RFR (L): 8003868: fix shark for latest HotSpot and LLVM [Was: Re: RFR: Fix shark for latest Hotspot and LLVM]
Roman Kennke
rkennke at redhat.com
Thu Nov 29 06:26:19 PST 2012
Am Dienstag, den 27.11.2012, 10:59 -0800 schrieb Christian Thalinger:
> On Nov 27, 2012, at 9:24 AM, Roman Kennke <rkennke at redhat.com> wrote:
>
> > Am Montag, den 26.11.2012, 17:14 -0800 schrieb Christian Thalinger:
> >> On Nov 26, 2012, at 4:44 PM, Roman Kennke <rkennke at redhat.com> wrote:
> >>
> >>> Am Montag, den 26.11.2012, 15:43 -0800 schrieb Christian Thalinger:
> >>>> On Nov 26, 2012, at 3:18 PM, Roman Kennke <rkennke at redhat.com> wrote:
> >>>>
> >>>>> Am Montag, den 26.11.2012, 11:55 -0800 schrieb Christian Thalinger:
> >>>>>> On Nov 23, 2012, at 9:14 AM, Roman Kennke <rkennke at redhat.com> wrote:
> >>>>>>
> >>>>>>> Am Mittwoch, den 21.11.2012, 14:52 -0800 schrieb Christian Thalinger:
> >>>>>>>> On Nov 21, 2012, at 2:22 PM, Christian Thalinger <christian.thalinger at oracle.com> wrote:
> >>>>>>>>
> >>>>>>>>>
> >>>>>>>>> On Nov 21, 2012, at 2:17 PM, Christian Thalinger <christian.thalinger at oracle.com> wrote:
> >>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> On Nov 21, 2012, at 12:54 PM, Roman Kennke <rkennke at redhat.com> wrote:
> >>>>>>>>>>
> >>>>>>>>>>> Am Mittwoch, den 21.11.2012, 12:47 -0800 schrieb Christian Thalinger:
> >>>>>>>>>>>> On Nov 21, 2012, at 11:43 AM, Christian Thalinger <christian.thalinger at oracle.com> wrote:
> >>>>>>>>>>>>
> >>>>>>>>>>>>> On Nov 21, 2012, at 9:31 AM, Roman Kennke <rkennke at redhat.com> wrote:
> >>>>>>>>>>>>>
> >>>>>>>>>>>>>> Hi there,
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> during the last days I worked on fixing the Shark compiler for Hotspot
> >>>>>>>>>>>>>> to get it to build and run again, with the latest Hotspot code and LLVM.
> >>>>>>>>>>>>>> Here are some details:
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> - A lot of changes are just to make it build and the compiler happy. For
> >>>>>>>>>>>>>> example, I had to remove a lot of 'const' qualifiers because of API
> >>>>>>>>>>>>>> changes in LLVM.
> >>>>>>>>>>>>>> - Most other changes have to do with the split of the oop and metadata
> >>>>>>>>>>>>>> class hierarchies in Hotspot.
> >>>>>>>>>>>>>> - Then there have been a few changes caused by LLVM changes and
> >>>>>>>>>>>>>> improvements, most notably the LLVM intrinsics for atomic operations
> >>>>>>>>>>>>>> (memory barrier and cmpxchg) have been removed and now have a
> >>>>>>>>>>>>>> representation directly in LLVM's IR. This makes our code a little
> >>>>>>>>>>>>>> nicer.
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> I tested this by running a number of applications, most notably Eclipse
> >>>>>>>>>>>>>> (which is notoriously difficult on VMs), Java2Demo, SwingSet2 and a
> >>>>>>>>>>>>>> bunch of other stuff.
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> I would like to get this integrated into OpenJDK now if possible. You
> >>>>>>>>>>>>>> can find the full webrev here:
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> http://cr.openjdk.java.net/~rkennke/shark/webrev.00/
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> The changes seem to touch almost only shark files so these should be fine. One question though:
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> + develop(bool, SharkShowCompiledMethods, false, \
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> Isn't PrintCompilation doing that already?
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> The shared code changes look good. I filed:
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> 8003868: fix shark for latest HotSpot and LLVM
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> -- Chris
> >>>>>>>>>>>>>
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> There are also a very minor change required in JDK:
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> http://cr.openjdk.java.net/~rkennke/shark/webrev-jdk-00/
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> In order to build it, apply the patches on hsx/hotspot-comp 's hotspot
> >>>>>>>>>>>>>> and jdk repositories respectivly. Find my build script here:
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> http://cr.openjdk.java.net/~rkennke/shark/Build8-zero-shark
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> (Review and adjust variables to your settings, most notably you will
> >>>>>>>>>>>>>> need to change LLVM_CONFIG to point to your LLVM 3.1 installation.)
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> Please let me know if there are any issues or how we can get this
> >>>>>>>>>>>>>> integrated into Hotspot.
> >>>>>>>>>>>>
> >>>>>>>>>>>> Finally I installed LLVM on one of our machines to be able to do a Shark build once in a while. When I try to do a jvmgshark build I get:
> >>>>>>>>>>>>
> >>>>>>>>>>>> In file included from /usr/local/include/llvm/Support/PointerLikeTypeTraits.h:18,
> >>>>>>>>>>>> from /usr/local/include/llvm/ADT/PointerIntPair.h:17,
> >>>>>>>>>>>> from /usr/local/include/llvm/Use.h:28,
> >>>>>>>>>>>> from /usr/local/include/llvm/Value.h:17,
> >>>>>>>>>>>> from /usr/local/include/llvm/Argument.h:17,
> >>>>>>>>>>>> from /home/cthaling/8003868/src/share/vm/shark/llvmHeaders.hpp:39,
> >>>>>>>>>>>> from /home/cthaling/8003868/src/share/vm/shark/sharkEntry.hpp:29,
> >>>>>>>>>>>> from /home/cthaling/8003868/src/share/vm/compiler/disassembler.cpp:51:
> >>>>>>>>>>>> /usr/local/include/llvm/Support/DataTypes.h:53:3: error: #error "Must #define __STDC_CONSTANT_MACROS before " "#including Support/DataTypes.h"
> >>>>>>>>>>>>
> >>>>>>>>>>>> and:
> >>>>>>>>>>>>
> >>>>>>>>>>>> In file included from /usr/local/include/llvm/Attributes.h:18,
> >>>>>>>>>>>> from /usr/local/include/llvm/Argument.h:18,
> >>>>>>>>>>>> from /home/cthaling/8003868/src/share/vm/shark/llvmHeaders.hpp:39,
> >>>>>>>>>>>> from /home/cthaling/8003868/src/share/vm/shark/sharkEntry.hpp:29,
> >>>>>>>>>>>> from /home/cthaling/8003868/src/share/vm/compiler/disassembler.cpp:51:
> >>>>>>>>>>>> /usr/local/include/llvm/Support/MathExtras.h: In function ‘bool llvm::isInt(int64_t)’:
> >>>>>>>>>>>> /usr/local/include/llvm/Support/MathExtras.h:38: error: there are no arguments to ‘INT64_C’ that depend on a template parameter, so a declaration of ‘INT64_C’ must be available
> >>>>>>>>>>>> /usr/local/include/llvm/Support/MathExtras.h:38: note: (if you use ‘-fpermissive’, G++ will accept your code, but allowing the use of an undeclared name is deprecated)
> >>>>>>>>>>>> /usr/local/include/llvm/Support/MathExtras.h:38: error: there are no arguments to ‘INT64_C’ that depend on a template parameter, so a declaration of ‘INT64_C’ must be available
> >>>>>>>>>>>> /usr/local/include/llvm/Support/MathExtras.h: In function ‘bool llvm::isUInt(uint64_t)’:
> >>>>>>>>>>>> /usr/local/include/llvm/Support/MathExtras.h:64: error: there are no arguments to ‘UINT64_C’ that depend on a template parameter, so a declaration of ‘UINT64_C’ must be available
> >>>>>>>>>>>> /usr/local/include/llvm/Support/MathExtras.h: In function ‘bool llvm::isIntN(unsigned int, int64_t)’:
> >>>>>>>>>>>> /usr/local/include/llvm/Support/MathExtras.h:96: error: ‘INT64_C’ was not declared in this scope
> >>>>>>>>>>>>
> >>>>>>>>>>>> Not sure if the latter is because of the former one. Have you seen this before?
> >>>>>>>>>>>
> >>>>>>>>>>> Yes, it's caused by the former. And yes, I have seen it before. IIRC,
> >>>>>>>>>>> this happens when certain cflags are not set correctly. The script
> >>>>>>>>>>> jdk/make/jdk_generic_profile.sh will call llvm-config to figure out the
> >>>>>>>>>>> correct flags. In order for this to work, you need to have the full path
> >>>>>>>>>>> to the llvm-config script set in the LLVM_CONFIG env variable. Were you
> >>>>>>>>>>> using the build script that I provided?
> >>>>>>>>>>
> >>>>>>>>>> No. I took your script and got the important environment variables. But I missed the LLVM_* ones. Usually we only build hotspot so we don't have this jdk script.
> >>>>>>>>>>
> >>>>>>>>>> Now that I have the LLVM_* variables it's even worse:
> >>>>>>>>>>
> >>>>>>>>>> /home/cthaling/8003868/src/share/vm/oops/oop.hpp:72: error: cast from type ‘markOopDesc* const volatile*’ to type ‘markOopDesc**’ casts away constness
> >>>>>>>>>>
> >>>>>>>>>> It's probably this guy: -Wcast-qual
> >>>>>>>>>
> >>>>>>>>> Got it:
> >>>>>>>>>
> >>>>>>>>> $ java -version
> >>>>>>>>> java version "1.8.0-ea"
> >>>>>>>>> Java(TM) SE Runtime Environment (build 1.8.0-ea-b65)
> >>>>>>>>> OpenJDK 64-Bit Shark VM (build 25.0-b11-internal-jvmg, mixed mode)
> >>>>>>>>
> >>>>>>>> I ran the compiler regression tests and Shark crashes in 5091921:
> >>>>>>>>
> >>>>>>>> cthaling at intelsdv03.us.oracle.com:~/8003868/test$ jtreg -workDir:$EXPORTHOME/jtreg -reportDir:$EXPORTHOME/jtreg -testjdk:$JAVA_HOME -verbose:summary compiler/5091921/
> >>>>>>>> Directory "/export/twisti/jtreg/scratch" not found: creating
> >>>>>>>> Passed: compiler/5091921/Test5091921.java
> >>>>>>>> Passed: compiler/5091921/Test6186134.java
> >>>>>>>> Passed: compiler/5091921/Test6196102.java
> >>>>>>>> Passed: compiler/5091921/Test6357214.java
> >>>>>>>> Passed: compiler/5091921/Test6559156.java
> >>>>>>>> Passed: compiler/5091921/Test6753639.java
> >>>>>>>> Passed: compiler/5091921/Test6850611.java
> >>>>>>>> Passed: compiler/5091921/Test6890943.java
> >>>>>>>> Passed: compiler/5091921/Test6897150.java
> >>>>>>>> Passed: compiler/5091921/Test6905845.java
> >>>>>>>> Passed: compiler/5091921/Test6931567.java
> >>>>>>>> /net/sqenfs-1.us.oracle.com/export1/comp/vm/tool/jtreg/execution/linux/bin/jtreg: line 139: 27784 Segmentation fault (core dumped) "${JT_JAVA}" $javaOpts -Dprogram=`basename "$0"` -jar "${JT_HOME}/lib/jtreg.jar" $jtregOpts
> >>>>>>>>
> >>>>>>>> You can also run all them with a simple make in test/ by setting:
> >>>>>>>>
> >>>>>>>> PRODUCT_HOME=$JAVA_HOME
> >>>>>>>> TESTDIRS=compiler
> >>>>>>>
> >>>>>>> Alright, I found another fairly grave bug that I would like to include a
> >>>>>>> fix for: apparently, a 'fence' is not enough of a memory barrier for
> >>>>>>> volatile putfield/getfield (duh). Which basically broke all of j.u.c.
> >>>>>>> LLVM offers atomic loads and stores, which seem to be exactly what is
> >>>>>>> needed for volatile field access. However, it does not provide helper
> >>>>>>> functions for those in llvm::IRBuilder so I wrote my own. This should be
> >>>>>>> the final patch for now (unless you find something else).
> >>>>>>>
> >>>>>>> http://cr.openjdk.java.net/~rkennke/shark/webrev.03/
> >>>>>>
> >>>>>> Hmm. Maybe I did something wrong but I've already rebuilt twice:
> >>>>>>
> >>>>>> $ java -Xcomp -version
> >>>>>> Value type size is target-dependent. Ask TLI.
> >>>>>> UNREACHABLE executed at /usr/local/src/llvm-3.1.src/include/llvm/CodeGen/ValueTypes.h:257!
> >>>>>> Stack dump:
> >>>>>> 0. Running pass 'X86 DAG->DAG Instruction Selection' on function '@"java.lang.System::getProperty"'
> >>>>>> Aborted (core dumped)
> >>>>>
> >>>>> Arg! The last couple of changes I did only with LLVM3.2, where the
> >>>>> problem disappears. Apparently, LLVM3.1 (and pre) don't deal well with
> >>>>> atomic load/store :-( I re-introduced the CreateMemoryBarrier call and
> >>>>> use that for SHARK_LLVM_VERSION <= 31.
> >>>>>
> >>>>> http://cr.openjdk.java.net/~rkennke/shark/webrev.04/
> >>>>>
> >>>>> Hope that works better :-)
> >>>>
> >>>> I'm so sorry but...
> >>>>
> >>>> /export/twisti/build/8003868/build/linux_amd64_shark/product/libjvm.so: undefined reference to `SharkBuilder::memory_barrier()'
> >>>
> >>> Gaaa, what the... I thought I did clean rebuilds with both llvm3.2 and
> >>> llvm3.1, but apparently not (maybe I shouldn't work after 1am). This
> >>> (hopefully final final) patch re-instates the missing memory_barrier()
> >>> method:
> >>>
> >>>
> >>> http://cr.openjdk.java.net/~rkennke/shark/webrev.05/
> >>>
> >>> Sorry for the messy back-and-forth.
> >>
> >> Again, so sorry:
> >>
> >> $ java -Xcomp -version
> >> LLVM ERROR: Program used external function 'llvm.memory.barrier' which could not be resolved!
> >>
> >> Send a new patch tomorrow after some sleep ;-)
> >
> > Yeah, apparently 'replaced by' means that the old thing (the intrinsics)
> > are indeed gone ;-)
> >
> > The problem is that the correct way to implement it (atomic load/store)
> > doesn't work, the 'old way' (the memory_barrier() intrinsic call)
> > doesn't work either, I also tried CreateAtomicRMW() which is probably
> > not 100% correct, but would have done the job, but that doesn't work
> > either (it throws the same error as the atomic load/store ...). The
> > problem seems to only appear on 64bit volatile access.
> >
> > I don't know a really good solution that doesn't require jumping through
> > big hoops, and I don't feel like working around LLVM bugs like this,
> > especially since LLVM 3.2 (which should be released real soon now) works
> > just fine. If you have a suggestion, please let me know, otherwise I
> > suggest the following patch, which gets rid of all the LLVM31 blocks
> > again and creates atomic load/store instructions (and requires LLVM 3.2
> > which can be found here
> > http://llvm.org/svn/llvm-project/llvm/branches/release_32/ ).
> >
> > http://cr.openjdk.java.net/~rkennke/shark/webrev.06/
>
> That's a reasonable thing to do given the tentative release date of December 16th. While running the compiler regression tests I got a couple of failures. You might want to address them with separate bugs.
Hi Twisti,
I see that you committed the patch, thanks a lot for this! However,
there is still the small jdk patch missing:
http://cr.openjdk.java.net/~rkennke/shark/webrev-jdk-00/
Can I push this myself or do you want to do it?
Cheers,
Roman
More information about the hotspot-dev
mailing list