Re: Passing arguments of function through registers

[Andrew Haley <aph at redhat dot com>]

sumanth wrote:

> I successfully passed function arguments through registers in gcc-3.3.
> Every thing seems okay except there is a reshuffling of registers
> happening once the arguments are passed in registers.
> 
> see the below example code snippet.
> 
> 
> int add(short int x,short int y,short int z)
> {
> return x+y+z;
> }
> main()
> {
> int a=5,b=6,c=7,d;
> d=add(a,b,c);
> }
> 
> and the objdump of add function is
> 00000134 <_add>:
> 134:   c3 30           000030c3     mov        r3,r0
> 136:   c3 01           000001c3     mov        r0,r1
> 138:   c3 12           000012c3     mov        r1,r2
> 13a:   00 0c           00000c00     add        r0,r3,r0
> 13c:   80 00           00000080     add        r0,r0,r1
> 13e:   04 08           00000804     rts
> 
> 
> The arguments are clearly passed in to r0,r1,r2 but my compiler
> reshuffled them to r3,r0,r1
> my return value of a function goes to r0. since i declared r0-r3 as 1 in
> CALL_USED_REGISTERS ,r0-r3
> are not pushed in stack.
> 
> Can any one help me to avoid the reshuffling of registers ..the idle
> case should be
> 00000134 <_add>:
> 134:  add  r0, r0,r1  ==> r0=r0+r1
> 138:  add r0, r0, r2==> r0 = r0+r2
> 13c:   rts
> 
> since the required values are already in r0,r1,r2.

We'd need to know what processor this is, and what optimization options
you used.

> PS: I am using gcc-3.3

That's very bad.  Even if there is a bug in that compiler, it's
so very old that it's unlikely to be fixed.

Andrew.