Ping: Switch initializations conversion

+ info.index_expr = SWITCH_CND (swtch);
+ index_type = TREE_TYPE (info.index_expr);

The use of the SWITCH_CND's type as the index type is problematic in Ada (and
maybe in other languages) because it badly interacts with the optimization in
tree-ssa-forwprop.c that gobbles casts within SWITCH_CND and you may end up
doing arithmetics in the wrong type.

Ada testcase attached, compile with "gcc -c p.adb".

+ BUG
| 4.4.0 20080704 (experimental) [trunk revision 137450] (i586-suse-linux-gnu)
GCC error:|
| in set_value_range, at tree-vrp.c:379 |
| Error detected around p.adb:12

In .028t.ccp:

op_kind.0_7 = (q__internal_element_kinds) op_kind_2(D);
switch (op_kind.0_7)
{
case 244: goto <L4>;
case 246: goto <L3>;
default : goto <L5>;
}

op_kind is , which is an
subtype (you cannot do arithmetics in it), so it is converted to the base
type q__internal_element_kinds before the switch.

In 029t.forwprop1 the conversion is removed:

op_kind.0_7 = (q__internal_element_kinds) op_kind_2(D);
switch (op_kind_2(D))
{
case 244: goto <L4>;
case 246: goto <L3>;
default : goto <L5>;
}

Then in 040t.switchconv arithmetics in the subtype is created:

csui.6_8 = op_kind_2(D);
csui.6_7 = csui.6_8 - 244;
if (csui.6_7 <= 2)
goto <bb 10(<L10>);
else
goto <bb 9(<L9>);

which doesn't make sense since the range of the subtype in [226; 246] and
ultimately leads to the VRP failure because it deduces the [226;2] range
for the first branch of the condition.


It seems to me that, since the pass generates new arithmetics for the
SWITCH_EXPR, it should do it in the SWITCH_EXPR's type and the condition be
converted back to it if necessary. What do you think? Thanks in advance.

--
Eric Botcazou