[Bug middle-end/23090] [4.0/4.1 Regression] gcc.c-torture/execute/20050713-1.c -Os fails

["dje at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

------- Additional Comments From dje at gcc dot gnu dot org  2005-07-27 22:03 -------
The testcase passes three structs with three members each for a total of nine
words.  AIX, Darwin, and PPC64 Linux pass structs by value in GPRs r3-r10
allowing a maximum of 8 words.  The ninth argument is passed on the stack.  This
bug appears to be a problem with either argument passing or scheduling.

struct S a = { 3, 4, 5 }, b = { 6, 7, 8 }, c = { 9, 10, 11 };

main() calls baz3 (c, a, b) and the arguments are correct

9, 10, 11, 3, 4, 5, 6, 7 in GPRs and 8 on stack at 56(r1)

baz3 (struct S x, struct S y, struct S z)
{
  return foo3 (y, z, x);
}

is compiled as:

        addi r2,r1,24
        addi r11,r1,36
        stswi r3,r2,12 <== store baz3 argument x to r1+24,28,32
        addi r2,r1,48  <== address of foo3 second argument
        stswi r6,r11,12
        lwz r0,32(r1)  <== load x.c
        stw r9,48(r1)
        mr r9,r3
        stw r10,52(r1)
        stw r0,56(r1)  <== store x.c to third argument last member on stack
        lwz r10,28(r1)
        lswi r3,r11,12
        lswi r6,r2,12  <== load foo3 second argument from r2+48,52,56
        b _foo3

GCC stores the argument passed on the stack for foo3 before loading the second
argument from the same location.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23090