On Wed, Aug 03, 2022 at 09:46:05AM +0800, Jinhao Fan wrote: > at 4:54 PM, Klaus Jensen <i...@irrelevant.dk> wrote: > > > I am unsure if the compiler will transform that division into the shift > > if it can infer that the divisor is a power of two (it most likely > > will be able to). > > > > But I see no reason to have a potential division here when we can do > > without and to me it is just as readable when you know the definition of > > DSTRD is `2 ^ (2 + DSTRD)`. > > OK. I will send a new patch with shifts instead of divisions. BTW, why do we > want to avoid divisions?
Integer division is at least an order of magnitude more CPU cycles than a shift. Some archs are worse than others, but historically we go out of the way to avoid them in a hot path, so shifting is a more familiar coding pattern. Compilers typically implement division as a shift if you're dividing by a a power of two integer constant expression (ICE). This example here isn't an ICE, but it is a shifted constant power-of-two. I wrote up a simple test to see what my compiler does with that, and it looks like gcc will properly optimize it, but only if compiled with '-O3'.
