remove maxaccessed from general bitmaps

src/arena-meta.c

@@ -77,7 +77,7 @@ static mi_meta_page_t* mi_meta_page_zalloc(void) {
   // initialize the page
   mpage->memid = memid;
   mi_bbitmap_init(&mpage->blocks_free, MI_META_BLOCKS_PER_PAGE, true /* already_zero */);
-  const size_t mpage_size  = offsetof(mi_meta_page_t,blocks_free) + mi_bitmap_size(MI_META_BLOCKS_PER_PAGE, NULL);
+  const size_t mpage_size  = offsetof(mi_meta_page_t,blocks_free) + mi_bbitmap_size(MI_META_BLOCKS_PER_PAGE, NULL);
   const size_t info_blocks = _mi_divide_up(mpage_size,MI_META_BLOCK_SIZE);
   mi_assert_internal(info_blocks < MI_META_BLOCKS_PER_PAGE);
   mi_bbitmap_unsafe_setN(&mpage->blocks_free, info_blocks, MI_META_BLOCKS_PER_PAGE - info_blocks);
@@ -142,7 +142,7 @@ mi_decl_noinline void _mi_meta_free(void* p, size_t size, mi_memid_t memid) {
     mi_meta_page_t* mpage = (mi_meta_page_t*)memid.mem.meta.meta_page; 
     mi_assert_internal(mi_meta_page_of_ptr(p,NULL) == mpage);
     mi_assert_internal(block_idx + block_count < MI_META_BLOCKS_PER_PAGE);
-    mi_assert_internal(mi_bitmap_is_clearN(&mpage->blocks_free, block_idx, block_count));
+    mi_assert_internal(mi_bbitmap_is_clearN(&mpage->blocks_free, block_idx, block_count));
     // we zero on free (and on the initial page allocation) so we don't need a "dirty" map
     _mi_memzero_aligned(mi_meta_block_start(mpage, block_idx), block_count*MI_META_BLOCK_SIZE);
     mi_bbitmap_setN(&mpage->blocks_free, block_idx, block_count);

src/arena.c

@@ -1258,7 +1258,7 @@ static size_t mi_debug_show_chunks(const char* header, size_t slice_count, size_
 
     char chunk_kind = ' ';
     if (chunk_bins != NULL) {
-      switch (chunk_bins[i]) {
+      switch (mi_atomic_load_relaxed(&chunk_bins[i])) {
         // case MI_BBIN_SMALL: chunk_kind  = 'S'; break;
         case MI_BBIN_MEDIUM: chunk_kind = 'M'; break;
         case MI_BBIN_LARGE: chunk_kind  = 'L'; break;

src/bitmap.c

@@ -45,6 +45,14 @@ static inline bool mi_bfield_find_least_bit(mi_bfield_t x, size_t* idx) {
   return mi_bsf(x,idx);
 }
 
+// find the most significant bit that is set.
+// return false if `x==0` (with `*idx` undefined) and true otherwise,
+// with the `idx` is set to the bit index (`0 <= *idx < MI_BFIELD_BITS`).
+static inline bool mi_bfield_find_highest_bit(mi_bfield_t x, size_t* idx) {
+  return mi_bsf(x, idx);
+}
+
+
 
 // find each set bit in a bit field `x` and clear it, until it becomes zero.
 static inline bool mi_bfield_foreach_bit(mi_bfield_t* x, size_t* idx) {
@@ -873,17 +881,9 @@ static bool mi_bchunk_bsr(mi_bchunk_t* chunk, size_t* pidx) {
  bitmap chunkmap
 -------------------------------------------------------------------------------- */
 
-static void mi_bitmap_chunkmap_set_max(mi_bitmap_t* bitmap, size_t chunk_idx) {
-  size_t oldmax = mi_atomic_load_relaxed(&bitmap->chunk_max_accessed);
-  if mi_unlikely(chunk_idx > oldmax) {
-    mi_atomic_cas_strong_relaxed(&bitmap->chunk_max_accessed, &oldmax, chunk_idx);
-  }
-}
-
 static void mi_bitmap_chunkmap_set(mi_bitmap_t* bitmap, size_t chunk_idx) {
   mi_assert(chunk_idx < mi_bitmap_chunk_count(bitmap));
   mi_bchunk_set(&bitmap->chunkmap, chunk_idx);  
-  mi_bitmap_chunkmap_set_max(bitmap, chunk_idx);
 }
 
 static bool mi_bitmap_chunkmap_try_clear(mi_bitmap_t* bitmap, size_t chunk_idx) {
@@ -898,7 +898,6 @@ static bool mi_bitmap_chunkmap_try_clear(mi_bitmap_t* bitmap, size_t chunk_idx)
     mi_bchunk_set(&bitmap->chunkmap, chunk_idx);
     return false;
   }
-  mi_bitmap_chunkmap_set_max(bitmap, chunk_idx);
   return true;
 }
 
@@ -1107,25 +1106,16 @@ typedef bool (mi_bitmap_visit_fun_t)(mi_bitmap_t* bitmap, size_t chunk_idx, size
 // If it returns `true` stop the search.
 static inline bool mi_bitmap_find(mi_bitmap_t* bitmap, size_t tseq, size_t n, size_t* pidx, mi_bitmap_visit_fun_t* on_find, void* arg1, void* arg2)
 {
-  // we space out threads to reduce contention
-  const size_t cmap_max_count  = _mi_divide_up(mi_bitmap_chunk_count(bitmap),MI_BFIELD_BITS);
-  const size_t chunk_acc       = mi_atomic_load_relaxed(&bitmap->chunk_max_accessed);
-  const size_t cmap_acc        = chunk_acc / MI_BFIELD_BITS;
-  const size_t cmap_acc_bits   = 1 + (chunk_acc % MI_BFIELD_BITS);
-
-  // create a mask over the chunkmap entries to iterate over them efficiently
-  mi_assert_internal(MI_BFIELD_BITS >= MI_BCHUNK_FIELDS);
-  const mi_bfield_t cmap_mask  = mi_bfield_mask(cmap_max_count,0);
-  const size_t cmap_cycle      = cmap_acc+1;
-  mi_bfield_cycle_iterate(cmap_mask, tseq, cmap_cycle, cmap_idx, X)
-  {
+  const size_t chunkmap_max = _mi_divide_up(mi_bitmap_chunk_count(bitmap), MI_BFIELD_BITS);
+  for (size_t i = 0; i < chunkmap_max; i++) {
     // and for each chunkmap entry we iterate over its bits to find the chunks
-    mi_bfield_t cmap_entry = mi_atomic_load_relaxed(&bitmap->chunkmap.bfields[cmap_idx]);
-    size_t cmap_entry_cycle = (cmap_idx != cmap_acc ? MI_BFIELD_BITS : cmap_acc_bits);
-    mi_bfield_cycle_iterate(cmap_entry, tseq%8, cmap_entry_cycle, eidx, Y) // reduce the tseq to 8 bins to reduce using extra memory (see `mstress`)
+    const mi_bfield_t cmap_entry = mi_atomic_load_relaxed(&bitmap->chunkmap.bfields[i]);
+    size_t hi;
+    if (mi_bfield_find_highest_bit(cmap_entry, &hi)) {
+      mi_bfield_cycle_iterate(cmap_entry, tseq%8, hi+1, eidx, Y) // reduce the tseq to 8 bins to reduce using extra memory (see `mstress`)
       {
         mi_assert_internal(eidx <= MI_BFIELD_BITS);
-      const size_t chunk_idx = cmap_idx*MI_BFIELD_BITS + eidx;
+        const size_t chunk_idx = i*MI_BFIELD_BITS + eidx;
         mi_assert_internal(chunk_idx < mi_bitmap_chunk_count(bitmap));
         if ((*on_find)(bitmap, chunk_idx, n, pidx, arg1, arg2)) {
           return true;
@@ -1133,7 +1123,7 @@ static inline bool mi_bitmap_find(mi_bitmap_t* bitmap, size_t tseq, size_t n, si
       }
       mi_bfield_cycle_iterate_end(Y);
     }
-  mi_bfield_cycle_iterate_end(X);
+  }
   return false;
 }
 
@@ -1478,7 +1468,7 @@ static inline bool mi_bbitmap_try_find_and_clear_generic(mi_bbitmap_t* bbitmap,
         // only in the current size class!
         const mi_bbin_t chunk_bin = (mi_bbin_t)mi_atomic_load_acquire(&bbitmap->chunk_bins[chunk_idx]);
         if // (bin >= chunk_bin) { 
-           (bin == chunk_bin || (bin <= MI_BBIN_SMALL && chunk_bin <= MI_BBIN_SMALL)) {
+           ((mi_bbin_t)bin == chunk_bin || (bin <= MI_BBIN_SMALL && chunk_bin <= MI_BBIN_SMALL)) {
           mi_bchunk_t* chunk = &bbitmap->chunks[chunk_idx];
           size_t cidx;
           if ((*on_find)(chunk, n, &cidx)) {

src/bitmap.h

@@ -98,8 +98,7 @@ typedef mi_bchunk_t mi_bchunkmap_t;
 // An atomic bitmap
 typedef mi_decl_align(MI_BCHUNK_SIZE) struct mi_bitmap_s {
   _Atomic(size_t)  chunk_count;         // total count of chunks (0 < N <= MI_BCHUNKMAP_BITS)
-  _Atomic(size_t)  chunk_max_accessed;  // max chunk index that was once cleared or set
-  size_t           _padding[MI_BCHUNK_SIZE/MI_SIZE_SIZE - 2];    // suppress warning on msvc
+  size_t           _padding[MI_BCHUNK_SIZE/MI_SIZE_SIZE - 1];    // suppress warning on msvc
   mi_bchunkmap_t   chunkmap;
   mi_bchunk_t      chunks[MI_BITMAP_DEFAULT_CHUNK_COUNT];        // usually dynamic MI_BITMAP_MAX_CHUNK_COUNT
 } mi_bitmap_t;

src/heap.c

@@ -139,6 +139,8 @@ static void mi_heap_collect_ex(mi_heap_t* heap, mi_collect_t collect)
   // collect retired pages
   _mi_heap_collect_retired(heap, force);
   
+  // if (_mi_is_main_thread()) { mi_debug_show_arenas(true, false, false); }
+  
   // collect all pages owned by this thread
   mi_heap_visit_pages(heap, &mi_heap_page_collect, &collect, NULL);
   

src/page-map.c

@@ -17,7 +17,7 @@ static mi_memid_t  mi_page_map_memid;
 
 
 // (note: we need to initialize statically or otherwise C++ may run a default constructors after process initialization)
-static mi_bitmap_t mi_page_map_commit = { MI_ATOMIC_VAR_INIT(MI_BITMAP_DEFAULT_CHUNK_COUNT), MI_ATOMIC_VAR_INIT(0),
+static mi_bitmap_t mi_page_map_commit = { MI_ATOMIC_VAR_INIT(MI_BITMAP_DEFAULT_CHUNK_COUNT),
                                           { 0 }, { {MI_ATOMIC_VAR_INIT(0)} }, {{{ MI_ATOMIC_VAR_INIT(0) }}} };
 
 bool _mi_page_map_init(void) {

src/page.c

@@ -872,10 +872,14 @@ void* _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero, size_t huge_al
   mi_assert_internal(mi_heap_is_initialized(heap));
 
   // call potential deferred free routines
-  // _mi_deferred_free(heap, false);
+  _mi_deferred_free(heap, false);
 
-  // free delayed frees from other threads (but skip contended ones)
-  // _mi_heap_delayed_free_partial(heap);
+  // collect every N generic mallocs
+  /*static long count = 0;
+  if (count++ > 100000) {
+    count = 0;
+    _mi_heap_collect_retired(heap,false);
+  }*/
 
   // find (or allocate) a page of the right size
   mi_page_t* page = mi_find_page(heap, size, huge_alignment);

test/test-stress.c

@@ -262,7 +262,7 @@ static void test_stress(void) {
     #if !defined(NDEBUG) || defined(MI_TSAN)
     if ((n + 1) % 10 == 0) { 
       printf("- iterations left: %3d\n", ITER - (n + 1)); 
-      //mi_debug_show_arenas(true, false, false);
+      mi_debug_show_arenas(true, false, false);
       //mi_collect(true);
       //mi_debug_show_arenas(true, false, false);
     }
@@ -352,7 +352,7 @@ int main(int argc, char** argv) {
   mi_collect(true);
   mi_debug_show_arenas(true,false,false);
   #else
-  //mi_collect(true);
+  mi_collect(false);
   mi_debug_show_arenas(true,false,false);
   // mi_stats_print(NULL);
   #endif