I'm wondering why you wouldn't use intrinsics to begin with, since most architectures have multiple load/store operations for vector types and using intrinsics you get to specify which load/store operation you use.

In the case I'm speaking of, I can often handle the vector writes with intrinsics --- GCC, CLang, and ICC will accept unaligned pointers without problem. But the reason I'm using this write strategy (advancing output by a variable) is to avoid branch prediction errors by converting a jump table to a conditional move. Often (half the time?) I can't figure out a way to get the intrinsics to produce the assembly I want, much less across all three of those compilers. Combining the intrinsics with the conditional moves is what requires the inline assembly.

Other places this comes up are addressing modes, operands that work directly on memory, and register allocation. For example, Haswell is capable of doing 2 loads and a store in a single cycle, but only if the store address is a "simple" base plus offset (no indexing). Convincing a compiler to produce a particular addressing mode is somewhat like teaching a blind person to drive a stick shift while you too are wearing a blindfold. Particularly if you are optimizing for a particular processor (and don't need to support Windows), inline assembly is considerably more reliable than depending on a compiler to consistently generate sane code.