Using Large Memory Model, Microsoft C/C++, & Windows 3.1 (90294)

MORE INFORMATION

1. Large model applications can have multiple data segments (this is always the case when compiling with Microsoft C 6.0). In Windows 3.0, these extra segments are loaded as FIXED and page-locked. Having a large amount of FIXED page-locked memory results in a serious degradation of the memory manager's performance. FIXED memory is allocated from the bottom of the global heap, which means it normally lies in conventional memory (below 1 MB), which is needed by the Windows loader when it loads a new module into memory. For more information on this topic, query this knowledge base on the following words:

   using and memory and below and megabyte

   This problem has been corrected in Windows 3.1. In Windows 3.1, an application's code and data segments are always loaded as MOVEABLE, regardless of the segment attributes specified in the application's definition (.DEF) file. A DLL's code and data segments are loaded exactly as they are specified in the DLL's .DEF file.
2. Under both Windows 3.0 and Windows 3.1, only one instance of an application with multiple data segments can be run at one time. This is because the Windows loader can't fix-up multiple instances of a far pointer because code is shared among all instances of an application.
   
   In most cases, this problem may be resolved by using the /Gx and /Gt compile options with the Microsoft C/C++ 7.0 compiler. The /Gx option causes the compiler to force static data into the default data segment, which results in the application having only one data segment. Note, however, that this works only if the application's static data, string literals, stack, and heap all fit into a 64K segment.
   
   The /Gt switch specifies the size a static object must be to be allocated its own data segment. When trying to create a single data segment, this threshold value should be large enough to ensure that no data objects are allocated outside of the default data segment. If the /Gt option is not specified, the data threshold will be 32767.
3. The final problem with using large model under both Windows 3.0 and Windows 3.1 is the performance degradation that occurs when an application uses far pointers to access its data.
   
   When running under the protected mode of the Intel 286, 386, and 486 processors, a far address is a combination of a selector and an offset. A selector is essentially an index into an array called a descriptor table. A descriptor is an 8-byte value that contains information about the segment, such as its base address, size, read/write privileges, and so forth.
   
   Whenever a far variable is referenced in protected mode, the processor must load the descriptor into one of the segment registers' descriptor cache and mark the descriptor as being accessed. This means that a reference to a far variable requires at least two reads from and one write to memory just to obtain the variable's linear address.
   
   Adding this delay to the time needed for performing far pointer arithmetic amounts to a significant loss of performance when compared to the medium model, which uses near pointers for data. One method to increase the performance of a large model application is to explicitly declare commonly used global or static variables as being NEAR, which causes the compiler to allocate the variable in the default data segment, and to use a near pointer when referencing the variable.

Conclusion