extern __shared__ unsigned char smem[];
int* cache = (int*)(&smem[0])

Illegal access to address (@global)0x100002d detected

#include <cstdio>

__device__
void myPrint()
{
    extern __shared__ unsigned char cache[];
    int* r1 = (int*)(&cache[0]);
    int* r2 = &r1[3];
    int* r3 = &r2[3];

    for(int i = 0; i < 9; i += 1)
    {
        cache[i] = i;
    }

    printf("%dn", r1[0]);
    printf("%dn", r2[0]);
    printf("%dn", r3[0]);
}

__global__
void childKernel()
{
    myPrint();
}

__global__
void parentKernel()
{
    childKernel<<<1,1, sizeof(int) * 3 * 3>>>(); // 3x3 matrix, row major
    cudaDeviceSynchronize();
}

int main(void)
{

    parentKernel<<<1,1>>>();
    cudaDeviceSynchronize();

    return 0;
}

extern __shared__ int cache[];