6480539: BigDecimal.stripTrailingZeros() should specify no-op on zero BigDecimals (original) (raw)

Joe Darcy joe.darcy at oracle.com
Fri Feb 15 05:54:18 UTC 2013

On 02/08/2013 02:37 PM, Stephen Colebourne wrote:

We've established that it causes problems (me and Paul) and that likely workarounds of the bug would not be broken by a fix.

The adaptation you and Paul have for the current behavior would not be impacted by a change in the behavior of stringTrailingZeros. That just leaves all the other Java programmers to check with ;-)

Since any angry emails on changing the behavior would come the way of Brian and I (probably in a few years time), Brian has agreed to do some additional research on how stripTrailingZeros is used in existing code to gauge what impact there could be to altering the long-standing behavior. If the impact is small, we can make the change in JDK 8.

-Joe

The getMethods case is very different as the spec clearly allowed the behavior. Here the spec and method name are also clear on the expected behavior. The proposal would make the behavior out of line with the method name and of no earthly use to users. ie, as a user of that method, I would never be calling it for any reason other than to remove all the trailing zeroes. Mostly the JDK team gets it right. But not this time. Stephen

On 8 Feb 2013 20:32, "Joe Darcy" <joe.darcy at oracle.com> wrote: Hello,

On 2/6/2013 11:32 PM, Bruce Chapman wrote: Stephen, In your case(s) would the workaround fail to work if the bug was fixed? Working around a bug is quite different to taking advantage of the buggy behaviour. If fixing the bug would break code that works around it that can be seen as a problem, while breaking code that relies on the bug is IMHO much less of an issue since anyone that does that is taking a known risk, or a risk that should reasonably be expected to be known. I am finding it hard to imagine a genuine attempt at a workaround that would not still work (though redundantly) if the bug was fixed. Also bear in mind that there are other implementations, and the signature and the javadoc are the spec. If you find behaviour that differs and take advantage of that behaviour then you are opening the possibility of it changing if either: you run with another implementation, or the bug gets fixed. While it is easy to contrive an example that would break if this bug were fixed, and it is possible (on the grounds that I cannot prove it is impossible) that some real code might break, it is hard to imagine a scenario where the author/owner of that broken code has any morally legitimate grounds for complaint in that case. I guess if you take the "This is one of those unfortunate cases where a bug can become a feature." approach to its logical conclusion then no bugs get fixed because there are no bugs, just a nice online list of newly discovered unexpected features. As noted earlier in this thread, we use a nuanced compatibility policy in evolving the JDK: http://cr.openjdk.java.net/~darcy/OpenJdkDevGuide/OpenJdkDevelopersGuide.v0.777.html In particular, besides looking after source and binary compatibility, we also look to manage behavioral compatibility, that is, to be mindful of changing what a method does at runtime when called, even when the specification gives us leeway to do so. Let me relate an example of behavioral compatibility from JDK 7. The method Class.getMethods returns an array of Method objects for the Class and in part its javadoc has long stated: "The elements in the array returned are not sorted and are not in any particular order." Therefore, any caller of Class.getMethods relying on or assuming a particular order has a bug according to the specification. As a side-effect of permgen removal in JDK 7, the long-standing (and mostly stable) order of Method objects returned by HotSpot changed. As expected, some user applications and tests "broke" after this change went in. We received requests to "fix" the ordering of Class.getMethods, which we declined to do given the benefits of permgen removal and the clear specification that no ordering should be relied upon. Even though that change in getMethods is allowed by specification, it is out-of-bounds of what we would do an an update release but in-bounds for a platform release like JDK 7. The reason for this conservatism is because we value keeping the broad usage of the JDK working :-) Getting back to BigDecimal.stripTrailingZeros, we cannot inspect all usages of this method today, nor can we inspect all the future usages of BigDecimal.stripTrailingZeros that will be around before JDK 8 is adopted for the code in question. We know not everyone migrates to a new JDK release promptly; within the past two years I fielded a query/complaint about the behavior change in BigDecimal.toString made between 1.4.2 and JDK 5 and later. For these sorts of reasons, the default resolution when the specification and implementation conflict is to make the specification match the implementation. There are exceptions to this default. Given sufficient evidence that changing the behavior of BigDecimal.stripTrailingZeros would not have adverse consequences on fielded code, we could change its behavior despite being implemented that way for about 9 years. -Joe

Bruce

On 7/02/2013 12:16 p.m., Stephen Colebourne wrote: On 5 February 2013 09:09, Paul Sandoz <paul.sandoz at oracle.com> wrote: This is one of those unfortunate cases where a bug can become a feature. I really don't see how.

The method name is absolutely clear about its purpose. "Strip trailing zeros". Anyone relying on it not stripping zeroes for zero needs their head examining. This particular one just happens to be one that I've run across twice and in both cases it required a workaround. I'd argue that there are more people with undiscovered bugs in their code because the method is buggy than people who would break were the method fixed. What bothers me even more is the desire expressed in this thread to simply wish away bugs by redefining the documentation. If the method name is clear enough, like in this case, then its a bug, and a documentation change simply isn't the right solution. Stephen

More information about the core-libs-dev mailing list