@@ -342,12 +342,13 @@ void mi_bitmap_clear_once_set(mi_subproc_t* subproc, mi_bitmap_t* bitmap, size_t
342342
343343
344344// Visit all set bits in a bitmap.
345+ // Acquire even cleared bits: heap deletion must synchronize with concurrent page releases.
345346// todo: optimize further? maybe use avx512 to directly get all indices using a mask_compressstore?
346347bool _mi_bitmap_forall_set (mi_bitmap_t * bitmap , mi_forall_set_fun_t * visit , mi_arena_t * arena , void * arg ) {
347348 // for all chunkmap entries
348349 const size_t chunkmap_max = mi_divide_up (mi_bitmap_chunk_count (bitmap ), MI_BFIELD_BITS );
349350 for (size_t i = 0 ; i < chunkmap_max ; i ++ ) {
350- mi_bfield_t cmap_entry = mi_atomic_load_relaxed (& bitmap -> chunkmap .bfields [i ]);
351+ mi_bfield_t cmap_entry = mi_atomic_load_acquire (& bitmap -> chunkmap .bfields [i ]);
351352 size_t cmap_idx ;
352353 // for each chunk (corresponding to a set bit in a chunkmap entry)
353354 while (mi_bfield_foreach_bit (& cmap_entry , & cmap_idx )) {
@@ -356,7 +357,7 @@ bool _mi_bitmap_forall_set(mi_bitmap_t* bitmap, mi_forall_set_fun_t* visit, mi_a
356357 mi_bchunk_t * const chunk = & bitmap -> chunks [chunk_idx ];
357358 for (size_t j = 0 ; j < MI_BCHUNK_FIELDS ; j ++ ) {
358359 const size_t base_idx = (chunk_idx * MI_BCHUNK_BITS ) + (j * MI_BFIELD_BITS );
359- mi_bfield_t b = mi_atomic_load_relaxed (& chunk -> bfields [j ]);
360+ mi_bfield_t b = mi_atomic_load_acquire (& chunk -> bfields [j ]);
360361 size_t bidx ;
361362 while (mi_bfield_foreach_bit (& b , & bidx )) {
362363 const size_t idx = base_idx + bidx ;
0 commit comments