RFR (S): 8201326: Renaming ThreadLocalAllocationBuffer end to fast_path_end (original) (raw)

Daniel D. Daugherty daniel.daugherty at oracle.com
Wed Apr 18 13:29:08 UTC 2018

Greetings,

The idea of splitting this change off from "8171119: Low-Overhead Heap Profiling" came up during the design review. It might have been me that suggested the split off or it was someone else. Sorry I don't remember.

In any case, the rename of "end" to "fast_path_end" is just a clarity change to the existing code and I think that change can easily stand on its own. That is particularly true if the accessors are renamed at the same time. I think having the accessor names match the field names is a pretty well known HotSpot rule so I'm not sure why we're talking about not renaming the accessors...

Personally, I prefer "_fast_path_end" over "_current_end".

Dan

On 4/18/18 4:04 AM, Stefan Johansson wrote:

Hi,

On 2018-04-18 02:02, Vladimir Kozlov wrote:  > I think I like better not to add it. If no one is using it, there should be  > no reason to add it? By not adding it, it also makes any future wish to  > have it a nice indicator of: hey why do you want to see this? Same as  > hardend btw which is not visible. Am I missing something?

I say "may" ;) You don't need new function if there is no use.  >  > So to summarize, the current consensus:  >    - end can be renamed either to currentend or fastpathend; that is  > still up to a vote and choice :) Please, use currentend. I was thinking about sampleend but it does not reflect double usage. I'm not sure if you have seen the discussion about naming on hotspot-gc-dev, but I and others think that currentend doesn't describe the usage at all. Naming it fastpathend would clearly state what it is, sampleend or something similar also crossed my mind but I think the code reads a lot better in the compiler with the name fastpathend.

 >    - the access method end() and tlabendoffset() remain the same to not  > modify JIT/interpreter codes I would find this very unfortunate, having accessors that don't match the members can easily lead to misunderstanding, especially if we have three different *end members. Why do you think this is the best way forward? My first thought was also that it would be nice to avoid all the compiler changes, but then we would have to stick with end being the sample/current/fast-path end and I'm not sure that is a better solution. I don't see the big problem with changing those accessors to what they really access and I think the compiler code reads good even after the change. For example: cmpptr(end, Address(thread, JavaThread::tlabfastpathendoffset())); jcc(Assembler::above, slowcase);  >  > If all agree to this, then the consequences are:  >    - JDK-8201326 becomes useless  >    - The work for JEP-331 becomes smaller in terms of file changes  >    - I'll still be needing a decision on the renaming of the TLAB end field  > (current suggestions are currentend and fastpathend). We'll see where we end up. If JDK-8201326 ends up being a separate change I think it should be pushed at the same time as the rest of the JEP changes. To me it only makes sense to have it in the code base if we also have the rest of the changes. Thanks, Stefan Sounds good to me. Thanks, Vladimir On 4/17/18 4:51 PM, JC Beyler wrote: Hi Vladimir and Dean,

@Dean: seems that Vladimir is in agreement with you for renaming just the field and not the method names so ack to your answer that came at the same time :)

Yes, from the beginning such changes should be discussed on common hotspot-dev list since all Hotspot's parts are affected. Sorry, being new to the scene, most of the conversation had been going on in serviceability-dev.

Graal specific question could be send to hotspot-compiler-dev list with [Graal] in subject. I looked on JEP's changes http://cr.openjdk.java.net/~jcbeyler/8171119/webrev.02/ to understand how it works. Few questions about proposed JEP changes so I can understand it. You introducing new TLAB end for sampling and adjust it so that allocations in JITed code and Interpreter it will trigger slow path allocation where you do sampling. Right? Yes that is correct; if sampling is enabled of course. Btw, this is the current webrev <http://cr.openjdk.java.net/~jcbeyler/8171119/heapevent.15/>. Do all changes to end, actualend and other TLAB fields happen during slow allocation? Yes, to that effect, with Robbin's help, we finalized deprecating the FastTLabRefill via JDK-8194084 <https://bugs.openjdk.java.net/browse/JDK-8194084>. Seems like I/we missed that Graal does this as well. I filed GRAAL-64 <https://bugs.openjdk.java.net/browse/GRAAL-64> to not forget that Graal would need to get that fixed.

I am concern about concurrent accesses to these fields from other threads if you have them (I don't see). Yes that is why we deprecated the FastTlabRefill. Other threads should not be changing the thread local data structure so I don't see an issue there. The major issue was that the fast paths could change the tlab without going via the slowpath. I had a fix to detect this issue but removed it once JDK-8194084 was done; Graal was missed in that work so that is why if sampling were enabled with Graal, there is a chance things would break currently. That will get fixed via GRAAL-64 <https://bugs.openjdk.java.net/browse/GRAAL-64> whether it is rendering the code also obsolete and going to the slowpath or whether it is adding my fix again to detect a fastpath TLAB reallocation happened. Renaming. I am fine with renaming if it helps to understand code better. I agree with proposed changes to fields name: currentend allocationend But, as Dean, I would suggest to keep names of access functions. This way we can say that allocation code in Interpreter and JITed code stay the same. Sounds good to me, then in that case, this webrev will disappear, which also makes me happy, since it simplifies a lot of things (and reduces code change). You may need only new method to access allocationend in places which look for real end of TLAB. I think I like better not to add it. If no one is using it, there should be no reason to add it? By not adding it, it also makes any future wish to have it a nice indicator of: hey why do you want to see this? Same as hardend btw which is not visible. Am I missing something? So to summarize, the current consensus: - end can be renamed either to currentend or fastpathend; that is still up to a vote and choice :) - the access method end() and tlabendoffset() remain the same to not modify JIT/interpreter codes If all agree to this, then the consequences are: - JDK-8201326 becomes useless - The work for JEP-331 becomes smaller in terms of file changes - I'll still be needing a decision on the renaming of the TLAB end field (current suggestions are currentend and fastpathend). Thanks for your help! Jc Regards, Vladimir On 4/16/18 4:42 PM, JC Beyler wrote: Hi Dean, I think perhaps this is also an attempt to figure out the naming of all this because naming (or renaming like here) is often the start of big conversations :). Originally, in the JEP-331 work, I had left the end field as is and, by doing so, this whole renaming webrev goes away. However, if we do that, then the TLAB code contains: end: the fast path end, can be the allocation end or the to-be-sampled end allocationend: the actual allocation end of the current TLAB hardend: allocationend + reserved space With an early iteration of a webrev for JEP-331, a conversation occurred about whether or not that was clear or not and it was determined that this renaming was more clear: currentend: the fast path end allocationend: the actual allocation end of the current TLAB reservedend: allocationend + reserved space Because I'm trying to reduce the footprint of files changed, I pulled out the renaming changes into this webrev. While talking about it with the GC team, the renaming suggested became: fastpathend: the fast path end allocationend: the actual allocation end of the current TLAB hardend: allocationend + reserved space Now, to be honest, any renaming or no renaming is fine by me. I like not renaming the fields to not change the code of every backend and Graal; I also like the fastpathend rename due to it making the backend code easy to read as mentioned in the GC mailing lis thread. So there are two questions really: - Should we rename the end field and thus provoke the changes in this webrev? - If we do want to change the field, should/could it go in an initial webrev or should it go in the JEP-331 webrev whenever/ifever it goes in? And if we do wait, could we at least converge on a renaming now so that this does not become a point of contention for that webrev's review? If I read your answer correctly: - You are saying that we should keep the end field altogether; or at least only rename the field not the methods to access it, thus reducing the change scope. - You are also saying to wait for the JEP-331 webrev's final iteration and integrate it in one step Have I understood your answer correctly? Again, I am fine with renaming to whatever or not renaming at all. I just am trying to get forward progress here :) Thanks for your help! Jc On Mon, Apr 16, 2018 at 3:29 PM <dean.long at oracle.com> wrote: Hi JC.  Others might disagree, but I would vote "no" on doing this renaming now, before the JEP, and even in the context of the JEP, I don't think renaming the field requires renaming all the uses.  The compiler code is only interested in the fast path, so it's implicitly understood.  Also, if there is a fastpathend(), then isn't it logical to have fastpathstart() paired with it?  ("start" is already overloaded, but nobody is suggesting adding allocationstart()/currentstart()/hardstart() are they?).  My opinion is that it's fine the way it is. dl On 4/16/18 1:43 PM, JC Beyler wrote: Hi all, I've left the mail thread from the hotspot-gc-dev below for tracking purposes but will start a new request here. - I'm trying to push a renaming of a TLAB field to make my work for JEP-331 <http://openjdk.java.net/jeps/331> easier - There is an understanding that if JEP-331 does not make it, this might be useless and I'll revert if that is what the community wants; however the new name seems better anyway so perhaps not? - The webrev renames a field used across the various back-ends and Graal - The webrev is here: http://cr.openjdk.java.net/~jcbeyler/8201326/webrev.04/ Could I please get some feedback on this? Thanks all for your help, Jc

On Fri, Apr 13, 2018 at 2:37 AM Stefan Johansson <_ _stefan.johansson at oracle.com> wrote: Hi JC, I don't have a name, but I've looked at bit more at the failures and I think they are unrelated and one of the local compiler engineers agree. I also ran some local testing and was not able to get any error with you latest changes, but verified that it doens't work without the graal parts. So they seem good. It might still be good to switch this over to the general hotspot-dev list to let someone with Graal knowledge to look at the change and make sure everything is correct. Thanks, Stefan On 2018-04-12 17:26, JC Beyler wrote: Hi Stefan, Thanks for testing :). I've renamed the bug title in the JBS and will emit a new webrev shortly. Do you have the name of a compiler engineer in mind or should I perhaps now move this conversation to the general hotspot-dev mailing list and ask there? The alternative is to add the compiler-mailing list to this email thread and ask there before sending to the general list. What do you think is best? Thanks for all your help, Jc On Thu, Apr 12, 2018 at 2:09 AM Stefan Johansson <stefan.johansson at oracle.com_ _<mailto:stefan.johansson at oracle.com>> wrote: On 2018-04-11 17:48, JC Beyler wrote: > Hi Stefan, > > Sorry about that, I must have missed adding the files or something. Here > is a webrev that added the changes for the SA file: > http://cr.openjdk.java.net/~jcbeyler/8201326/webrev.03/ > No problem, this looks good. I kicked of a run in our test system to get some more coverage and have tried to include some Graal testing. I'll let you know the results once they are done. Please also update the bug title both in JBS and the changeset. Cheers, Stefan > I changed the method name, which propagated a change to: > src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/oops/ObjectHeap.java > > I tried testing your test file. It exists in my branch (if the same) under: > hotspot/jtreg/serviceability/sa/ClhsdbJhisto.java > > and interestingly (which generally means I did something wrong), it > passed before/after the change so I could not verify that this is a test > showing that all is well in the world on my side. Any ideas of what I > did wrong? > > I did again test it for hotspot/jtreg/compiler/aot/ and > hotspot/jtreg/serviceability/jvmti and it passes those. > > Thanks for all your help, > Jc > > > > On Wed, Apr 11, 2018 at 7:44 AM Stefan Johansson > <stefan.johansson at oracle.com <mailto:_ _stefan.johansson at oracle.com> <mailto:stefan.johansson at oracle.com_ _<mailto:stefan.johansson at oracle.com>>> wrote: > >     Hi JC, > >     On 2018-04-11 00:56, JC Beyler wrote: >      > Small update: >      > >      > Here is the fixed webrev for the '{' that were out of alignment. >     This >      > passed release build JTREG for hotspot/jtreg/compiler/jvmti (just >     to run >      > something as a smoke screen) and hotspot/jtreg/compiler/aot/ (to >     test >      > Graal). >      > http://cr.openjdk.java.net/~jcbeyler/8201326/webrev.02/ >      > >     I think this looks better, I agree that leaving end is tempting to >     avoid a lot of change, but I think this will be better in the long run. > >     I still miss the changes to make the SA work. To see a problem you >     can run: >     make CONF=fast run-test > TEST=open/test/hotspot/jtreg/serviceability/ClhsdbJhisto.java > >     Cheers, >     Stefan > >      > Let me know what you think, >      > Jc >      > >      > On Tue, Apr 10, 2018 at 3:21 PM JC Beyler <jcbeyler at google.com <mailto:jcbeyler at google.com> >     <mailto:jcbeyler at google.com_ _<mailto:jcbeyler at google.com_ _>      > <mailto:jcbeyler at google.com <mailto:_ _jcbeyler at google.com_ _<mailto:jcbeyler at google.com <mailto:jcbeyler at google.com>>>> wrote: >      > >      >     Hi Karen and Stefan, >      > >      >     @Stefan: Naming is hard :) >      >     @Karen: thanks for the Graal comment, I fixed it in the new >     webrev, >      >     let me know what you think :) >      > >      >     I think the naming convention suggested in this webrev came from >      >     conversations in for JEP-331 and I am actually relatively >      > indifferent to the final result (as long as we have some form of >      >     forward progress :)). To be honest, I'd also be happy to just >     leave >      >     end as is for all architectures and Graal and have a new >      > allocationend. However, during initial reviews of JEP-331 >     it was >      >     deemed complicated to understand: >      > >      >     end -> the end or sampling end >      >     allocationend -> end pointer for the last possible allocation >      >     hardend -> allocation end + reserved space >      > >      >     That is how this naming came up and why hardend went to > "reservedend". >      > >      >     I'm really indifferent, so I offer as a perusal: >      > http://cr.openjdk.java.net/~jcbeyler/8201326/webrev.01/ >      > >      >     I noticed a few problems of alignement of '{' so I'll go fix >     that. >      >     Basically this webrev does the following: >      > >      >     - Uses fastpathend instead of end >      >     - Reverts hardend back to where it was >      >     - Adds the changes to Graal; there is a bunch of changes in Graal >      >     because Graal still contains a bit of code doing fasttlabrefills. >      > >      >     Let me know what you think! >      > >      >     Thanks, >      >     Jc >      > >      > >      >     On Tue, Apr 10, 2018 at 6:56 AM Karen Kinnear >      > <karen.kinnear at oracle.com_ _<mailto:karen.kinnear at oracle.com> <mailto:_ _karen.kinnear at oracle.com_ _<mailto:karen.kinnear at oracle.com>> > <mailto:karen.kinnear at oracle.com_ _<mailto:karen.kinnear at oracle.com> <mailto:_ _karen.kinnear at oracle.com_ _<mailto:karen.kinnear at oracle.com>>>> >     wrote: >      > >      >         Hi JC, >      > >      >         A comment about Graal. The impact on Graal for this >     particular >      >         change would be to break it - so you’ll need >      >         to complete the Graal changes for this renaming. That isn’t >      >         optional or something that could be a follow-on. It > >         is not ok to break a feature, even an experimental one. >     We will >      >         discuss in the other thread potential phasing of adding >     sampling. >      > >      >         I did not do a thorough search -you can do that - I did find >      > src/jdk.internal.vm.compiler/share/classes/ >      > > org.graalvm.compiler.hotspot/src/org/graalvm/compiler/hotspot/GraalHotSpotVMConfig.java: > >            public final int threadTlabOffset = >      > getFieldOffset("Thread::tlab", Integer.class, > >         "ThreadLocalAllocBuffer"); >      > > org.graalvm.compiler.hotspot/src/org/graalvm/compiler/hotspot/GraalHotSpotVMConfig.java: > >            private final int threadLocalAllocBufferStartOffset = >      > getFieldOffset("ThreadLocalAllocBuffer::start", > Integer.class, >      >         "HeapWord*"); >      > > org.graalvm.compiler.hotspot/src/org/graalvm/compiler/hotspot/GraalHotSpotVMConfig.java: > >            private final int threadLocalAllocBufferEndOffset = >      > getFieldOffset("ThreadLocalAllocBuffer::end", Integer.class, >      >         "HeapWord*"); >      > > org.graalvm.compiler.hotspot/src/org/graalvm/compiler/hotspot/GraalHotSpotVMConfig.java: > >            private final int threadLocalAllocBufferTopOffset = >      > getFieldOffset("ThreadLocalAllocBuffer::top", Integer.class, >      >         "HeapWord*"); >      > > org.graalvm.compiler.hotspot/src/org/graalvm/compiler/hotspot/GraalHotSpotVMConfig.java: > >            private final int threadLocalAllocBufferPfTopOffset = >      > getFieldOffset("ThreadLocalAllocBuffer::pftop", > Integer.class, >      >         "HeapWord*"); >      > > org.graalvm.compiler.hotspot/src/org/graalvm/compiler/hotspot/GraalHotSpotVMConfig.java: > >            private final int > threadLocalAllocBufferSlowAllocationsOffset >      >         = getFieldOffset("ThreadLocalAllocBuffer::slowallocations", >      >         Integer.class, "unsigned"); >      > > org.graalvm.compiler.hotspot/src/org/graalvm/compiler/hotspot/GraalHotSpotVMConfig.java: > >            private final int > threadLocalAllocBufferFastRefillWasteOffset >      >         = > getFieldOffset("ThreadLocalAllocBuffer::fastrefillwaste", > >         Integer.class, "unsigned"); >      > > org.graalvm.compiler.hotspot/src/org/graalvm/compiler/hotspot/GraalHotSpotVMConfig.java: > >            private final int > threadLocalAllocBufferNumberOfRefillsOffset >      >         = > getFieldOffset("ThreadLocalAllocBuffer::numberofrefills", > >         Integer.class, "unsigned"); >      > > org.graalvm.compiler.hotspot/src/org/graalvm/compiler/hotspot/GraalHotSpotVMConfig.java: > >            private final int >      > threadLocalAllocBufferRefillWasteLimitOffset = >      > getFieldOffset("ThreadLocalAllocBuffer::refillwastelimit", >      >         Integer.class, "sizet"); >      > > org.graalvm.compiler.hotspot/src/org/graalvm/compiler/hotspot/GraalHotSpotVMConfig.java: > >            private final int > threadLocalAllocBufferDesiredSizeOffset = >      > getFieldOffset("ThreadLocalAllocBuffer::desiredsize", >      >         Integer.class, "sizet"); >      > > org.graalvm.compiler.hotspot/src/org/graalvm/compiler/hotspot/GraalHotSpotVMConfig.java: > >            public final int tlabAlignmentReserve = >      > > getFieldValue("CompilerToVM::Data::ThreadLocalAllocBufferalignmentreserve", > >         Integer.class, "sizet”); >      > >      >         hope this helps, >      >         Karen >      > >      >>         On Apr 10, 2018, at 7:04 AM, Stefan Johansson > >>         <stefan.johansson at oracle.com_ _<mailto:stefan.johansson at oracle.com> > <mailto:stefan.johansson at oracle.com_ _<mailto:stefan.johansson at oracle.com>> >      >> <mailto:stefan.johansson at oracle.com_ _<mailto:stefan.johansson at oracle.com> > <mailto:stefan.johansson at oracle.com_ _<mailto:stefan.johansson at oracle.com>>>> wrote: >      >> >      >>         Hi JC, >      >> >      >>         I realize that the names have been discussed before but I'm >      >>         leaning towards suggesting something new. We discussed this >      >>         briefly here in the office and others might have different >      >>         opinions. One point that came up is that if we do this >     change >      >>         and change all usages of JavaThread::tlabendoffset() it >      >>         would be good to make sure the new name is good. To us >      >>         currentend is not very descriptive, but naming is hard and >      >>         the best we could come up with is fastpathend which would >      >>         give the code: >      >>          cmpptr(end, Address(thread, >      >> JavaThread::tlabfastpathendoffset())); >      >> jcc(Assembler::above, slowcase); >      >> >      >>         I think this reads pretty good and is fairly clear. If we do >      >>         this rename I think you can re-use end in JEP-331 >     instead of >      >>         calling it allocationend. But that is a later review :) >      >> >      >>         Also, is there a good reason for renaming hardend() to >      >>         reservedend()? >      >> >      >>         One additional thing, you need to update the SA to reflect >      >>         this change. I think the only place you need to fix is in: >      >> > src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/runtime/ThreadLocalAllocBuffer.java >      >> >      >>         Thanks, >      >>         Stefan >      >> >      >>         On 2018-04-09 19:24, JC Beyler wrote: >      >>>         Hi all, >      >>>         Small pre-amble to this request: >      >>>         In my work to try to get a heap sampler in OpenJDK (via JEP >      >>>         331 > <https://bugs.openjdk.java.net/browse/JDK-8171119>), I'm > >>>         trying to reduce the footprint of my change so that the >      >>>         integration can be easier. I was told that generally a JEP >      >>>         webrev should be feature complete and go in at-once. >     However, >      >>>         with the change touching quite a bit of various code >     pieces, >      >>>         I was trying to figure out what could be separated as not >      >>>         "part of the feature". >      >>>         I asked around and said that perhaps a solution would be to >      >>>         cut up the renaming of TLAB's end field that I do in that >      >>>         webrev. Because I'm renaming a field in TLAB used by >     various >      >>>         backends for that work, I have to update every architecture >      >>>         dependent code to reflect it. >      >>>         I entirely understand that perhaps this is not in the >     habits >      >>>         and very potentially might not be the way things are > >>>         generally done. If so, I apologize and let me know if you >      >>>         would not want this to go in separately :) >      >>>         Final note: there is still a chance JEP-331 does not go in. >      >>>         If it does not, we can leave the new name in place or I'll >      >>>         happily revert it. I can even create an issue to track this >      >>>         if that makes it easier for all. >      >>>         End of the pre-amble. >      >>>         The 33-line change webrev in question is here: > >>> http://cr.openjdk.java.net/~jcbeyler/8201326/webrev.00/ > >>>         I fixed all the architectures and JVMCI and ran a few >     sanity >      >>>         tests to ensure I had not missed anything. >      >>>         Thanks for your help and I hope this is not too much >     trouble, >      >>>         Jc >      >>>         Ps: there is a graal change that needs to happen but I was >      >>>         not sure who/where <https://teams.googleplex.com/u/where> <https://teams.googleplex.com/u/where> <https://teams.googleplex.com/u/where> <https://teams.googleplex.com/u/where> > <https://teams.googleplex.com/u/where> > <https://teams.googleplex.com/u/where> to >      >>>         ask about it. I was told it could happen in a separate >      >>>         webrev. Can anyone point me to the right direction? >     Should it >      >>>         just be hotspot-compiler-dev? >      > >

More information about the hotspot-dev mailing list