This bug seems to have been exposed by the combined m->m load/store
instructions available on M68k (these instructions are not available on
i386, which the M68k backend is based on). This meant that token factors
were inserted which could lead to distinct call sequence chains,
increasing the nesting level and preventing the matching callseq_start
from being identified during scheduling.
The patch addresses this by not allowing combined loads/stores when the
folded operation would result in a new chain dependency on a different
call sequence.
closes#146213 and #175472
Had been doing this piece by piece, but makes more sense to do it in a
single PR. Adds support for `ARID`, `PCI`, `PCD`, `AL`, and `ARD`
addressing modes for atomic operations, along with a variety of tests.
The `CodeModel` tests have been rearranged, as some of the new
addressing modes are only exercised under some combinations of
`CodeModel` and relocation mode
