Re: bug in internal PUSH expanding (splitting)

[Denis Chertykov <denisc at overta dot ru>]

Jeffrey A Law <law@cygnus.com> writes:

> I think something else is wrong.
> 
> I would think you should be referencing the virtual outgoing
> argument pointer,
> not the stack pointer here.  That will be turned into SP + 
> STACK_POINTER_OFFSET.

May be you right. May be no.
My main question: What gcc mean push is ?
1. *(--sp) = arg;
2. *((--sp) + STACK_POINTER_OFFSET) = arg,

In first case you right.
In second no.

gcc don't use virtual_outgoing_args_rtx because on any level above
than emit_move_insn_1 gcc don't check a `movdi' pattern.

> If you look at the c4x port, which has similar characteristics to your port
> you'll see something like this in the .rtl dump:

c4x port have all movXX patterns.
(really movqi and movhi but QImode = 32 bit and HImode = 64 bit)

My port don't have a movdi pattern and gcc try to expand push command
for DImode himself (emit_move_insn_1).

My new example:
void boo (int, ...);
struct {long int lo,hi;} di_struct;
foo ()
{
  boo (1,0x1234567890LL /* number in DImode. Wrong code */);
  boo (1, di_struct /* here gcc generates right code.
                       Here gcc uses another method
		       (sp-=sizeof (di_struct),
                        memcpy (sp + STACK_POINTER_OFFSET, di_struct, ...))
                     */ );
}

IMHO: Solution is in emit_move_insn_1.
old_sp = sp;
sp-=sizeof (arg);
*old_sp-- = arg[n]; /* MSB of arg */
*old_sp-- = arg[3];
*old_sp-- = arg[2];
*old_sp-- = arg[1];
*old_sp-- = arg[0]; /* LSB of arg */

Today's gcc-2.95.2:
sp-=sizeof (arg);
*sp = arg;

With my patch:
sp-=sizeof (arg);
*(sp + STACK_POINTER_OFFSET) = arg;

???

Denis.
PS: Is my bad English a problem for understanding my message ?