On Friday 16 December 2005 02:35, Andreas Schwab wrote:
> Paul Brook <[EMAIL PROTECTED]> writes:
> > In m68k.md we have:
> >
> > ;; Speed up pushing a single byte but leaving four bytes of space.
> >
> > (define_peephole
> > [(set (mem:QI (pre_dec:SI (reg:SI SP_REG)))
> > (match_operand:QI 1 "general_operand" "dami"))
> > (set (reg:SI SP_REG) (minus:SI (reg:SI SP_REG) (const_int 2)))]
> >
> >
> > The (const_int 2) looks very suspect. Shouldn't this be (const_int 3) ?
> > If not, could someone please explain to me what bit of arcane magic I'm
> > missing?
>
> Pushing a byte will actually push a word, i.e "move.b x,-(%sp)" modifies
> %sp by two, not one.
Ah, I see. So all post_inc/pre_dec of stack_pointer_rtx should be modified by
PUSH_ROUNDING?
In dwarf2out.c we use the size of the destination mode without any adjustment:
case POST_DEC:
if (XEXP (src, 0) == stack_pointer_rtx)
{
offset = GET_MODE_SIZE (GET_MODE (dest));
Is this a bug in dwarf2out.c?
If so the peephole is incorrect for ColdFire, where a byte push only modifies
the address register by one. It's very doubtful whether the pattern would
ever match, but still.
Paul
