Re: [avr-gcc-list] Add builtins in avr target.

[Dmitry K.]

On Friday 18 April 2008 12:56, Weddington, Eric wrote:
> > > 2008/4/17, Wouter van Gulik <address@hidden>:
> > >>  For a 2 cycles delays an rjmp can be used. Saves an instruction!
> > >
> > > And so on:
> > >
> > > 3 cycles:
> > >   rjmp .
> > >   nop
> >
> > I shall try to replace 'nop' with 'rjmp .' for two cycle
> > delay.  It is necessary to check that the linker relaxation
> > pass will not remove the  'rjmp .' instruction. Now I do not
> > wish to complicate a code and to optimize 3..6 cycles delays.
>
> AFAIK, linker relaxation will do JMP->RJMP transformations only. Going
> the other way is not a size optimization.

Alas, not only.

Linker relaxation replaces the (R)CALL,RET sequence
into (R)JMP,RET.  Note, the linker does not differ
the C and ASM functions and can spoil the delay routine
like:

  long_delay:
        rcall .
  middle_delay:
        rcall .
  short_delay:
        ret

or may destroy the program where the 'return to next
level' trik is used, like Avr-libc's old (1.4 and
early) float library.

The Binutils-2.17 makes such replacement unconditionally.
The 2.18 has an option to disable it: '--no-call-ret-replacement'.

IMHO, such replacement must be disabled by default.
Note, GCC for some targets can itself optimize tailing call.
This is better, as the second instruction (RET) is deleted.

Regards,
Dmitry.