Extended stack code genration bugs

There are some bugs in code generation for extended stack operands int eh z80-related ports. While testcases tend to be hard to create, as they typically require functions with hundreds of local variables, running the normal regression tests compiling using --fomit-frame-pointer combined with a really low value for --max-allocs-per-node will detect them too, by forcing most variables onto the stack and stack access to be done without frame pointers, which often results in the same mechanism as the one for the extended stack to be used. I already fixed some such bugs, but there's more. Here's the current (revison #7854) results of the regression tests using --fomit-frame-pointer --max-allocs-per-node 100:

Summary for 'ucz80': 114 failures, 6908 tests, 1540 test cases, 4285735 bytes, 13713292 ticks
Failure: gen/ucz80/bug1714204/bug1714204
Failure: gen/ucz80/bug1723128/bug1723128
Failure: gen/ucz80/bug3521024/bug3521024
Failure: gen/ucz80/gcc-torture-execute-20010910-1/gcc-torture-execute-20010910-1
Failure: gen/ucz80/gcc-torture-execute-20080424-1/gcc-torture-execute-20080424-1
Failure: gen/ucz80/gcc-torture-execute-950929-1/gcc-torture-execute-950929-1
Failure: gen/ucz80/libmullong/libmullong_type_asm
Failure: gen/ucz80/libmullong/libmullong_type_c

Philipp