This avoids code duplication and actually works correctly.
This fixes opt/convs.c.

Computing (post)dominators removes tuples, which in turn invalidates out edges, if any tuples were removed.

It currently uses dominance checks, though the use looks questionable.
backend/lower_call4.c is the only test case which triggers this check.
The check was added in bd1b1f0e.
Though later in 5f4cf6e0 another check was added, which would also handle this case for different reasons (target is not on the frame).
So at least the test case needs to be updated validate the dominance check.
Maybe the dominance check should be replaced by a dependence check.

This should fix opt/jumpthreading4.c and opt/place.c when debug (in particular verification) is off.

The dominator calculation assures the same.
This fixes the optimize build for several test cases.

Before it always said 8, because it took the callee perspective, which adds the artificial parameters.
This caused calls to variadic functions to get 8 passed in %eax, which is unnecessary for most calls.

We now also consider post dominance to find SESE regions

Also fix some missing or wrong uses of (post)dominance property methods.

'ia32_create_address_mode()' (actually 'x86_create_address_mode()') sets the variant already.

This ensures that blocks with a label get their label printed.
In particular this fixes computed goto on arm.

This fixes C/gnu99/combo_load_mode.c, C/gnu99/computed_goto.c, C/gnu99/ijmp.c, C/gnu99/vlabel2.c and C/gnu99/vlabel3.c.

This fixes C/gnu99/combo_load_mode.c, C/gnu99/computed_goto.c, C/gnu99/ijmp.c, C/gnu99/vlabel2.c and C/gnu99/vlabel3.c.

The verifier checks this already.

When compound types are passed as pointer, then not only the mode, but also the type, needs to be adjusted to a pointer.

They are passed as a pointer, not the value itself, so an extra load is needed.

This was broken when rewriting stack frame handling. (b6787e36)
Though the below test case did not work before either, because the correct code path was not entered due to inspecting the wrong type.
The latter was accidentally corrected while getting rid of higher types. (ae42d217)
This fixes gcc/va-arg-12.c.

Unused attributes may be uninitialized.

The negation accidentally got lost when using be_mode_needs_gp_reg().

When the Add was marked as non-adress-mode node it tried to recursively transform the Add again.
Just use 'x86_create_am_force' as it is done in the ia32 backend in the same situation.

The verifier would have complained earlier, if an Address node does not have reference mode.