c/859: sparc-specific pessimal code for simple loop

[Joe Buck]

See the GNATS case c/859 for details.  I was trying to understand the
strange results for the Stepanov abstraction penalty on the Sparc, and
the problem is that we generate very bad code for some simple
fortran-style loops on the Sparc, but the ix86 back end doesn't have
this problem.  I originally saw the problem in C++, but it's in C as well.

The problem code is below.  We get a rather wasteful 8-instruction inner
loop.

--------------------------
void check(double);

void test0(double* first, double* last)
{
  int n, pdiff = last - first;
  double result = 0;
  for (n = 0; n < pdiff; ++n) result += first[n];
  check(result);
}