I neglected to turn off cyclic gc while performing the timings. Doing so, the overhead is measured at about 5%. I've also tried some experiments with traversing a binary tree. For very small trees the yield-from version starts off slightly slower overall than the for-loop version. Break-even occurs at about 31 nodes in the tree, and after that the yield-from version steadily gains ground, until at 1 million nodes it's 1.6 times faster. The columns below are: depth of the tree, for-loop time, yield-from time and the ratio of the times (> 1 meaning yield-from is faster). 1 0.000049 0.000060 0.805263 2 0.000051 0.000072 0.706534 3 0.000069 0.000108 0.637564 4 0.000130 0.000148 0.878364 5 0.000265 0.000253 1.04675 6 0.000457 0.000458 0.997223 7 0.000988 0.000920 1.07379 8 0.002025 0.001743 1.162 9 0.003985 0.003437 1.15947 10 0.008345 0.006848 1.2185 11 0.017398 0.013902 1.25145 12 0.036223 0.027844 1.30091 13 0.075175 0.055327 1.35874 14 0.154206 0.110290 1.39819 15 0.318121 0.220421 1.44324 16 0.663929 0.442529 1.50031 17 1.369077 0.891843 1.53511 18 2.827645 1.798144 1.57254 19 5.883996 3.661280 1.60709 20 12.093364 7.424631 1.62882 def forloop(node): if node == 1: yield node else: for x in forloop(node[0]): yield x for x in forloop(node[1]): yield x def yieldfrom(node): if node == 1: yield node else: yield from yieldfrom(node[0]) yield from yieldfrom(node[1]) -- Greg