On Wednesday, 21 December 2016 at 17:49:22 UTC, kinke wrote:
> Basic stuff such as this is appropriately tested. The named return value optimization is enforced by D (incl. unoptimized builds), so behavior doesn't change by this optimization. It's you who changed the behavior by removing the if. Due to the `if`, the compiler doesn't know whether it should construct `a` or `b` directly into the memory (sret pointee) provided by the caller. When omitting the `if`, it's clear that `b` is returned in all cases, so the compiles constructs `a` on the local stack (and destructs it before exiting the function), but emplaces `b` into `*sret` (i.e., the caller's stack) and can thus elide 1x postblit + 1x dtor.

Thanks a lot. It makes sense. It seemed just weired that a conditional return value causes such a change. But I begin to understand the background.