Attached is another revision of a patch I posted last July to keep
the pretty-printer from crashing on MEM_REFs with void* arguments:
   https://gcc.gnu.org/pipermail/gcc-patches/2020-July/549746.html

Besides avoiding the ICE and enhancing the MEM_REF detail and
improving its format, this revision implements the suggestions
in that discussion.  To avoid code duplication it moves
the handling to the C pretty-printer and changes the C++ front
end to delegate to it.  In addition, it includes a cast to
the accessed type if it's different from/incompatible with
(according to GIMPLE) that of the dereferenced pointer, or if
the object is typeless.  Lastly, it replaces the <unknown> in
the output with either VLA names or the RHS of the GIMPLE
expression (this improves the output when for dynamically
allocated objects).

As an aside, In my experience, MEM_REFs in warnings are limited
to -Wuninitialized.  I think other middle end warnings tend to
avoid them.  Those that involve invalid/out-of-bounds accesses
replace them with either the target DECL (e.g., local variable,
or FIELD_DECL), the allocation call (e.g., malloc), or the DECL
of the pointer (e.g., PARM_DECL), followed by a note mentioning
the offset into the object.  I'd like to change -Wuninitialized
at some point to follow the same style.  So I see the value of
the MEM_REF formatting enhancement mainly as a transient solution
until that happens.

Martin