Long and short jumps

[Mike Stump]

On Friday, May 21, 2004, at 10:50  AM, Nathan Sidwell wrote:
I have this architecture which has limited range short jumps.
Everybody does...

Long range jumps require a register
Yup.

Then have a machine dependent reorg pass that fixes the long 
conditional jumps to be short conditional jumps to unconditional long 
jump stubs.  If possible I'd like the opportunity before register 
allocation to remove the link clobbers, if the function is small 
enough that a short jump can always be used.
It would be nice to have this type of functionality generally available 
in the compiler.

The recent work with hot/cold partitioning faces these same type of 
problems, as do the various longcall bits.  I think it'd be nice to 
unify and expand the functionality and move more and more of it into 
the mi parts of the compiler.

On our platform, we have a jmp instruction that takes two places to 
jump, one clobbers a register and gets there even if it is far away, 
and the other is just the destination.  Later (as or ld time), when we 
know how far the code will be away, we use the short form of the branch 
instruction, if the distance is short enough, else we jump to the 
branch island (that's the one that kills the register and does the 32 
bit branch).

The advantage is that when the final code is small enough, the fast 
instructions are used in places where we otherwise would have to 
generate the slower instructions.  We avoid the extra branch, the space 
in the branch predictor, the fact that it is an indirect jump.  We 
don't recover the extra register.