…tomaticAllocationScope

This doesnt affect their behavior when they are just
called from the command line, since these switches
will target the ops nested within the module (which
before where func.func, now can be others).

Fix affine tests

Fix anchor for some passes

TODO better fix would be to improve the PassManager
to support passes that can run on an arbitrary nesting
level. Then these passes could continue to target the
root op properly. The fix with Nesting::ImplicitAny is
really bad because it changes the anchor of OperationPass<>
passes (any anchor) to target children of the root instead
of the root.

Fix pass manager nesting again

And canonicalize a single-element memref copy
into a load and a store. This allows SROA to
scalarize the copied elements. Currently
SROA cannot handle forwarding of one memory slot
to another.

causing relinking of many libs when
a change to Pass.cpp is made

It had a bug whereby the regions of an
op are marked as dead code, and their
successors are not properly populated,
if any operand of a RegionBranchOpInterface
is not folded to a constant. The interface
already is meant to support partially-folded
operands, so we should use that instead of giving up.

Some changes are also just to make debug prints
more helpful. Some of them were printing pointer
addresses.

The problem is that the live in
set is not necessarily already
populated when we ask for the
start op

in the case where the subview op uses
an index of an affine loop as eg, the
offset. The previous implementation
always generated symbols, and the
verifier failed, although the
transformation is valid if you just
generate dims for the variables that
are not symbols, but are valid dims.

and not just plain dims and symbols. This makes it possible
to fold memref.subview ops that use an affine expression of
valid symbol and dims as an offset, even if that expression
is computed by arith ops like muli and addi.