RFR: JDK-8198243: Add build time check for global operator new/delete in object files

[Erik Joelsson]

Hello,

Here is a new version of this patch, reworked in several ways. It now 
supports gcc, clang and solstudio. It uses nm instead of objdump, which 
is more readily available in all our current build environments. The 
check now uses mangled symbol names for all toolchain types which makes 
things considerably faster. I also added a check for only finding 
symbols classified as undefined. Otherwise we get false positives in 
operator_new.o in debug builds.

I left the objdump addition to the devkit in there because it's a good 
thing to have anyway for compare.sh.

Webrev: http://cr.openjdk.java.net/~erikj/8198243/webrev.02/

I have run this patch against current jdk/jdk and jdk/hs in Mach5. In 
jdk/jdk the build fails as expected on Solaris, Linux and Mac and in 
jdk/hs, where the fixes we are detecting are present, all builds succeeds.

/Erik


On 2018-02-23 05:27, Magnus Ihse Bursie wrote:
> On 2018-02-22 20:41, Erik Joelsson wrote:
>>
>>
>> On 2018-02-21 21:06, David Holmes wrote:
>>> On 22/02/2018 4:07 AM, Erik Joelsson wrote:
>>>> Hello,
>>>>
>>>>
>>>> On 2018-02-20 21:33, David Holmes wrote:
>>>>>
>>>>> a) how much time it adds to the build?
>>>>>
>>>> I have not done extensive testing, but on my Linux workstation with 
>>>> 32 hw threads, building just hotspot release build from a clean 
>>>> workspace increased maybe 1 or 2 seconds (at around 90s total), but 
>>>> the variance was around the same amount as that.
>>>>> b) why this doesn't work for Solaris Studio?
>>>>>
>>>> I didn't put a lot of effort into trying to figure it out. The 
>>>> check used was provided by Kim Barrett, for Linux only. I figured 
>>>> it would be simple enough to get it to work on mac and succeeded 
>>>> there. It should certainly be possible to implement a similar check 
>>>> on Solaris, but is it worth the time to do it? Both development 
>>>> time and increased build time on one of the slower build platforms?
>>>
>>> Depends how concerned we are with detecting this problem in OS 
>>> specific source code?
>>>
>> I investigated this some more. I was able to do it successfully, but 
>> the build time cost is way too large here. The culprit is c++filt on 
>> Solaris which is incredibly costly, and the gnu version does not 
>> demangle Solaris Studio symbols. I don't think we should do this on 
>> Solaris.
> I agree, it's not worth it.
>
> Not all programmer's mistakes are reasonable to catch in technical 
> traps. It we *should* spend time on getting automatic tool for keeping 
> code quality up (and, yes, I really do think we should), there's most 
> likely to be much better areas to spend that effort in, in making a 
> lot of prone-to-break scripts for catching a single kind of problem.
>
> /Magnus
>
>>
>> We could grep for the mangled strings for the operators instead, 
>> which is super fast. Problem is just figuring out all the possible 
>> combinations.
>>
>> /Erik
>>> To be honest I'm not sure this pulls its own weight regardless.
>>>
>>> David
>>>
>>>> /Erik
>>>>> Thanks,
>>>>> David
>>>>>
>>>>> On 21/02/2018 4:05 AM, Erik Joelsson wrote:
>>>>>> Hello,
>>>>>>
>>>>>> This patch adds a build time check for uses of global operators 
>>>>>> new and delete in hotspot C++ code. The check is only run with 
>>>>>> toolchains GCC and Clang (Linux and Macos builds). I have also 
>>>>>> modified the Oracle devkit on Linux to add the necessary symlink 
>>>>>> so that objdump will get picked up by configure.
>>>>>>
>>>>>> This change is depending on several fixes removing such uses that 
>>>>>> are currently in jdk/hs so this change will need to be pushed 
>>>>>> there as well.
>>>>>>
>>>>>> Bug: https://bugs.openjdk.java.net/browse/JDK-8198243
>>>>>>
>>>>>> Webrev: http://cr.openjdk.java.net/~erikj/8198243/webrev.01/
>>>>>>
>>>>>> /Erik
>>>>>>
>>>>
>>
>