r7-2111 introduced maybe_constant_value in cp_fully_fold.
maybe_constant_value uses cxx_eval_outermost_constant_expr, which
can clear TREE_CONSTANT:
6510   else if (non_constant_p && TREE_CONSTANT (r))
[...]
6529       TREE_CONSTANT (r) = false;

In this test the array size is '(long int) "h"'.  This used to be
TREE_CONSTANT but given the change above, the flag will be cleared
when we cp_fully_fold the array size in compute_array_index_type_loc.
That means we don't emit an error in the
10391   else if (TREE_CONSTANT (size)
block in the same function, and we go on.  Then we compute ITYPE
using cp_build_binary_op and use maybe_constant_value on it and
suddenly we have something that's TREE_CONSTANT again.  And then we
crash in reshape_init_array_1 in tree_to_uhwi, because what we have
doesn't fit in an unsigned HWI.

icc accepts this code, but since we used to reject it, I see no desire
to make this work, so don't use the folded result when we've lost
the TREE_CONSTANT flag while evaluating the size.

2020-02-26  Marek Polacek  <polacek@redhat.com>

	PR c++/93789 - ICE with invalid array bounds.
	* decl.c (compute_array_index_type_loc): Don't use the folded
	size when folding cleared TREE_CONSTANT.

	* g++.dg/ext/vla22.C: New test.