Better handling of register clobbers in functions that adjust the stack pointer

Vague title, but I don't really know what the root issue is.

Repro: http://cyberjes.us/bash-ppc.elf @ 0x100c0a00

current_function.clobbered_regs lists just about every register. The one that's particularly bothering me is r2, which is the global pointer register defined in the calling convention. This is breaking a bunch of references to constant data (r2 + offset).

For context, 100c0a00 is part of the fairly standard "restore GPRs" sled compilers generate for PPC. Rather than doing your typical push for each callee-saved register used in a function, PPC compilers tend to insert a sled of stores and a sled of loads, then functions will just branch to the right offset in the sled, e.g. save_r20_to_r31(); actual function code; restore_r20_to_r31();

Start with the PPC repro at http://cyberjes.us/bash-ppc.elf at 0x100c0a00, and inspect how current_function.clobbered_regs is populated for the restore-GPR sled. Done means the analysis no longer incorrectly marks registers such as r2 as clobbered in this pattern, while preserving correct clobber reporting.