In 16-bit versions of Windows (1.x, 2.x, 3.x), each module (application or DLL) typically has its own data segment (DGROUP) limited to 64 KB. This segment contains the stack, the local heap, and the atom table. Local heap functions manage memory within this segment using near pointers (offsets). They are exported by the kernel (KRNL386.EXE). Internally, the local heap is managed through a set of data structures that reside inside the segment itself.

A key part of this management is the Instance Data (also called the NULL segment) located at the very beginning of DGROUP, which holds pointers to the heap, atom table, and stack information. The field at offset 6 (pLocalHeap) contains a near pointer to the HeapInfo structure that heads the local heap. This pointer is set by LocalInit() and must be validated by checking the signature word (li_sig) at offset 22h (Standard mode) or 28h (KRNL386) within the presumed heap.

The local heap itself is organized as a series of arenas (headers) that precede each block. The two low bits of the la_prev field in an arena are used as flags: bit 0 indicates whether the block is in use (1) or free (0); bit 1 indicates whether the block is MOVEABLE (1) or FIXED (0). Free blocks are linked via la_free_prev and la_free_next. For MOVEABLE blocks, a separate handle table (pointed to by hi_htable in HeapInfo) stores the actual address, lock count, and flags; each handle is an offset into this table.

The first 16 (10h) bytes of the default data segment (DGROUP) are reserved for system use and are collectively referred to as the Instance Data or NULL segment. This area is present only if the first WORD at offset 0 is zero; otherwise, the structure is not present. The layout is as follows:

Important Notes: * The field at offset 6 (pLocalHeap) is the primary way to locate the local heap structures given only the DGROUP selector. * When LocalInit() is called on a globally allocated block (non‑DGROUP), the WORD at offset 6 of that block is also set to point to the local heap information structure for that block. * Similarly, if InitAtomTable() is called on a global block, offset 8 points to the atom table, and offset 6 will point to the associated local heap (since atoms are stored in the local heap).

Every local heap begins with an instance of the HeapInfo structure, which is identical to the one used by the global heap and is defined in WINKERN.INC. Its location is given by the pLocalHeap field at offset 6 of the Instance Data. Immediately following the HeapInfo structure are additional fields that, together with HeapInfo, form the LocalInfo structure.

In the enhanced mode (krnl386), the HeapInfo structure occupies 1Eh bytes and has the following format (according to “Windows Internals” documentation, Table 2-2):

In standard mode (krnl286), the HeapInfo structure occupies 18h bytes. Due to the 16‑bit segmented architecture, near pointers are used instead of far pointers, and all size fields are limited to 64 KB. The layout is as follows:

Note: In krnl286, all pointer fields are near (16‑bit offsets) because the local heap resides in a single 64‑KB segment.

For krnl386 (HeapInfo size of 1Eh), the LocalInfo structure has the following layout:

For krnl286, the LocalInfo structure immediately follows the HeapInfo structure and contains additional fields that manage heap notifications, locking, sizing, and a signature word. Its layout begins at offset 18h from the start of the local heap information block:

Important: In krnl286, the li_sig field is located at offset 22h from the beginning of the combined HeapInfo + LocalInfo structure.

The total size of the local heap information block (HeapInfo + LocalInfo) for krnl286 is therefore 24h bytes (18h + 0Ch), though only the first 22h bytes are used up to the signature. The signature itself occupies bytes 22h–23h.

Every block in the local heap is preceded by an arena (header) that contains management information. Arenas always start on a 4‑byte boundary, so the two low bits of every arena address are zero. These bits are reused as flags in the la_prev field of each arena. The two low bits of la_prev have the following meaning:

* Bit 0 (least significant): Set if the block is in use (FIXED or MOVEABLE); cleared if the block is free. * Bit 1: Set if the block is MOVEABLE; cleared if the block is FIXED (only meaningful when bit 0 is set).

Thus, to obtain the real address of the previous arena, the two low bits must be masked off.

There are three arena formats, corresponding to the three possible block states:

For a FIXED block, the handle is the address of the block itself. The arena can be found by subtracting 4 from the block address.

The la_handle field provides two‑way mapping between a MOVEABLE block and its handle table entry. Given the block address, subtract 6 to get the arena; the la_handle field gives the offset of the handle entry. Given a handle entry, the first field (lhe_address) gives the block address.

Free blocks are threaded in a doubly linked list using la_free_prev and la_free_next. The start of this free list is stored in the la_free_next field of the first block in the heap (see below).

The first block in the local heap is special. It resides in memory before the LocalInfo structure. Although its la_prev field has the low bit set (indicating a FIXED, in‑use arena), the local heap routines treat this arena as if it were free. The la_free_next field of this first block points to the first truly free block in the heap.

For MOVEABLE blocks, handles are offsets into a handle table that resides in its own local heap blocks. The first handle table is located via the hi_htable field in the HeapInfo structure. Each handle table begins with a WORD specifying how many handle entries follow. After all entries, the last WORD is the offset of the next handle table (or 0 if none). Free handle entries are linked together for quick allocation.

Format of Handle Table Entries (in-use):

Format of Handle Table Entries (free):

* Allocation (LocalAlloc) walks the free list, splitting blocks if necessary, and sets up the appropriate arena. For MOVEABLE blocks, it also allocates a handle table entry. * Compaction (LocalCompact) coalesces adjacent free blocks and may move or discard unlocked MOVEABLE blocks. When a block is moved, its lhe_address is updated. * Locking (LocalLock/LocalUnlock) manipulates the lhe_count field of the handle entry for MOVEABLE blocks; for FIXED blocks, no count is maintained. * Discarding (LocalDiscard) frees the memory of a MOVEABLE block but keeps the handle entry alive with the LHE_DISCARDED flag set.

Atom tables provide a mechanism for efficiently storing and sharing strings (atoms) in 16-bit Windows. Each module may have its own local atom table located in its local heap. In addition, there is a single global atom table that serves the entire system.

The field at offset 08h in the Instance Data (`pAtomTable`) contains a near pointer to the atom table header for the current module. This pointer is initialized by calling `InitAtomTable()`. If no atom table exists, `pAtomTable` is zero.

Physically, an atom table resides inside the local heap of some data segment. Therefore, before creating an atom table, the segment must be initialized as a local heap by calling `LocalInit()`.

Each string atom is stored as an `ATOMENTRY` structure in the local heap. The structure has the following form:

Windows supports two fundamentally different types of atoms: string atoms and integer atoms. Their handling is completely distinct.

String atoms are created by passing an ordinary string to `AddAtom` or `GlobalAddAtom`. They are stored in the atom table as `ATOMENTRY` structures.

* Range: `0xC000` to `0xFFFF` (encoded pointer). * Storage: Allocated in the local heap as `ATOMENTRY`, inserted into the hash table. * Reference count: Yes (`usage` field). * String representation: The original string. * Creation: `AddAtom(“MyString”)`.

Encoding: A string atom value is derived from the near pointer to its `ATOMENTRY`. Since the pointer is 4‑byte aligned, the low two bits are zero. The atom is formed by shifting the pointer right by 2 bits and ORing with `0xC000`. This guarantees the range `0xC000–0xFFFF`.

#define HANDLETOATOM(handle) ((ATOM)(0xc000 | ((handle) >> 2)))
#define ATOMTOHANDLE(atom)   ((HANDLE16)(atom) << 2)

Integer atoms are created by passing a string of the form `“#dddd”` (or by using `MAKEINTATOM` with a value ≤ 0xBFFF). They are not stored in the atom table and have no associated `ATOMENTRY` structure.

* Range: `0x0001` to `0xBFFF` . * Storage: None; the value is used directly as the atom. * Reference count: Not applicable. * String representation: Generated on the fly as `“#dddd”` when `GetAtomName` is called. * Creation: `AddAtom(“#1234”)` or `AddAtom(MAKEINTATOM(0x04D2))`.

How it works: When a string of the form '#dddd' is passed, the function parses the decimal number and, if it is less than 0xC000, returns it directly without accessing the atom table. Similarly, `FindAtom` for such a string or for a `MAKEINTATOM` value simply returns the number without any lookup. Integer atoms are always considered “found” because any value in the range is valid.

The `MAKEINTATOM` macro is defined as:

#define MAKEINTATOM(i)  (LPTSTR)((DWORD)((WORD)(i)))

This macro casts a 16‑bit integer value to a pointer type. When this “pointer” is passed to atom functions, it is interpreted as an integer atom (if the value is ≤ `0xBFFF`) or as a string atom (if ≥ `0xC000`).

* For values ≤ `0xBFFF`, the function treats it as an integer atom and returns the value directly. * For values ≥ `0xC000`, the function assumes it is an encoded pointer to an `ATOMENTRY` and will dereference it (after shifting left by 2 bits) to access the atom entry.

Important: `MAKEINTATOM` does not create a string or allocate any memory; it is simply a type-punning convenience to pass integer atoms to functions that formally expect a string pointer.

* Local atom tables: Bound to a specific data segment (e.g., an application's DGROUP). Created by calling `InitAtomTable()`. Used for a module's internal needs. Access is only possible when the DS register points to that segment. * Global atom table: A system-wide table accessible to all applications via `GlobalAddAtom`, `GlobalFindAtom`, and `GlobalDeleteAtom` . Physically, it resides not in an application's data segment but in a special USER data segment (part of the so-called “global atom and text heap”) . Its structure is identical to a local atom table. The `Global…` functions internally switch DS to the USER segment and call the ordinary `AddAtom`/`FindAtom`.

Since all atom operations work with the current segment pointed to by DS, you can create and use an atom table in any arbitrary data segment by following three steps:

1. Create a local heap in the target segment using `LocalInit(Selector, Start, End)`. 2. Switch the DS register to that segment. 3. Call `InitAtomTable(size)` to initialize the atom table in the newly created heap.

After that, any subsequent call to `AddAtom`, `FindAtom`, etc., will operate on the custom table if DS is temporarily set to the correct segment.

Although each module typically has a default local heap in its DGROUP, it is possible to create additional local heaps in other segments or in globally allocated memory. This is useful for managing private memory pools within a large global block, or for creating atom tables in separate memory areas.

The `LocalInit()` function initializes a local heap within a specified segment. Its prototype is:

WORD LocalInit(WORD wSegment, WORD pStart, WORD pEnd);

* `wSegment` – Selector of the segment where the heap will be created. * `pStart` – Offset of the first byte of the heap area (must be paragraph‑aligned, i.e., a multiple of 16). * `pEnd` – Offset of the last byte of the heap area (inclusive). The heap will manage memory from `pStart` to `pEnd`.

If successful, `LocalInit()` returns a non‑zero value. It sets up the `HeapInfo` and `LocalInfo` structures at the beginning of the heap area (starting at `pStart`) and updates the segment’s instance data at offset 06h (`pLocalHeap`) to point to that `HeapInfo` structure. However, if the segment is not a default data segment (i.e., not DGROUP), the instance data at offset 0 must also contain a zero word to indicate that the NULL segment structure is present; otherwise, the heap may not be recognized by some routines.

Example: Creating a local heap in a globally allocated block of memory (64 KB):

; 1. Allocate a 64KB global memory block
GlobalAlloc GMEM_FIXED, 0x10000
mov dx, ax          ; DX = selector of allocated block
 
; 2. Temporarily set DS to that segment to access its instance data
push ds
push dx
pop ds
 
; 3. Initialize the NULL segment (Instance Data) at offset 0.
;    The first word must be zero (wMustBeZero = 0).
xor ax, ax
mov word ptr [0], ax
; The other fields (pLocalHeap, etc.) will be filled by LocalInit.
 
; 4. Define the heap area: start at offset 16 (0x0010) to preserve
;    the 16-byte Instance Data, end at 0xFFFF (the last byte of the segment).
mov bx, 16          ; pStart = 16
mov cx, 0xFFFF      ; pEnd = 0xFFFF
 
; 5. Restore DS (if no longer needed)
pop ds
 
; 6. Call LocalInit to create the heap in segment dx, from pStart to pEnd.
push dx
push bx
push cx
call LocalInit      ; returns non-zero on success
 
; After the call, the Instance Data at offset 6 in segment dx
; contains a valid near pointer to the HeapInfo structure located at offset 16.

After this call, the global block can be used with local heap functions (`LocalAlloc`, `LocalFree`, etc.) by using the selector in `DX` and near pointers (offsets) within that segment.

Important Considerations: * The heap structures themselves occupy space at the beginning of the heap area. The first block (sentinel) resides at `pStart + size of (LocalInfo)`. * The segment’s instance data (at offset 0) must be properly set up, especially the zero word at offset 0, to avoid confusion with other structures. * Custom local heaps are not automatically enlarged if they run out of space; they are limited to the range specified in `LocalInit`. * The `HEAPSIZE` setting in the module’s .DEF file only affects the default DGROUP heap.

As already described, to create an atom table in an arbitrary memory area, you must first initialize a local heap there (as above) and then, after switching DS, call `InitAtomTable`. This technique allows fully isolated atom tables for special purposes.

* Use `LocalInit` on a segment to establish a local heap anywhere in memory. * The segment must have a valid NULL segment structure (zero word at offset 0) for the heap to be recognized. * After `LocalInit`, you can use `LocalAlloc`, `LocalLock`, etc., with near pointers within that segment. * To create an atom table in a custom heap, switch DS to that segment and call `InitAtomTable`. * All subsequent atom operations must be performed with DS set appropriately (or via wrapper functions). * Custom heaps and atom tables are useful for isolating memory pools, implementing resource managers, or working with large data structures without polluting the default DGROUP.

1. Schulman, A., Maxey, D., Pietrek, M. Undocumented Windows. Addison-Wesley, 1992. 2. Pietrek, M. Windows Internals. Addison-Wesley, 1993. 3. Chen, R. The Old New Thing (blog). Microsoft Developer Blogs. 4. Microsoft OS/2 Version 1.1 Programmer's Reference, Volume 1.