zinc's "ioregs" macro system takes care of that sort of optimization quite nicely these days, actually. For each peripheral register, one of the things it generates is an "update" structure that the optimizer can eliminate, which allows you to chain updates to multiple fields in a single write. For example, in my stm32f4-discovery "blinky" example it does the bsrr writes to flip 2 LEDs in a single write (I've inspected the generated asm to verify this).
EDIT: nevermind, I realize I misunderstood; you're referring to the optimizer being able to see that both reg addresses have the same high 16 bits. I thought you were referring to the actual interactions with the register contents. Yea, that makes sense.
EDIT 2: for what it's worth, i did a quick test just now and found that moving the addresses into the source actually made the generated code significantly worse. I'm just tinkering around on a lunch break and I need to get back to real work, so I can't investigate too deeply, but the 2 biggest factors I see are that (1) it simply doesn't perform the optimization you describe, and (2) it makes things even worse because rather than addressing relative to the peripheral base address it loads a complete new constant address for every operation.