On Wednesday, 15 January 2014 at 07:08:14 UTC, deadalnix wrote:
Actually this is a useful optimization technique when you need to skip a lot of objects that way. See http://www.onversity.com/load/d-loop.pdf
Well, yes, expect for the fact that the case discussed here isn't an instance of the optimization. Even the "pre-skip" loop performs two conditional jumps per array element. And worse, the form the loop is in causes LLVM to miss two additional optimization opportunities, compared to the plain C (D) version:
First, LLVM actually optimizes away the even/odd conditional branch in the plain version, replacing it with a shl/sar/and combination. The rotated loop structure in the filter version seems to cause the optimizer to miss that opportunity.
And second, as a consequence the loop in the filter version is no longer a good candidate for vectorization, whereas LLVM will happily emit AVX/… instructions in the plain case if you let it.
David
