When the arrays given to std.numeric.dotProduct are very small, its performance compared to an inlined foreach loop becomes quite bad (a simple benchmarks on request).

This is a hard problem to solve in general (you probably need profile-driven optimizations, that maybe future LLVM versions will perform with -O5 or -O6 level optimization or more), but there is an important special case where I think there is a way to improve the situation:


int[3] a, b;
auto r = dotProduct(a, b);


In this situation dotProduct may just recognize (with template constraints) that both a and b are fixed-sized arrays, that their length is the same (so no need to test it again at runtime), and that this length is small (less than 6? 8?). In this case a function template overload for dotProduct performs a "static foreach" on the items of a and b (in theory you are also allowed to use inline asm with few instructions that use SIMD registers, but currently with DMD this kills inlining, so it's not good enough). I think in this case DMD will *inline* this very simple dotProduct function overload, allowing good enough performance.

This low performance problem has caused problems in my code.

See also bug 4393