Yes. And for the same reason, that's why you sometimes have to do things in assembly. No compiler is going to go against its own grain when generating code. Yet, sometimes, that's what you must do to eke out the necessary performance.
As an example, while doing code for a 6809-based video game (see [1] for full war story), I had to use every single register on the processor, including the system-stack register and the direct-page register, to speed up a graphics routine. This meant that, while my code was running, the system effectively had no stack and couldn't use certain addressing modes. Further, it also meant that interrupts would corrupt whatever memory the stack register happened to be pointing to.
All these problems were solvable – and the pain we had to go through to solve them was, for us, absolutely worth the 70% speed-up it bought us – but no compiler writer is going to incorporate tactics this weird into a compiler's optimization logic. If you need to go "full weird," assembly is often the only way.
[1] http://blog.moertel.com/posts/2013-12-14-great-old-timey-gam...