<AWT Dev> [12] RFR : JDK-8198000 : java/awt/List/EmptyListEventTest/EmptyListEventTest.java debug assert on Windows
Ambarish Rapte
ambarish.rapte at oracle.com
Wed Nov 7 17:19:19 UTC 2018
Hi Phil,
> I don't think so. I already pointed out that it is apparently returning -1 in your test
> when HIWORD is 1. I am pretty sure it isn't expecting you to interpret 0xFFFF stored
> in 16 bits as positive.
The value 0xFFFF is -1 when saved into a signed short and 65535 when assigned to unsigned short.
I was trying to find any supporting documentation or source code from MSDN. But cannot be certain of the behavior based on documentation.
So I also think that we can go ahead with the earlier version of fix as you suggested. : http://cr.openjdk.java.net/~arapte/8198000/webrev.04/
Please take a look at the fix, The webrev has fix only for JDK-8198000.
The fix change for issue #2 i.e. cmath::pow function is removed from webrev.04
> These are platform UI components. They can behave differently w.r.t things like
> how they respond to mouse clicks, all dependent on what the platform norms are.
> I see no reason to go out of our way to "make windows mouse clicks behave like mac".
This looks good to me as well. With the changes in webrev.04, an item gets selected only when mouse is clicked on the item.
> Not understanding you. How can you tell if 0x0001 is supposed to be "1" or "65536" ?
Sorry for not explaining the fix in earlier email with webrev.03.
I have removed this change from the webrev.
I should not have mixed fix of the two issues. If the issue #2 needs to be addressed, I request if we can work on this issue separately.
> So if the user clicks outside we do nothing. What is wrong with that ?
I too see nothing wrong in this behavior. Mouse click outside the client area can be neglected under this fix.
Please take a look at the updated webrev.04 which includes fix only for JDK-8198000
Regards,
Ambarish
From: Philip Race
Sent: Tuesday, November 06, 2018 10:21 PM
To: Ambarish Rapte <ambarish.rapte at oracle.com>
Cc: awt-dev at openjdk.java.net
Subject: Re: <AWT Dev> [12] RFR : JDK-8198000 : java/awt/List/EmptyListEventTest/EmptyListEventTest.java debug assert on Windows
On 11/5/18 8:51 AM, Ambarish Rapte wrote:
Hi Phil,
The little abstraction or lack of clarity in documentation is leading us to make some assumptions.
Just trying to connect the pieces,
I referred these two doc pages to learn LOWORD behavior:
LOWORD : HYPERLINK "https://msdn.microsoft.com/en-us/library/windows/desktop/ms632659%28v=vs.85%29.aspx"https://msdn.microsoft.com/en-us/library/windows/desktop/ms632659(v=vs.85).aspx
WORD : https://docs.microsoft.com/en-us/windows/desktop/winprog/windows-data-types
And the documentation of LB_ITEMFROMPOINT does not mention LB_ERR(-1)
Based on these documentation, it looks safe to conclude that LOWORD would be positive.
I don't think so. I already pointed out that it is apparently returning -1 in your test
when HIWORD is 1. I am pretty sure it isn't expecting you to interpret 0xFFFF stored
in 16 bits as positive.
Issue #1 :
Yes, the HIWORD is 1, when the mouse is clicked inside the list's client area but not on the item.
Once we separate the HIWORD and LOWORD, LOWORD gives correct index of the last item, and this issue gets solved.
Checking if HIWORD is 1 can be additional check. But as we extract correct index, this check can be ignored for allowing selection.
Or the other way, we can use HIWORD to avoid selection/deselection when it is 1.
Additionally, The behavior on Ubuntu and mac is different,
Ubuntu: Last item does not get selected & it also does not select any other item. User must click on the item to select or deselect.
Mac: Selects the last item
Windows: Focused item gets selected / deselected on multiple clicks, 3-4
After this fix[Allowing selection], Mac and Windows will have same behavior.
These are platform UI components. They can behave differently w.r.t things like
how they respond to mouse clicks, all dependent on what the platform norms are.
I see no reason to go out of our way to "make windows mouse clicks behave like mac".
Issue #2 :
This looks a very corner case of list of size more than 65535 size, but the item selection and deselection works fine with keyboard.
Also List behaves correctly on Ubuntu and Mac.
Just for a look, I am including a fix for this issue in webrev.03: HYPERLINK "http://cr.openjdk.java.net/%7Earapte/8198000/webrev.03/"http://cr.openjdk.java.net/~arapte/8198000/webrev.03/
As you have already mentioned we should not really change a long running behavior, but I could not hold on from suggestion.
If you think this fix is not needed, I shall just change/remove it.
Not understanding you. How can you tell if 0x0001 is supposed to be "1" or "65536" ?
I also don't like std::pow .. it is unnecessary here.
Not withstanding the lack of clarity on how you are claiming to know that,
I was expecting a fix to look more like
LONG count = GetCount();
if (count > 0) {
LONG item = static_cast<LONG>(SendListMessage(LB_ITEMFROMPOINT, 0, msg->lParam));
if (HIWORD(item) == 0) {
item = LOWORD(item);
if (item > 0 && item < GetCount()) {
...
So if the user clicks outside we do nothing. What is wrong with that ?
-phil.
Regards,
Ambarish
From: Philip Race
Sent: Saturday, November 03, 2018 10:59 PM
To: Ambarish Rapte HYPERLINK "mailto:ambarish.rapte at oracle.com"<ambarish.rapte at oracle.com>
Cc: HYPERLINK "mailto:awt-dev at openjdk.java.net"awt-dev at openjdk.java.net
Subject: Re: <AWT Dev> [12] RFR : JDK-8198000 : java/awt/List/EmptyListEventTest/EmptyListEventTest.java debug assert on Windows
- I am not sure you can assume the LOWORD value is non-negative.
It seems to me that the 0XFFFF we got back is meant to be interpreted as "-1"
which is what I wrote below.
I do note that it appears that in issue #2 you are seeing that up to 65534 items
may be allowed and it wraps if you add more than 65535 ..
So *perhaps* we can interpret 0xFFFF as meaning positive 65535 but I think
*only* if HIWORD is "0".
- Issue #1 : I am not sure that a "1" in HIWORD automatically means it is off the
end of the list ... just "outside the client area". I think if HIWORD is 1 we just
bail don't we ?
Then you don't have to worry about whether 0xFFFF meant -1 or 65536
- Issue #2 : There is always *some* limit in cases like this. 32767 (2^15-1) or 65535 (2^16-1)
are very typical in these platform APIs. Often the platform doesn't explicitly document it
and you have to infer it from the data type. I think it was all very moot when these APIs
were designed because you'd run out of memory before you could get that many items :-)
I'd be surprised if there were not already some open bug pointing out that we accept "int"
for index and don't have any feedback when exceeding platform limits.
>From a compatibility point of view, I don't think it is worth doing anything that would
break ancient applications to provide that feedback.
-phil.
On 11/2/18, 10:23 AM, Ambarish Rapte wrote:
Hi Phil,
Thanks for guiding with the documentation.
The fix is modified after a relook at the documentation, and observed two issues [mentioned below].
webrev: HYPERLINK "http://cr.openjdk.java.net/%7Earapte/8198000/webrev.02"http://cr.openjdk.java.net/~arapte/8198000/webrev.02
Fix Justification:
LB_ITEMFROMPOINT documentation(https://docs.microsoft.com/en-us/windows/desktop/controls/lb-itemfrompoint):
Message LB_ITEMFROMPOINT returns a LONG value.
There is no mention of LB_ERR as return value, like it is mentioned clearly for some of the other messages as LB_GETCOUNT, LB_SETITEMHEIGHT, LB_GETTEXT, LB_GETCURSEL.
So the existing comparison against LB_ERR is incorrect.
The two parts LOWORD and HIWORD of return value are of type WORD(unsigned short), so the return value can never be negative.
It is another reason to not to compare the returned index value with LB_ERR which is defined as (-1).
Fix:
Extract the index from LOWORD into a WORD variable and verify only if the index is smaller than the list size. (webrev.02)
This fixes both JDK-8198000 & below mentioned Issue 1.
2 other Issues:
Issue 1: NON selection of an item.
Currently the returned LONG value is used without extracting the LOWORD.
As far as the HIWORD is 0, the LONG return value would be same as index of item(LOWORD).
But when HIWORD is 1, the LONG return value would be a large unexpected value. If it is used as an index, then it would result in NON selection of the item.
Test program: HYPERLINK "http://cr.openjdk.java.net/%7Earapte/8198000/ListSize3.java"http://cr.openjdk.java.net/~arapte/8198000/ListSize3.java
Steps:
Compile and execute the test program with release build JDK.
Click in the list's client area below the last item, i.e. do not click on any item.
Expected behavior: Last item should get selected.
Actual behavior: Last item does not get selected on first click. But the focused item gets selected after few clicks.
Issue 2: Incorrect selection when list size exceeds sizeof(WORD) [0xFFFF].
Steps:
Compile and execute the program with release build JDK (with or without the fix) : HYPERLINK "http://cr.openjdk.java.net/%7Earapte/8198000/List65544.java"http://cr.openjdk.java.net/~arapte/8198000/List65544.java
Click the first item in list, press End key.
Click the last item 65544.
Expected behavior: The item 65544 should get selected.
Actual behavior: The item 65544 does not get selected and instead item 8 gets selected.
Verified only windows behavior.
I suggest to file a new JBS for this issue.
Regards,
Ambarish
From: Phil Race
Sent: Thursday, November 01, 2018 1:51 AM
To: Ambarish Rapte HYPERLINK "mailto:ambarish.rapte at oracle.com"<ambarish.rapte at oracle.com>; HYPERLINK "mailto:awt-dev at openjdk.java.net"awt-dev at openjdk.java.net
Subject: Re: <AWT Dev> [12] RFR : JDK-8198000 : java/awt/List/EmptyListEventTest/EmptyListEventTest.java debug assert on Windows
That adds what I suggested, but I had also suggested you leave in what you
had added as it also adds some protection.
Additionally I read the MS docs and they do explain the 131071 return value.
The message this code is sending is LB_ITEMFROMPOINT and the docs say here
https://docs.microsoft.com/en-us/windows/desktop/controls/lb-itemfrompoint that
>The return value contains the index of the nearest item in the HYPERLINK "https://msdn.microsoft.com/library/windows/desktop/ms632659"LOWORD. The HYPERLINK "https://msdn.microsoft.com/library/windows/desktop/ms632657"HIWORD is zero if the > specified point is in the client area of the list box, or one if it is outside the client area"
You got 131071 which is, in hex 0X1FFFF.
So you got "1" for hi-word, meaning "outside client area" and "-1" for loword,
meaning the index. And a return index of "-1" doubtless means an error ..
-phil.
On 10/31/18 12:49 PM, Ambarish Rapte wrote:
Hi Phil & Sergey,
This issue was introduced with the fix for HYPERLINK "https://bugs.openjdk.java.net/browse/JDK-6806217"JDK-6806217, in 7b55, which modified AwtList::HandleEvent(), so It was not observed with JDK6.
Please review the updated change as discussed offline: HYPERLINK "http://cr.openjdk.java.net/%7Earapte/8198000/webrev.01/"http://cr.openjdk.java.net/~arapte/8198000/webrev.01/
Regards,
Ambarish
From: Phil Race
Sent: Tuesday, October 30, 2018 2:09 AM
To: Ambarish Rapte HYPERLINK "mailto:ambarish.rapte at oracle.com"<ambarish.rapte at oracle.com>; HYPERLINK "mailto:awt-dev at openjdk.java.net"awt-dev at openjdk.java.net
Subject: Re: <AWT Dev> [12] RFR : JDK-8198000 : java/awt/List/EmptyListEventTest/EmptyListEventTest.java debug assert on Windows
On 10/29/18 7:31 AM, Ambarish Rapte wrote:
Hi,
Please review this Windows only fix,
JBS: https://bugs.openjdk.java.net/browse/JDK-8198000
Webrev: HYPERLINK "http://cr.openjdk.java.net/%7Earapte/8198000/webrev.00/"http://cr.openjdk.java.net/~arapte/8198000/webrev.00/
Issue:
Test asserts with debug build jdk, only on windows.
Only the debug build turns on asserts.
But I think JDK 6 always turned on asserts, and this test was introduced in 6,
so something must have changed else we'd have seen this test fail a long time ago.
Can you identify what it was ?
Also same comment as the other bug - you need to add the bug id to the test.
Assert at Line no 77, awt_List.h :: IsItemSelected()
awt_List.cpp :: AwtList::HandleEvent() calls IsItemSelected() with an incorrect index value.
Why ?
In AwtList::HandleEvent() , the call SendListMessage(LB_ITEMFROMPOINT, 0, msg->lParam) returns an arbitrary value 131071, which gets passed to IsItemSelected().
Could not find any relevance to the value 131071, from LB_ITEMFROMPOINT doc.
That is (128*1024)-1, so it is probably not "arbitrary".
Please add the eval above to the bug report .. once we have a complete understanding.
-phil.
Fix: Index should be verified before making call to IsItemSelected() : HYPERLINK "http://cr.openjdk.java.net/%7Earapte/8198000/webrev.00/"http://cr.openjdk.java.net/~arapte/8198000/webrev.00/
Verification: All list tests pass.
Regards,
Ambarish
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/awt-dev/attachments/20181107/eda5e525/attachment-0001.html>
More information about the awt-dev
mailing list