[juliusgeo]

I was considering doing it all in a ternary statement, but I feel that the current form is also branchless because it is simply a multiply and add. The extra bounds-checking condition can probably be omitted, but I haven't tested that.

  for (step >>= 1; step != 0; step >>=1) {
          if ((next = begin + step) < size) {
              begin += PyObject_RichCompareBool(PyList_GetItem(list_obj, next),     value, Py_LT) * step;
          }
      }

[yarg]

My point was, anywhere there's a hidden 'if' can be branching.

If there's no calculation being done, it'll simplify.

    value = (test) ? const0 : const1;

But if calculations are being done, it won't.

    value = (test) ? calc0() : calc1();

If you want non-branching where the ternary options are calculated, you need to calculate both.

This matters most with SIMD operations.

Look at section 2.5.1 (Branch-Equivalent SIMD Processing)

http://ftp.cvut.cz/kernel/people/geoff/cell/ps3-linux-docs/C...

[juliusgeo]

Ah, yeah I see what you mean. If I'm understanding you correctly, the fact that we are calling the Python interpreter internal functions during that calculation makes it branch because it is not pre-calculated?

[yarg]

Pretty much (at least as far as I understand it).

There's probably something at the instruction level which allows the constant ternary expressions to be non-branching.