[RFC] Multiple memory-spaces in GCC feature request

["Svein E. Seldal" <Svein dot Seldal at solidas dot com>]

Hi,

There is a specific feature that I'm missing from GCC: The ability to 
address multiple memory-spaces. Some embedded processors like the very 
popular AVR, the 8051, and a couple of high-speed DSP's, uses two (or 
possibly more) independent memoryspaces. One for code and one for data. 
I.e. address 0x1000 is a valid addresses in both spaces, yet they point 
to two entirely different locations. The CPU usually differentiates 
between the memory-spaces using completely different INSN sets.

IMHO what GCC needs is the following features. I can post a longer 
discussion and explanations of how and why if you'd like me to.

a) The ability to store variables into codespace. This is already 
partially supported by the AVR target using the progmem attribute:
     int data __attribute__((progmem)) = 42;

b) The ability to declare type attributes on pointers (which belongs to 
the typedef)
     int __attribute__((progmempointer)) * ptr;

c) GCC should be able to switch which INSN set to use when accessing the 
given memoryspace indicated by the appropriate attribute.
     ptr = &data;   // Valid because of the attributes on both decl.
     temp = *ptr;   // Will access the programspace and retrieve 42.

d) The ability to store strings into codespace as well:
     const char __attribute__((progmempointer) * string
     __attribute__((progmem)) =  "Hello World!";

(The progmempointer attribute declares a pointer to a code-address, 
while the progmem attributes tells the compiler to place the literal 
string into codespace.)

These features should apply only to those target that requires this of 
course. And the names "codepointer" is of course a matter of discussion.

My problem is that I do not have the faintest idea of where to start to 
implementing this. I would surely like somee help and/or some pointer to 
how to attack this challenge. Esp. those parts concerning the global GCC 
issues, like type attributes on pointers.

I would surely appreciate some input and comment about this topic.


Regards,
Svein