Re: commutative asm operands

[Alan Modra <amodra at bigpond dot net dot au>]

On Mon, Aug 05, 2002 at 07:16:10AM +0930, Alan Modra wrote:
> On Sun, Aug 04, 2002 at 06:29:19PM +0200, Roman Zippel wrote:
> > Alan Modra wrote:
> > >         * recog.c (constrain_operands): Handle commutative operands.
> > 
> > What exactly does this patch fix?
> 
> The ICE reported at the beginning of this thread, and the following
> similar one.
> 
> > What is the assembly output for this test case (register names might
> > need a change):
> 
> Thanks, this testcase is useful.
> 
> void f (void)
> {
>   register int a asm ("%r3");
>   register int b asm ("%r4");
>   register int c asm ("%r5");
>   register int d asm ("%r6");
> 
>   asm volatile ("addc %0, %2, %3\n\tadde %1, %4, %5"
> 		: "=r" (a), "=r" (b)
> 		: "%0" (a), "r" (c), "%1" (b), "r" (d));
>   asm volatile ("addc %0, %2, %3\n\tadde %1, %4, %5"
> 		: "=r" (a), "=r" (b)
> 		: "%0" (c), "r" (a), "%1" (d), "r" (b));
> }
> 
> produces, with my patch:
> 
> .f:
> #APP
>         addc 3, 3, 5
>         adde 4, 4, 6
>         addc 3, 5, 3
>         adde 4, 4, 6
> #NO_APP
>         blr

and with yours:

.f:
#APP
        addc 3, 3, 5
        adde 4, 4, 6
        addc 3, 3, 5
        adde 4, 4, 6
#NO_APP
        blr

I suspect your fix would never generate the "addc 3,5,3" case, which
is probably a good thing in that the rest of the compiler never needs
to worry about "%" constraints.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre