Re: [Amforth] Compiling Amforth 4.9 - forward labels in macros

[Marcin C. <sa...@sa...>]

On Fri, 17 Aug 2012, Erich Waelde wrote:

> Hello Marcin,
> 
> I have made tests
> 1. compile avra from git
> commit a6e8b2957953810dae6467eeb4905bfc5ea6c33e
> Author: Marcin Cieslak <sa...@sa...>
> Date:   Wed Aug 8 17:01:30 2012 +0200

> Does this help?

Yes, thank you.

It seems the problem described in this thread:

http://thread.gmane.org/gmane.comp.lang.forth.amforth/1353/focus=1354

is related to calculation of forward referenced labels in a macro.


Consider the following code (correctly assembled):

                .device atmega32
          PFA_COLD:
C:000000 c004           rjmp    PFA_DOPLUSLOOP3
C:000001 0000           nop
C:000002 0000           nop
C:000003 0000           nop
C:000004 0000           nop
          PFA_DOPLUSLOOP3:      
C:000005 9508           ret
Used memory blocks:
   code      :  Start = 0x0000, End = 0x0005, Length = 0x0006 (6 words), Overlap=N


Segment usage:
   Code      :         6 words (12 bytes)
   Data      :         0 bytes
   EEPROM    :         0 bytes


But this:

                .device atmega32
                  .macro jmp_
                         ; a more flexible macro
                        .ifdef @0
                        .if (@0-pc > 2040) || (pc-@0>2040)
                          jmp @0
                          .else   
                          rjmp @0
                          .endif
                          .else
                          jmp @0
                          .endif
                  .endmacro
          PFA_COLD:
C:000000   +    jmp_    PFA_DOPLUSLOOP3
         

          .ifdef PFA_DOPLUSLOOP3
          .if (PFA_DOPLUSLOOP3-pc > 2040) || (pc-PFA_DOPLUSLOOP3>2040)
:000000 c005      rjmp PFA_DOPLUSLOOP3
          .endif
          .else
C:000001 0000           nop
C:000002 0000           nop
C:000003 0000           nop
C:000004 0000           nop
          PFA_DOPLUSLOOP3:      
C:000005 9508           ret
Used memory blocks:
   code      :  Start = 0x0000, End = 0x0006, Length = 0x0007 (7 words), Overlap=N


Segment usage:
   Code      :         7 words (14 bytes)
   Data      :         0 bytes
   EEPROM    :         0 bytes

Assembly completed with no errors.



Generated jump goes one word too far (it is "jump 5 words forward"
and not "jump 4 words forward"). It can even be seen
that the length of a macro expansion is wrongly estimated
(it says the whole code is 7 words, while it is only 6).

avrasm2 generally does not allow using not-yet-defind
labels in macros like this.

There are two forward jumps in the code of amforth:

1. There is no problem with "jmp_ PFA_COLD" since the 
distance to PFA_COLD does to depend on the size of the
macro expanded.

2. Only "jmp_ PFA_DOPLUSLOOP3" is problematic since the label 
address depends on the size of the macro - and this can
change .. depending on the target address of the label.

Amforth 4.4 works because it had "rjmp PFA_DOPLUSLOOP3"
hardcoded and not jmp_ macro.


So there are two workarounds for amforth:

1) revert to amforth 4.4 hardcoded "rjmp" since we can
   assume the distance is pretty small here.

2) add a nop after "rjmp" in the "jmp_" macro to keep
   sizes of the macro expansion constant.


Solutions:

1) AVRA may disallow forward refences in the macros.

2) Detecting the impossibility to determine
   the value.

3) ...

//Marcin