optimization/3384: PowerPC -funroll-loops -fbranch-count-reg misoptimization

[Zoltan Hidvegi]

This may be related to PR 3384.  The following little function
generates this code with gcc -O2 (note that -O2 turns on
-fbranch-count-reg):

void
do_loop(unsigned long c, char *m)
{
    unsigned long i = 0;

    do {
	m[i] = 0;
    } while (++i != c);
}

The assembly on powerpc-ibm-aix4.3.3.0 with gcc-3.0 -O2:

        mr. 3,3
        li 9,0
        li 0,0
        beq- 0,L..8
L..7:
        mtctr 3
L..6:
        stbx 0,4,9
        addi 9,9,1
        bdnz L..6
        blr
L..8:
        li 3,1
        b L..7

So it checks if c==0, and in that case it only executes the loop once,
even though the loop should be executed 2^32 times, which is probably
not wat the caller wants, but still the program should just do what I
told in the source.  The correct code would be

        mtctr 3
        li 9,0
        li 0,0
L..6:
        stbx 0,4,9
        addi 9,9,1
        bdnz L..6
        blr