LLVM trunk emits ~40% larger code at -Oz compared to -O2

[dcci]

Reduced from a large example:

long patatino() {
    long x = 0;
    for (int i = 0; i < 5; ++i) {
        while (x < 37) {
            if (x % 4 == 0) {
                x += 4;
            }
        }
    }
    return x;
}

-Oz

patatino():
        xorl    %eax, %eax
        xorl    %ecx, %ecx
.LBB0_1:
        cmpl    $5, %ecx
        jne     .LBB0_4
        jmp     .LBB0_2
.LBB0_5:
        testb   $3, %al
        jne     .LBB0_4
        addq    $4, %rax
.LBB0_4:
        cmpq    $36, %rax
        jle     .LBB0_5
        incl    %ecx
        jmp     .LBB0_1
.LBB0_2:
        retq

-O2

patatino():
        movl    $40, %eax
        retq

Godbolt link: https://godbolt.org/z/fzxWM693x

[R-Goc]

What is happening here is that this is unoptimized at Oz, don't know why, never looked into Oz, but at O2 the compiler notices that this literally does nothing functional and just returns the value that this function would return. Even if we modify the function, so it operates on an argument passed to it, what happens in it is very stupid, and O2 can optimize it more than Oz, again not sure how Oz works (https://godbolt.org/z/zM6vM6T8z). At O2 the compiler notices the outer for is meaningless, and that if x>36 nothing happens. It also optimizes to reduce the number of jumps by using a cmov.