Agner`s CPU blog

[...]
Well, that's a shame. I've been looking at some 80x86 disassembly of a C .exe recently, that contained a lot of function calls. Most of them do a bunch of PUSHes, read some stack vars, followed by the function's actual purpose, and end with the POPs, and finally the return. And, of course that's in the majority of function calls: this constant pattern of register save, fill, use, reload, followed by return (another stack pop). Seems like there should be a better way. Maybe instead of a set for every program/thread, you could have just enough space for a few levels deep of registers. This would correspond with the "call stack". Since most calls come with PUSHes, and most returns are proceeded by POPs, it seems like a prime candidate for optimization, for both speed and code size.

Maybe that's still too complex or slow for hardware, though.

One option would be register windows like on SPARC, which acts as a hardware assisted stack. Here's an article about that:
http://ieng9.ucsd.edu/~cs30x/sparcstack.html

Another option would be load-dual and store-dual instructions, which let you combine two 64-bit register memory operations into a single 128-bit store (as long as your stack is 128 bit-aligned) and is probably as fast possible for saving/restoring a bunch of registers at the same time (especially if you can dual-issue it to a 2-port data cache).