RFR: JDK-8199781: Don't use naked == for comparing oops (original) (raw)

Roman Kennke rkennke at redhat.com
Wed Mar 28 14:03:27 UTC 2018

Hi Erik,

Thanks for this work!

without having looked too deeply at the code (it might be in vain anyway, what with all the template voodoo stuff going on :-P), I have a few conceptual questions:

Thanks, Roman

Hi,

The access.hpp support has reincarnated. Here is an incremental patch on Roman's work (rebased on latest jdk-hs) that makes it possible to include access.hpp to call Access<>::equals specifically. The infrastructure allows picking more accesses that can be used from the hpp file, but I have kept it down to the minimum requirements now, so for now it only works for equals. Perhaps in the future, we might want to move more operations to be reachable from hpp, but not today. I changed the involved header files to include access.hpp, and made it possible to inline this again for non-Shenandoah builds, by checking for the INTERNALBTTOSPACEINVARIANT build-time decorator. Builds that only have to-space invariant GCs inline equals to a normal native comparison of the pointers. Builds that have at least one to-space invariant GC instead generate a single function pointer call. This function pointer may be resolved in the access.cpp file, instead of the access.inline.hpp file. Here is an explanation how this works: The template pipeline for each operation consists of 5 steps, and they were previously all performed in the access.inline.hpp file. The steps are: 1) Set default decorators and decay types 2) Reduce types 3) Pre-runtime dispatch 4) Runtime-dispatch 5.a) Barrier resolution (check which GC backend accessor to be used upon the first invocation, then patch function pointer to 5.b) 5.b) Post-runtime dispatch I have moved steps 1-4 to the accessBackend.hpp file, which is included by access.hpp. This allows performing all steps until the backend barrier resolution in only hpp files. This makes it possible to use, e.g. raw accesses (although without compressed oops, atomic or order access) from only the access.hpp file. The resolution of barriers in the GC backends (comprising steps 4-5) remains in the access.inline.hpp file. But they can be instantiated from the access.cpp file for well chosen accesses and decorator combinations that can be pre-generated in the libjvm.so file. The function pointers used in builds that have not to-space invariant GCs will be resolved to point right into these functions in the access.cpp file at runtime. This change required moving the decorators to a new file called accessDecorators.hpp. I kind of like having them there in a separate file. Incremental webrev to Roman's patch: http://cr.openjdk.java.net/~eosterlund/8199781/webrev.0001/ Full webrev: http://cr.openjdk.java.net/~eosterlund/8199781/webrev.01/ Thanks, /Erik On 2018-03-27 09:14, Roman Kennke wrote: Am 27.03.2018 um 00:46 schrieb John Rose:

On Mar 26, 2018, at 2:21 PM, Roman Kennke <rkennke at redhat.com> wrote:

Am 26.03.2018 um 21:11 schrieb John Rose:

... This is clearly one of those cases.  Let's code against surprises (both present and future) by abstracting simpler operations with inlines. Thanks John for clarifying. I generally do agree with you that it makes sense to inline something like an operator that will compile to a single instruction.

Unfortunately, this new abstraction will generate a non-inlineable call. Thanks Erik's work with the Access API, it should be a straight call to the right implementation, not even a virtual call, but it's a call nonetheless. What would it take to inline the call, so there's no new performance hazard introduced? This might be a place to add #ifdefs, if the templates don't let us control the decision statically. (I confess I find the templates complex and hard to reason about. But that's a potential maintenance issue with the access API, not your work.) The Access API supports static inclusion/exclusion of stuff, and we've used it before to ensable barriers on primitive field accesses for Shenandoah-enabled builds. It works by passing -DSUPPORTBARRIERONPRIMITIVES to gcc. This is much better than #ifdefs and does the same thing. I can add something similar for object equality. The reason why I'm proposing it is that the GC might want to have a say about object equality. In the case of Shenandoah, there may be two copies of any one object floating around, and comparing the two copies might otherwise result in false negatives. 1. Of course we need a way for some GC algorithms to control object identity. 2. It's good coding discipline to find all the places where native op== is used and upgrade them to the needed control. I hope I did that with the proposed patch (it's based on years of finding all those places in Shenandoah land ;-) ). (2a. We also need a way to avoid having op== creep back in in an uncontrolled way.) We have something for this in Shenandoah too. It works by overriding == in oopsHierarch.hpp and running into ShouldNotReachHere() if both operands are != NULL. This catches all uses of naked == at runtime, when running with +CheckUnhandledOops. It's not perfect, because it doesn't catch uses in places that are not executed in tests, but it's still incredibly useful. The caveat here is that we also need another abstraction for code paths that actually want naked ==, e.g. in the GC. In Shenandoah we have an oopDesc::unsafeequals() for that. The other alternative would be to drop operators from oop to oopDesc* and compare that (which is exactly what oopDesc::unsafeequals() does). 3. We still need an unsurprising performance model for GC's which don't require the extra identity hook. Yeah. Unfortunately this is C++ and not Java, and the decision which GC to use is made at runtime. The 'best' that I can see is to rip out the indirection in Shenandoah-disabled builds, as proposed above. I don't like it much, but if this is the only way we can get Shenandoah in, then let's go for it? So we're not there yet, I think.  If there's no easy way to adjust the access API to get the last inlining, I suggest adding an #ifdef into oopDesc::equals. I suspect the happy final solution will be to templatize hot loops using making the inlinable adjusted op== be part of the "decorations" of those loops.  Then we will have two copies of hot loops in the JVM, one with the out-of-line hook and one without. For now, I think it's enough to have separate builds of the JVM, one with the #ifdef set "the old way" and one that allows more flexibility. (Is this the first time we've run into an occasion to make a separate Shenandaoah build?  Surely not.) Primitive field accesses required similar jumping through hoops ;-) So... with inlining oopDesc::equals() we will get one straight call (from the caller of oopDesc::equals() to the impl), instead of 2 (caller -> oopDesc::equals() -> impl). It's the best I could come up with that does what (Shenandoah) GC needs, and it hasn't shown up in any performance testing that we did, yet. Still good? I think it's too risky to out-of-line the "cmpq" instruction.  Let's find an agreeable way to handle that via a trade-off that keeps the old code shape in VM builds which don't need the new code shape.  I'm speaking up here because I don't believe this is the last time we'll have to consider adding something gross like an #ifdef to support new features, and I want to aim at trade-offs that keep the existing hot paths efficient, while still allowing new code shapes to become part of the mix. (I'd be happy to see something cleaner than an #ifdef, of course.  Seems to me that a constant non-product flag -XX:+ShortCircuitOopEquals would work too, with a binding of falseInShenandoah.) That would require coding 2 versions of hot loops that involve oop==oop and switch to one or the other based on that flag. I have checked all the uses of oopDesc::equals() in my patch, and none seemed particularily critical. I think this would be overkill. And notice also that none interpreter/c1/c2 compiled code will suffer this problem because we'll just generate the right code at runtime. Roman

-------------- next part -------------- A non-text attachment was scrubbed... Name: signature.asc Type: application/pgp-signature Size: 836 bytes Desc: OpenPGP digital signature URL: <https://mail.openjdk.org/pipermail/hotspot-gc-dev/attachments/20180328/2a0f0001/signature.asc>

More information about the hotspot-gc-dev mailing list