I’d say the behavior is mostly historical. Early GPU “compilers” [1] and their corresponding architectures really meant that “an if statement means go slow”. A

I’d say the behavior is mostly historical. Early GPU “compilers” [1] and their corresponding architectures really meant that “an if statement means go slow”. Also, IIRC, the original mask depth on Fermi was ... 32 branches, and afterwards you’d spill. So it was an adaptation to “don’t make the compiler work harder when you can fold the branch for it”. I always found this to be worse for autovectorization with icc. Every once in a while, it would impress you by calculating masks and then blending. Usually it would not. [1] Many of these “compilers” were really just blindly emitting code with no optimizations. The point was it ran correctly, and was effectively autovectorized. If you cared, you could make it faster.