I wonder if they've considered using libjpeg-turbo instead of libjpeg9a, which comes with a lot of SIMD optimization and is regarded to be 2x-5x as fast as IJG's implementation.[1]
As far as I understand, these intrinsics map to SSE instructions (they are 128-bit). One could use the later AVX (256-bit, found in server chips since 2011). Probably they decided to start with the lowest common denomintor of SSE, because they don't have AVX in all of their servers? The term SIMD is generic.
Note: we could also choose to use 256-bit wide YMM registers here, but that would only work on the newest CPUs, while gaining little performance.
I'm not saying they're right and you're wrong, but you make it sound as if they never even considered AVX.
Should that be an #ifdef ?
More useful, I think, would be a runtime check far enough up the stack (i.e., away from tight loops) that it doesn't affect performance. Probably doesn't even need to be that high, it'll branch-predict correctly every time but the first iteration, and the binary bloat probably won't hurt your icache because the number of instructions you ever actually execute stays more or less the same.