In the testcase, fold_indirect_ref_1 won't fold *(T*)(s1+10) to an 
ARRAY_REF because T != unsigned.  Even if it were just a typedef to 
unsigned, that isn't close enough, but in this case it's a typedef to 
const unsigned.

I'm not sure what the type coherence rules are for ARRAY_REF.  Is it 
really necessary that the type of the ARRAY_REF match exactly the 
element type of the array?

In any case, constexpr expansion can be more flexible about type 
coherence because it is just trying to get a constant value; if that 
doesn't work out, we throw it away and fall back on the original 
expression.  We already handle some cases in cxx_eval_indirect_ref that 
aren't appropriate for fold_indirect_ref_1, but this testcase 
demonstrates that we also want to adjust the cases that are handled by 
that function.

So my options would seem to be either duplicating the whole of 
fold_indirect_ref_1, which tempts undesirable divergence, or adding a 
flag to that function to enable the more permissive type checking that 
the constexpr code wants.

Does this seem like a reasonable thing to do?