Jinoh Kang (@iamahuman) commented about dlls/ntdll/heap.c:
+} + +static inline struct block *group_get_block( struct group *group, SIZE_T block_size, UINT index ) +{ + char *first_block = (char *)&group->first_block; + return (struct block *)(first_block + index * block_size); +} + +/* lookup a free block using the group free_bits, the current thread must own the group */ +static inline LONG group_find_free_block( struct group *group, SIZE_T block_size, struct block **block ) +{ + ULONG i, free_bits = ReadNoFence( &group->free_bits ); + /* free_bits & ~GROUP_FLAG_FREE will never be 0 as the group is unlinked when it's fully used */ + BitScanForward( &i, free_bits & ~GROUP_FLAG_FREE ); + *block = group_get_block( group, block_size, i ); + return InterlockedAnd( &group->free_bits, ~(1 << i) ) & ~(1 << i); Try not to use `InterlockedAnd`'s return value if you don't want it to transform to cmpxchg loop.
-- https://gitlab.winehq.org/wine/wine/-/merge_requests/1628#note_23444