = 0;
   do {
-    tfreex.value = tfree.value = page->thread_free.value;
-    head = (mi_block_t*)((uintptr_t)tfree.head << MI_TF_PTR_SHIFT);
-    tfreex.head = 0;
-  } while (!mi_atomic_compare_exchange((volatile uintptr_t*)&page->thread_free, tfreex.value, tfree.value));
+    tfreex = tfree = page->thread_free;
+    head = mi_tf_block(tfree);
+    tfreex = mi_tf_set_block(tfree,NULL);
+  } while (!mi_atomic_compare_exchange((volatile uintptr_t*)&page->thread_free, tfreex, tfree));
 
   // return if the list is empty
   if (head == NULL) return;
@@ -189,7 +189,7 @@ void _mi_page_free_collect(mi_page_t* page) {
     page->local_free = NULL;
   }
   // and the thread free list
-  if (page->thread_free.head != 0) {  // quick test to avoid an atomic operation
+  if (mi_tf_block(page->thread_free) != NULL) {  // quick test to avoid an atomic operation
     mi_page_thread_free_collect(page);
   }
 }
@@ -258,8 +258,25 @@ void _mi_heap_delayed_free(mi_heap_t* heap) {
   // and free them all
   while(block != NULL) {
     mi_block_t* next = mi_block_nextx(heap->cookie,block);
-    // use internal free instead of regular one to keep stats etc correct
-    _mi_free_delayed_block(block);
+    // get segment and page
+    const mi_segment_t* segment = _mi_ptr_segment(block);
+    mi_assert_internal(_mi_ptr_cookie(segment) == segment->cookie);
+    mi_assert_internal(_mi_thread_id() == segment->thread_id);
+    mi_page_t* page = _mi_segment_page_of(segment, block);
+    if (mi_tf_delayed(page->thread_free) != MI_DELAYED_FREEING) {
+      // use internal free instead of regular one to keep stats etc correct
+      _mi_free_delayed_block(page,block);
+    }
+    else {
+      // we might already start delayed freeing while another thread has not yet 
+      // reset the flag; in that case delay it further :-(
+      mi_block_t* dfree;
+      do {
+        dfree = (mi_block_t*)heap->thread_delayed_free;
+        mi_block_set_nextx(heap->cookie, block, dfree);
+      } while (!mi_atomic_compare_exchange_ptr((volatile void**)&heap->thread_delayed_free, block, dfree));
+
+    }
     block = next;
   }
 }
@@ -334,11 +351,9 @@ void _mi_page_free(mi_page_t* page, mi_page_queue_t* pq, bool force) {
   mi_assert_internal(mi_page_all_free(page));
   #if MI_DEBUG>1
   // check if we can safely free
-  mi_thread_free_t free;
-  free.value = page->thread_free.value;
-  free.delayed = MI_NEVER_DELAYED_FREE;
-  free.value = mi_atomic_exchange(&page->thread_free.value, free.value);
-  mi_assert_internal(free.delayed != MI_DELAYED_FREEING);
+  mi_thread_free_t free = mi_tf_set_delayed(page->thread_free,MI_NEVER_DELAYED_FREE);
+  free = mi_atomic_exchange(&page->thread_free, free);
+  mi_assert_internal(mi_tf_delayed(free) != MI_DELAYED_FREEING);
   #endif
 
   page->flags.has_aligned = false;
@@ -538,7 +553,7 @@ static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t block_size, mi
   mi_assert_internal(page->capacity == 0);
   mi_assert_internal(page->free == NULL);
   mi_assert_internal(page->used == 0);
-  mi_assert_internal(page->thread_free.value == 0);
+  mi_assert_internal(page->thread_free == 0);
   mi_assert_internal(page->thread_freed == 0);
   mi_assert_internal(page->next == NULL);
   mi_assert_internal(page->prev == NULL);