Change of assumptions/behavior with -ftree-vrp in gcc 4.8+?

[Vikram Mulukutla]

Hi,

The program listed below seems to invoke optimization behavior that 
produces different results pre 4.8 and 4.8+ versions of gcc. Using the 
-fno-tree-vrp option makes things consistent again. I make no claim of 
an understanding of what this flag really does.

The program is badly written since the first operand to && in the while 
loop conditional will cause an invalid array access. I'm assuming that 
optimization behavior is unpredictable when the assumption that there 
will be no invalid array accesses is broken (and thus execution/output 
is untrustworthy), but I am curious as to why there is a change with 
4.8+. The only concern I'd have is if this change may somehow affect 
other, correctly written code.

# gcc 4.8 produces this:
$ gcc-4.8 -O2 fk2.c && ./a.out
i after loop (array size is 5): 5
found at 5, answer is 0

# pre gcc 4.8 produces this:
$ gcc-4.4 -O2 fk2.c && ./a.out
i after loop (array size is 5): 5
This should be printed.
found at 4, answer is 6

$ gcc-4.6 -O2 fk2.c && ./a.out
i after loop (array size is 5): 5
This should be printed.
found at 4, answer is 6

Program:

#include <stdio.h>

#define ARRAY_SIZE(x) (sizeof(x)/sizeof(x[0]))

unsigned int numbers[] = {
        10,
        9,
        8,
        7,
        6,
};

/* find the element closest to elem */
int main()
{
        unsigned int elem = 0;
        unsigned int i = 0;

        while ((elem < numbers[i]) && (i < ARRAY_SIZE(numbers)))
                        i++;

        printf("i after loop (array size is %lu): %d\n", 
ARRAY_SIZE(numbers), i);

        if (i == ARRAY_SIZE(numbers)) {
                printf("This should be printed.\n");
                i--;
        }

        printf("found at %d, answer is %u\n", i, numbers[i]);

        return 0;
}

Thanks,
Vikram