RFR (S): 8024067: Missing replace_in_map() calls following null checks

[Roland Westrelin]

Thanks for taking the time to discuss this.

Is this better:
http://cr.openjdk.java.net/~roland/8024067/webrev.01/
?

Roland.

On Sep 4, 2013, at 12:15 AM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:

> Can you pass safe_for_replace into null_check_oop()? I want replace_in_map in one place.
> 
> Vladimir
> 
> On 9/3/13 2:49 PM, Roland Westrelin wrote:
>>> Then why not do that in null_check_oop()? And why we need safe_for_replace?
>> 
>> For instance, LibraryCallKit::inline_native_Class_query() for vmIntrinsics::_isInstance builds a region with 2 control paths. gen_instanceof() is used on one of the paths. gen_instanceof() calls null_check_oop(). If null_check_oop() does the replace_in_map then we can end up with a reference in the map after the isInstance that is only valid on one of the control paths. In general, null_check_oop() is called directly or indirectly from library_call.cpp where we build our own control flow and null_check_oop() can be only valid on one control path and so a replace_in_map cannot be safely performed.
>> 
>> Roland.
>> 
>>> PS: Roland, please, include previous mails text (with webrev link) in your responses.
>>> 
>>> On 9/3/13 1:48 PM, Roland Westrelin wrote:
>>>> 
>>>>> Why you need this when null_check_common() does this already?
>>>> 
>>>> GraphKit::null_check_oop() sets (*null_control) = top()
>>>> then GraphKit::null_check_common() does:
>>>> if (null_control != NULL) {
>>>> (*null_control) = null_true;
>>>> so (*null_control) is no longer top and:
>>>> 
>>>>    if (null_control == NULL || (*null_control) == top())
>>>>       replace_in_map(value, cast);
>>>> 
>>>> doesn't help. Back in GraphKit::null_check_oop():
>>>> 
>>>>   if (never_see_null && (*null_control) != top()) {
>>>> 
>>>> is where the uncommon trap is added and (*null_control) = top() is set back to top and nothing does the replace_in_map.
>>>> 
>>>> Roland.
>>>> 
>>