Poor strength reduction with autoincrements

[Michael Hayes <m dot hayes at elec dot canterbury dot ac dot nz>]

The one area that I'm finding that egcs is consistently generating
poor code (compared with gcc-2.7.2) is strength reduction for machines
with autoincrement addressing.  One would think that strength
reduction should improve the performance of autoincrement addressing
but I'm finding that I often get better code for the c4x by disabling
it.

Some of the problems stem from a poor choice of GIVs to combine.
Richard Henderson posted a patch a few months ago that fixed many of
these problems but this patch has bever been committed.

Another problem occurs when a new GIV is unecessarily created and
where the GIV it was derived from is used after the loop.  This leads
to a redundant increment within the loop and increased register
pressure.  The following contrived function demonstrates this:


int foo(unsigned short *dest, const unsigned char *src, int bytes)
{
  int i;
  
  for (i = 0; i < bytes; i++)
    {
      unsigned int tmp;
 
      tmp = *src++ & 0xff;
      tmp |= *src++;
      *dest++ = tmp;
    }
  return *src;
}

The generated code for the c4x is:

	<snip...>
	lda	ar1,ar0		<= new GIV initialised
L6:
	ldiu	*ar0,r0
	and	255,r0
	or3	*+ar0(1),r0,r0
	addi3	2,ar0,ar0	<= new GIV incremented
	addi3	2,ar1,ar1	<= old GIV incremented
	addi3	-1,r1,r1
	bned	L6
	sti	r0,*ar2++
	nop
	nop
L4:
	ldiu	*ar1,r0
	rets

Note that the new and old GIVs are equivalent and thus the new one is
redundant.

Does anyone have any suggestions where I should look in loop.c to fix
this problem?

Michael.