H8300H Tiny

[Torbjörn Lindh <toobii@gopta.se>]

Hitachi have generated a number of so called Tiny H8:s, for instance the 
H8/3664F. These have an h8300h cpu but are limited to a 64k address 
space.

The problem with gcc-3.1 (and as far as I can see, the current cvs version
too) is that if I use h8300 mode, I don't take advantage of the added
instructions, adressing modes etc of the h8300h cpu. But if I use the
h8300h mode the addressing gets far to long: 

	subs	#0x4,er7
	mov.l	er7,@-er7

should really be

	subs	#0x4,er7
	mov.w	r7,@er7

and a piece of code such as:

unsigned char a[5];

void f(void)
{
  unsigned char i;

  i = 0;
  while (i < 5) {
    if (a[i])
      whatever();
    i++;
  }
}

generates

	sub.b  r3l,r3l
l0	mov.b  r3l,r2l
	extu.w r2
	mov.w  r2,r0
	extu.l er0
	mov.l  er0,er1
	add.l  #0xfffbd8,er1
	mov.b  @er1,r2l
	beq    l1
	jsr    @whatever
l1	add.b  #0x1,r3l
	cmp.b  #0x5,r3l
	bls	l0

whereas I would prefer

        sub.b  r3l,r3l
l0      mov.b  r3l,r2l
        extu.w r2   
        mov.w  r2,r0
        mov.w  r0,r1
        add.w  #0xfbd8,r1	* Changed
        mov.b  @er1,r2l
        beq    l1   
        ...
l1      add.b  #0x1,r3l
        cmp.b  #0x5,r3l
        bls     l0

(or maybe even save a few register moves, but that is beyond this scope).

This might not seem like a big difference, but when you only have 32k for
your code, every wasted word is bad.

I have looked at the code, and one possible way seems to be to define a 
new target that differs slightly from the h8300h target.

Does anyone have alternative suggestions how I should go about this?