Problem code with Valgrind on ARM

test.c

#include <stdio.h>
#include <stdint.h>

#define N_SUBVALS 4
#define BOUNDARY 4

int main()
{
    
    // The code finds the next boundary alignment for each set of of 4
    // values in the following list, and sums them all together.
    uint32_t vals[] = {10,10,13,13,13,12,13,14,15,16,17,18,19,20,21,22};

    uint32_t counter = 0;
    uint32_t index = 0;

    for (uint32_t i=0; i < 4; ++i){
        
        uint32_t max_subval = 0;

        for (uint32_t j=0; j < N_SUBVALS; ++j){

            uint32_t val = vals[index];

            if (val > max_subval){
                max_subval = val;
            }

            index += 1;
        }

        // should get the next aligned value from the largest of the set 
        // of N_SUBVALS subvalues
        // The max sub vals are: 13, 14, 18 and 22
        // These align to: 16, 16, 20 and 24
        uint32_t aligned_subval = (max_subval + 
                (BOUNDARY - (max_subval % BOUNDARY)) % BOUNDARY);

        // 16 + 16 + 20 + 24 = 76        
        counter += aligned_subval;

    }

    // Should print 76
    //
    if (counter != 76){
        printf("wrong final counter: %d\n", counter);
    } else {
        printf("correct final counter: %d\n", counter);
    }

    return 0;
}

Run and tested with the following shell script:

set -o verbose

# System:
# ARM cortex A9
# valgrind 3.10.1
# gcc 5.2.0
# glibc 2.22
# kernel 3.19.0

gcc -o debug_works.os -c -std=c11 -mfpu=neon -g -Og test.c
gcc -o debug_works debug_works.os
gcc -o release_fail.os -c -std=c11 -mfpu=neon -g -O3 test.c
gcc -o release_fail release_fail.os

# Prints "wrong final counter: 71"
valgrind ./release_fail


# Prints "correct final counter: 76"
valgrind ./debug_works