Can I use GCC's __builtin_expect() with ternary operator in C

Question

The GCC manual only shows examples where __builtin_expect() is placed around the entire condition of an \'if\' statement.

I also noticed that GCC does not complain if I

Answer 1

It apparently works for both ternary and regular if statements.

First, let's take a look at the following three code samples, two of which use __builtin_expect in both regular-if and ternary-if styles, and a third which does not use it at all.

builtin.c:

int main()
{
    char c = getchar();
    const char *printVal;
    if (__builtin_expect(c == 'c', 1))
    {
        printVal = "Took expected branch!\n";
    }
    else
    {
        printVal = "Boo!\n";
    }

    printf(printVal);
}

ternary.c:

int main()
{
    char c = getchar();
    const char *printVal = __builtin_expect(c == 'c', 1) 
        ? "Took expected branch!\n"
        : "Boo!\n";

    printf(printVal);
}

nobuiltin.c:

int main()
{
    char c = getchar();
    const char *printVal;
    if (c == 'c')
    {
        printVal = "Took expected branch!\n";
    }
    else
    {
        printVal = "Boo!\n";
    }

    printf(printVal);
}

When compiled with -O3, all three result in the same assembly. However, when the -O is left out (on GCC 4.7.2), both ternary.c and builtin.c have the same assembly listing (where it matters):

builtin.s:

    .file   "builtin.c"
    .section    .rodata
.LC0:
    .string "Took expected branch!\n"
.LC1:
    .string "Boo!\n"
    .text
    .globl  main
    .type   main, @function
main:
.LFB0:
    .cfi_startproc
    pushl   %ebp
    .cfi_def_cfa_offset 8
    .cfi_offset 5, -8
    movl    %esp, %ebp
    .cfi_def_cfa_register 5
    andl    $-16, %esp
    subl    $32, %esp
    call    getchar
    movb    %al, 27(%esp)
    cmpb    $99, 27(%esp)
    sete    %al
    movzbl  %al, %eax
    testl   %eax, %eax
    je  .L2
    movl    $.LC0, 28(%esp)
    jmp .L3
.L2:
    movl    $.LC1, 28(%esp)
.L3:
    movl    28(%esp), %eax
    movl    %eax, (%esp)
    call    printf
    leave
    .cfi_restore 5
    .cfi_def_cfa 4, 4
    ret
    .cfi_endproc
.LFE0:
    .size   main, .-main
    .ident  "GCC: (Debian 4.7.2-4) 4.7.2"
    .section    .note.GNU-stack,"",@progbits

ternary.s:

    .file   "ternary.c"
    .section    .rodata
.LC0:
    .string "Took expected branch!\n"
.LC1:
    .string "Boo!\n"
    .text
    .globl  main
    .type   main, @function
main:
.LFB0:
    .cfi_startproc
    pushl   %ebp
    .cfi_def_cfa_offset 8
    .cfi_offset 5, -8
    movl    %esp, %ebp
    .cfi_def_cfa_register 5
    andl    $-16, %esp
    subl    $32, %esp
    call    getchar
    movb    %al, 31(%esp)
    cmpb    $99, 31(%esp)
    sete    %al
    movzbl  %al, %eax
    testl   %eax, %eax
    je  .L2
    movl    $.LC0, %eax
    jmp .L3
.L2:
    movl    $.LC1, %eax
.L3:
    movl    %eax, 24(%esp)
    movl    24(%esp), %eax
    movl    %eax, (%esp)
    call    printf
    leave
    .cfi_restore 5
    .cfi_def_cfa 4, 4
    ret
    .cfi_endproc
.LFE0:
    .size   main, .-main
    .ident  "GCC: (Debian 4.7.2-4) 4.7.2"
    .section    .note.GNU-stack,"",@progbits

Whereas nobuiltin.c does not:

    .file   "nobuiltin.c"
    .section    .rodata
.LC0:
    .string "Took expected branch!\n"
.LC1:
    .string "Boo!\n"
    .text
    .globl  main
    .type   main, @function
main:
.LFB0:
    .cfi_startproc
    pushl   %ebp
    .cfi_def_cfa_offset 8
    .cfi_offset 5, -8
    movl    %esp, %ebp
    .cfi_def_cfa_register 5
    andl    $-16, %esp
    subl    $32, %esp
    call    getchar
    movb    %al, 27(%esp)
    cmpb    $99, 27(%esp)
    jne .L2
    movl    $.LC0, 28(%esp)
    jmp .L3
.L2:
    movl    $.LC1, 28(%esp)
.L3:
    movl    28(%esp), %eax
    movl    %eax, (%esp)
    call    printf
    leave
    .cfi_restore 5
    .cfi_def_cfa 4, 4
    ret
    .cfi_endproc
.LFE0:
    .size   main, .-main
    .ident  "GCC: (Debian 4.7.2-4) 4.7.2"
    .section    .note.GNU-stack,"",@progbits

The relevant part:

Basically, __builtin_expect causes extra code (sete %al...) to be executed before the je .L2 based on the outcome of testl %eax, %eax which the CPU is more likely to predict as being 1 (naive assumption, here) instead of based on the direct comparison of the input char with 'c'. Whereas in the nobuiltin.c case, no such code exists and the je/jne directly follows the comparison with 'c' (cmp $99). Remember, branch prediction is mainly done in the CPU, and here GCC is simply "laying a trap" for the CPU branch predictor to assume which path will be taken (via the extra code and the switching of je and jne, though I do not have a source for this, as Intel's official optimization manual does not mention treating first-encounters with je vs jne differently for branch prediction! I can only assume the GCC team arrived at this via trial and error).

I am sure there are better test cases where GCC's branch prediction can be seen more directly (instead of observing hints to the CPU), though I do not know how to emulate such a case succinctly/concisely. (Guess: it would likely involve loop unrolling during compilation.)