How do I delete redundant prologue/epilogue instructions?

How do I delete redundant prologue/epilogue instructions?

[Daniel Towner]

Hi all,

I'm trying to overcome an optimisation problem in my port of gcc. When 
the prologue and epilogue are generated as RTL lists, the compiler 
notices that certain registers are live, and emits save/restore 
sequences for them, with the appropriate stack adjustments. Later on, 
register renaming is able to change the registers used by the function 
from callee-save to caller-save, removing any need for the 
save/restore/stack adjust code in the prologue/epilogue. However, the 
instructions which perform these operations are still emitted.

How do I delete these redundant instructions? Can anyone point me at 
another port which solves this problem?

thanks,

dan.

============================================================================
Daniel Towner
picoChip Designs Ltd., Riverside Buildings, 108, Walcot Street, BATH,
BA1 5BG
daniel.towner@picochip.com
07786 702589 

Re: How do I delete redundant prologue/epilogue instructions?

[James E Wilson]

Daniel Towner wrote:
> register renaming is able to change the registers used by the function
> from callee-save to caller-save, removing any need for the
> save/restore/stack adjust code in the prologue/epilogue. However, the
> instructions which perform these operations are still emitted.

I doubt that any port does this optimization.  You would have to redo
the bulk of reload: register elimination, assigning stack slot
addresses, recomputing frame size and register save offsets, etc.  Gcc
isn't set up to do this.

In the simplest case, if you have no locals, no frame pointer, no
alloca, no outgoing args on the stack, etc, then a simple RTL scan of
the function may prove that SP/FP/AP are never used, allowing us to
remove code in the prologue and epilogue.  However, I'm skeptical that
this would occur often enough to be useful.  And it would require some
target dependent code to figure out how to do the rewrites, which may
make this complicated.

It is probably easier to try to avoid this situation, by getting better
register allocation earlier.  Make sure that the caller-save registers
occur before the callee-save registers in REG_ALLOC_ORDER.  Verify that
the caller-save.c code which is used in reload is working for your
target.  We could try moving some of the regrename optimizations to
before reload.
-- 
Jim Wilson, GNU Tools Support, http://www.specifix.com