Why does gcc generate different code for s->a[i] and *(s->a+i)?

I compile this with gcc -O1:

#include <stdint.h>struct mystruct{    uint32_t arr[1];    uint32_t something_after_arr;};uint32_t sum1(const struct mystruct *s,uintptr_t n){    uint32_t result=0;    for(uintptr_t i=0;i<n;++i)result+=s->arr[i];    return result;}uint32_t sum2(const struct mystruct *s,uintptr_t n){    uint32_t result=0;    for(uintptr_t i=0;i<n;++i)result+=i[s->arr];    return result;}uint32_t sum3(const struct mystruct *s,uintptr_t n){    uint32_t result=0;    for(uintptr_t i=0;i<n;++i)result+=*(s->arr+i);    return result;}

compiler (https://godbolt.org/#…) generated this x86_64 assembly (assembler directives removed):

sum1:        mov     eax, 0        test    rsi, rsi        je      .L1        mov     eax, DWORD PTR [rdi].L1:        retsum2:        mov     eax, 0        test    rsi, rsi        je      .L4        mov     eax, DWORD PTR [rdi].L4:        retsum3:        test    rsi, rsi        je      .L10        mov     eax, 0        mov     edx, 0.L9:        add     edx, DWORD PTR [rdi+rax*4]        add     rax, 1        cmp     rsi, rax        jne     .L9.L7:        mov     eax, edx        ret.L10:        mov     edx, 0        jmp     .L7

Only sum3 has backwards jumps. sum1 and sum2 only have forward jump (which means no looping).

Why are sum3 and sum1 different and why no loop in sum1?

I expected sum1 and sum2 and sum3 have same code with loop (like clang (https://godbolt.org/#…)).