[jallmann]

Pedantic: "n>2" is a comparison operator, and doesn't incur a branch. if statements, ternary operators, etc are conditionals. The effect in the article is the same though, without the potential for a mispredict penalty.

Whether that's actually faster is debatable though. If you're writing in C, you're probably concerned about performance, and the function pointer invocation is going to be a lot slower than a branch mispredict. A sufficiently complex array dereference will slow it down too, but I wouldn't be surprised if the compiler can turn simple cases like this into an equivalent switch/jumptable.

Thus conditional-free programming (as presented here, in C) is mostly academic. Debug tracing is an interesting application, but readability and performance both suffer.

If a language could internalize conditional-free constructs natively, however, it could be an interesting study.

[adrusi]

`n > 2` could be a comparison operator in some architectures. For example, in dcpu16 (fictional, but I'm sure this was inspired from somewhere), `a = b > 2` would be written as:

    set A, 0
    ifg B, 2
      set A, 1

It conditionally evaluates a machine instruction, which is, indeed, branching.

[JOnAgain]

>> Thus conditional-free programming (as presented here, in C) is mostly academic

Do you think this is true because C code has traditionally relied so heavily on branching?

Put another way, perhaps branch-free programming is has greater speed potential, but the tools we use have been heavily optimized for branch-laden programming. Maybe putting sufficient resources into optimizing code for branch-free programming (and stripping out optimizations for branching) would yield faster execution.

I don't know the answer, just musing. I am curious, though.

[wtracy]

Well, the "branch-free" paradigm requires pushing a stack frame and making a jump to the passed function (requiring a pointer lookup) and a jump back to the calling function. In the branching paradigm, only one execution path requires a jump at all, and that jump is made based on an offset embedded directly into the machine code, so the additional pointer lookup is not needed.

So, I'm going to come down on the side of the traditional model being inherently more performant.