Possible bug in PowerPC structure passing

[Stan Shebs]

Apple is switching from using EGCS 1.1.2 to GCC 2.95.2 for building
MacOS X, and has encountered a number of regressions.  One of them
is that passing medium-sized structures by value doesn't work,
because memcpy's result steps on the save area (which then causes
OS X networking stack to fall over).  I've included a small test
case below.  I'm not 100% certain this is a GCC bug, because we
have a lot of local mods and no AIX systems to test with; however,
if I'm on a PowerPC Linux system and compile the test case with

   gcc -mcall-aix -O -c dek.c
   gcc dek.o

then 1.1.2 (the default on my Linux system) wins, while 2.95.2 and
yesterday's checkout both lose.  You need -mcall-aix to emulate the
MacOS calling convention, -mcall-sysv doesn't seem to have a problem.

So if some kind person could try this out on a real AIX system and
report back, I'd be very grateful.

Turly O'Connor <turly@apple.com> has worked up a skanky hack
to calls.c that gets our networking people going again, but it's
pretty ugly, so I won't pollute this list unless asked. :-)

Stan

dek.c:

typedef struct {
  char y;
  char x[32];
} X;

int z (void)
{
  X xxx;
  xxx.x[0] =
  xxx.x[31] = '0';
  xxx.y = 0xf;
  return f (xxx, xxx);
}

int main (void)
{
  extern int printf (const char *, ...);
  int val;

  val = z ();
  printf ("f(xxx) returns '%c' (==0x%x) [should be 0x60]\n", val, val);
  return 0;
}

int f(X x, X y)
{
  if (x.y != y.y)
    return 'F';

  return x.x[0] + y.x[0];
}