[RFA/Ada 1/2] Handle other cases than EVAL_NORMAL in the default case

[Jerome Guitton <guitton at adacore dot com>]

In the evaluation of an expression in Ada mode, the default case
unwraps the argument unconditionally. For an object of a variant
record type, this unwrapping builds a fixed type from the
specification of the variant type and the actual values of the
object's discriminants.  It means that unwrapping needs the "proper"
value for the object, not just a zero value with the proper type.

When not in EVAL_NORMAL, we cannot assume that the evaluation returns
such a proper value; it may well return a zero value of the
appropriate type e.g in EVAL_AVOID_SIDE_EFFECTS. It is wrong to try to
unwrap in that case.

In particular, a problem shows up when using expression of the form
{VARIANT_TYPE}OBJ. GDB first evaluates this expression in
EVAL_AVOID_SIDE_EFFECTS to compute the type, the evaluation of OBJ
in most cases returns a zero value of its type, and as UNOP_MEMVAL
is mapped to the default case its evaluation ends up trying to
read memory around address 0.

gdb/ChangeLog:

	* ada-lang.c (ada_evaluate_subexp): Unwrap only in EVAL_NORMAL.

OK to apply?
---
 gdb/ada-lang.c |    4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/gdb/ada-lang.c b/gdb/ada-lang.c
index 0621c79..ee2f765 100644
--- a/gdb/ada-lang.c
+++ b/gdb/ada-lang.c
@@ -9453,7 +9453,9 @@ ada_evaluate_subexp (struct type *expect_type, struct expression *exp,
     default:
       *pos -= 1;
       arg1 = evaluate_subexp_standard (expect_type, exp, pos, noside);
-      arg1 = unwrap_value (arg1);
+
+      if (noside == EVAL_NORMAL)
+	arg1 = unwrap_value (arg1);
 
       /* If evaluating an OP_DOUBLE and an EXPECT_TYPE was provided,
          then we need to perform the conversion manually, because
-- 
1.7.10.4