Why does GCC generate a faster program than Clang in this recursive Fibonacci code?

GCC 4.9.2 in compiler explorer really does loop-unrolling and inlines a lot of function calls while Clang 3.5.1 calls fib twice each iteration without even tail call optimization like below

fib(int):                                # @fib(int)
        push    rbp
        push    rbx
        push    rax
        mov     ebx, edi
        cmp     ebx, 2
        jge     .LBB0_1
        mov     eax, ebx
        jmp     .LBB0_3
.LBB0_1:
        lea     edi, dword ptr [rbx - 1]
        call    fib(int)       # fib(ebx - 1)
        mov     ebp, eax
        add     ebx, -2
        mov     edi, ebx
        call    fib(int)       # fib(ebx - 2)
        add     eax, ebp
.LBB0_3:
        add     rsp, 8
        pop     rbx
        pop     rbp
        ret

The GCC version is more than 10 times longer, with only a single fib call and 20+ labels for inlining the call, which also means that the last call has been tail-optimized into a jmp or GCC has converted some of the recursion into iteration (since it allocates a big array for storing intermediate values)

I've also brought ICC into perspective, and surprisingly it has 10 call instructions inside fib, and it also inlines fib calls 9 times inside main, but it doesn't convert the recursive code to iterative

Here's the compiler outputs for comparison

Note that you can modify the code like this to make the output easier to read

int fib(int n) {
    if (n<2) return n;
    int t = fib(n-1);
    return t + fib(n-2);
}

Now compiler explorer will highlight which source code line an instruction in the assembly output corresponds to with distinct colors, and you'll easily see how the two calls are made. The line return t + fib(n-2) is compiled by GCC to

.L3:
        mov     eax, DWORD PTR [rsp+112]  # n, %sfp
        add     edx, DWORD PTR [rsp+64]   # D.35656, %sfp
        add     DWORD PTR [rsp+60], edx   # %sfp, D.35656
        sub     DWORD PTR [rsp+104], 2    # %sfp,

I wouldn't say that gcc beats clang by miles. In my opinion, the performance difference (6.3 seconds vs 9 seconds) is rather small. On my FreeBSD system, clang requires 26.12 seconds and gcc requires 10.55 seconds.

However, the way to debug this is to use g++ -S and clang++ -S to get the assembly output.

I tested this on my FreeBSD system. The assembly language files are too long to post here, but it appears that gcc performs multiple levels of inlining in the Fibonacci calculation function (there were 20 fib() calls in there!) whereas clang simply calls fib(n-1) and fib(n-2) with no levels of inlining.

By the way, my gcc version was 4.2.1 20070831 patched [FreeBSD] and clang version was 3.1 (branches/release_31 156863) 20120523. These were the versions that come with the FreeBSD 9.1-RELEAESE base system. The CPU is AMD Turion II Neo N40L Dual-Core Processor (1497.54-MHz).