Re: Guidelines for instruction set design?

cyril.cressent@gmail.com writes:




> I was wondering if there are some general guidelines one should
> observe when designing an instruction set so that a C compiler can
> easily be ported to that CPU.


If your CPU supports the standard C datatypes (signed and unsigned
chars, shorts, integers and longs, single- and double precision reals
and pointers to byte addresses), you should not have that great
difficulties. You can make it simplest for yourself by choosing char
= 8 bits, short = int = long = pointer = float = 32 bits and double =
64 bits.


Having load/store instructions for all the sizes of char, integer,
long, pointer, float and double is also nice, as having to do a byte
store as a full-word load, a masking operation and a full-word store
is a pain. Doing a double-word load/store as two separate word
load/stores or a half-word load/store as two byte load/stores is less
of a problem, so you can make do with just byte and (32-bit) word
load/stores.


Shift left and right by arbitrary values is also useful, but if you
have single-bit shifts, you can do the rest in a loop. Rotates are
less useful, as C doesn't have rotates (though some compilers will
recognize certain combinations of shifts and and/ors as a rotate and
compile to this).


You should also be able to compare signed and unsigned numbers for <.


And there should be some support for procedure calls, i.e., a jump
that stores a return address in a register or memory and a jump to an
address stored in a register or in memory (so you can return).


Apart form the above, I can't think of anything essential.


You can implement floats and doubles in libraries using integer
instructions, as long as you have shifts and masks (and/or), but this
will be slower than having hardware for it.


Torben

Return to the comp.compilers page.
Search the comp.compilers archives again.