>I didn't know. the answer (as I faintly remember) was because ALU comes before the memory logics in the cpu pipeline. so if this can be solved by math calculations, the instructions can bail out early from the pipeline, hence the execution is faster.

This makes perfect sense and it occurred to me earlier , but I couldn't get neither gcc or clang to do as you said :) As far as I know, there is no way of getting such behavior using "regular" C.

>the second part of the question was asking why using XOR in the actual code (or why XOR is faster than using assignments).

From the benchmarks I performed it seems to me that XOR-ing is definitely slower, but nvidia folks claiming that it is faster makes me wonder if I made a mistake somewhere.

>also the article doesn't mention compiler flags. was optimization on?

Nope. It seems to me that passing -O3 makes the XOR approach behave equivalently to the approach with a tmp variable (https://godbolt.org/z/N_zGjv):)