Code demonstrating hand-crafted and compiler-generated migration of redundant loads/stores out of a loop

minimize-loadstore.c

void nonoptimizable_accumulate(int *accumulator, int const *array, int length) {
    for (int i = 0; i < length; i++) {
        *accumulator += array[i];
    }
}

void optimizable_accumulate(int *restrict accumulator, int const *restrict array, int length) {
    for (int i = 0; i < length; i++) {
        *accumulator += array[i];
    }
}

minimize-loadstore_arm64.asm

// -O1

nonoptimizable_accumulate:
        cmp     w2, 0
        ble     .L1
        mov     x3, x1
        add     x4, x1, w2, sxtw 2
.L3:
        ldr     w1, [x0]
        ldr     w2, [x3], 4
        add     w1, w1, w2
        str     w1, [x0]
        cmp     x3, x4
        bne     .L3
.L1:
        ret
optimizable_accumulate:
        cmp     w2, 0
        ble     .L5
        ldr     w4, [x0]
        mov     x3, x1
        add     x2, x1, w2, sxtw 2
.L7:
        ldr     w1, [x3], 4
        add     w4, w4, w1
        cmp     x3, x2
        bne     .L7
        str     w4, [x0]
.L5:
        ret

minimize-loadstore_handcrafted.c

void optimized_accumulate(int *restrict accumulator, int const *restrict array, int length) {
    int local_accumulator = *accumulator;
    for (int i = 0; i < length; i++) {
        local_accumulator += array[i];
    }
    *accumulator = local_accumulator;
}

minimize-loadstore_x86-64.asm

# -O1

nonoptimizable_accumulate:
        testl   %edx, %edx
        jle     .L1
        movq    %rsi, %rax
        movslq  %edx, %rdx
        leaq    (%rsi,%rdx,4), %rcx
.L3:
        movl    (%rax), %edx
        addl    %edx, (%rdi)
        addq    $4, %rax
        cmpq    %rcx, %rax
        jne     .L3
.L1:
        ret
optimizable_accumulate:
        testl   %edx, %edx
        jle     .L5
        movl    (%rdi), %ecx
        movq    %rsi, %rax
        movslq  %edx, %rdx
        leaq    (%rsi,%rdx,4), %rsi
.L7:
        addl    (%rax), %ecx
        addq    $4, %rax
        cmpq    %rsi, %rax
        jne     .L7
        movl    %ecx, (%rdi)
.L5:
        ret