[Bug c/109618] [12/13/14 Regression] ICE: tree check: expected class ‘type’, have ‘exceptional’ (error_mark) in generic_simplify_CONVERT_EXPR, at generic-match.cc since r12-3278-g823685221de986

[pinskia at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109618

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
This one seems harder to fix.

We have:
```
(gdb) p debug_tree(value)
 <nop_expr 0x7ffff79ad5c0
    type <integer_type 0x7ffff7822000 sizetype public unsigned DI
        size <integer_cst 0x7ffff7802f48 constant 64>
        unit-size <integer_cst 0x7ffff7802f60 constant 8>
        align:64 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type
0x7ffff7822000 precision:64 min <integer_cst 0x7ffff7802f78 0> max <integer_cst
0x7ffff78035e0 18446744073709551615>>
    readonly
    arg:0 <var_decl 0x7ffff7810c60 var_1 type <error_mark 0x7ffff7802f30>
        readonly used unsigned read SI t2.c:3:22
        size <integer_cst 0x7ffff7824198 constant 32>
        unit-size <integer_cst 0x7ffff78241b0 constant 4>
        align:32 warn_if_not_align:0 context <function_decl 0x7ffff79b3600 foo>
initial <integer_cst 0x7ffff79bbbe8 2>>>
```

Before calling fold.
So we turn var_1's type to error_mark when merging the 2 decls (after an
error).
But then the rest of the front-end/middle-end is still not ready for types all
the time to have an error_mark.

Note I don't think we should revert r12-3278 even though it has had a lot of
fall out because it produces better error recovery in general.

Note for this simplified testcase we could look through the NOP_EXPR in
c_sizeof_alignof but a more complex one which does s/var_5[var_1]/var_5[13 *
var_1]/, we can't as we get:
(sizetype) ((unsigned int) var_1 * 13)

And that one ICEs in tree_nonzero_bits.
    CASE_CONVERT:
      return wide_int::from (tree_nonzero_bits (TREE_OPERAND (t, 0)),
                             TYPE_PRECISION (TREE_TYPE (t)),
                             TYPE_SIGN (TREE_TYPE (TREE_OPERAND (t, 0))));