The last FFT convolution stage can often be smaller due to covering a shorter portion of the impulse-response buffer. A stage of the same size but with half the FFT size needs to perform the FFT twice as often but still theoretically wins by a log(n) reduction. That may not be much, but I expect there to also be improvements in cache hits through iterating over smaller buffers. Reducing the FFT size can also improve accuracy, including reducing the period of influence of any numerical errors or non-finite bad values in the input (see bug 1223520).

Kind of tested by https://hg.mozilla.org/mozilla-central/rev/51287ab3d66a