Building WebKit natives with debug symbols on 32 bit Linux

Peter Levart peter.levart at
Fri Jun 6 14:44:04 UTC 2014

Hi David,

Thanks for pointing me in the right direction. Although I've almost lost 
hope before finally succeeding. I tried various other approaches on the 
way, but finally, what I did was very simple. It could be possible to 
create a separate configuration (i.e. LINUX32) with some additional 
tweaks to some other scripts (some scripts use the configuration name as 
a logic token), but here I present a simpler variant with changes 
applied to current LINUX configuration.

In order to cross-compile on 64 bit Ubuntu 14.04 for 32 bit Linux target 
you need to patch some makefiles and build scripts:

Besides above changes to build scripts, the following environment 
variables have to be defined (in $HOME/.profile):


You need to install the i386 versions of runtime libraries needed to 
execute the 32 bit JVM. I don't have the names at hand, but the list 
could quickly be compiled by trying to run some Java GUI program using 
32 bit JVM and then using apt-file on the missing library file to 
find-out the name of the package. /Wash, rinse, repeat.../ until all 
i386 JDK runtime dependencies are installed.

In addition, i386 versions of *-dev packages needed for compiling 
openjfx have to be installed. This is the most tricky part to get right, 
since installing them pulls-in some i386 runtime library dependencies 
which conflict with x86_64 versions of libraries that might already be 
installed on 64 bit system. "apt-get install" is clever on Ubuntu 14.04 
and *REMOVES* the installed dependency sub-tree of x86_64 packages 
before installing a conflicting i386 version of requested package. For 
some of *-dev:i386 packages this means that even the tools needed for 
building (gcc, binutils) get de-installed, but you can re-install them 
later without conflicts (that's strange, but APT packaging seem to work 
with at least 3 levels of dependency strengths). The most important part 
is that the following packages:

     gcc, gcc-4.8, g++, g++-48, binutils (with dependencies)

are kept at x86_64 architecture (that's the whole purpose of 
cross-compilation). In addition, you also need to install the following 
packages, to be able to produce 32bit binaries with 64bit tools:

gcc-multilib, gcc-4.8-multilib, g++-multilib, g++-4.8-multilib

I think that's all. With this changes I managed to produce 32 bit with debugging symbols included (it's nearly 1.7 GiB 
!!!). You need at least 10 GiB of free RAM when attempting such build 
(the "ld" process grows to about 8 GiB of resident space when collecting 
object files into a shared library - no wonder it can't be done with 32 
bit "ld").

With 32bit debug build of openjfx I reproduced the crash mentioned in 
( and got the following 
backtrace from gdb:

#7  <signal handler called>
#8  0x6591cc7e in WTFCrash () at 
#9  0x65950c3f in WTF::OSAllocator::reserveUncommitted (bytes=126976, 
usage=WTF::OSAllocator::UnknownUsage, writable=true, executable=false, 
     at ../../../../src/main/native/Source/WTF/wtf/OSAllocatorPosix.cpp:58
#10 0x65ac9033 in WTF::PageAllocationAligned::allocate (size=65536, 
alignment=65536, usage=WTF::OSAllocator::UnknownUsage, writable=true)
#11 0x65b5bff7 in JSC::DFG::Allocator<JSC::DFG::Node>::allocateSlow 
(this=0x70df9f78) at 
#12 0x65b5a7c0 in JSC::DFG::Allocator<JSC::DFG::Node>::freeListAllocate 
(this=0x70df9f78) at 
#13 0x65b58e4a in JSC::DFG::Allocator<JSC::DFG::Node>::allocate 
(this=0x70df9f78) at 
#14 0x65b53d53 in operator new (size=84, allocator=...) at 
#15 0x65b5bbb1 in JSC::DFG::Graph::addNode<JSC::DFG::NodeType, 
JSC::CodeOrigin, JSC::DFG::OpInfo> (this=0x68a9effc, type=0, 
_DFG_value1=@0x68a9de3c: JSC::DFG::Phi,
     _DFG_value2=..., _DFG_value3=...) at 
#16 0x65ce34a7 in JSC::DFG::CPSRethreadingPhase::addPhiSilently 
(this=0x68a9dfd8, block=0x88d5970, codeOrigin=..., variable=0x89501b0)
#17 0x65ce532c in 
(this=0x68a9dfd8, block=0x88d5970, codeOrigin=..., variable=0x89501b0, 
#18 0x65ce5492 in 
(this=0x68a9dfd8, codeOrigin=..., variable=0x89501b0, index=5)
#19 0x65ce4c43 in 
(this=0x68a9dfd8, node=0x43f389d0, variable=0x89501b0, idx=5)
#20 0x65ce3559 in JSC::DFG::CPSRethreadingPhase::canonicalizeGetLocal 
(this=0x68a9dfd8, node=0x43f389d0)
#21 0x65ce373c in 
JSC::DFG::CPSRethreadingPhase::canonicalizeLocalsInBlock (this=0x68a9dfd8)
#22 0x65ce3828 in 
JSC::DFG::CPSRethreadingPhase::canonicalizeLocalsInBlocks (this=0x68a9dfd8)
#23 0x65ce30f4 in JSC::DFG::CPSRethreadingPhase::run (this=0x68a9dfd8) 
#24 0x65ce5f5e in JSC::DFG::runAndLog<JSC::DFG::CPSRethreadingPhase> 
(phase=...) at 
#25 0x65ce50bd in JSC::DFG::runPhase<JSC::DFG::CPSRethreadingPhase> 
(graph=...) at 
#26 0x65ce2cfd in JSC::DFG::performCPSRethreading (graph=...) at 
#27 0x65b4e08a in JSC::DFG::compile 
(compileMode=JSC::DFG::CompileFunction, exec=0x6343bf48, 
codeBlock=0x43ce1308, jitCode=..., jitCodeWithArityCheck=0x5f52bb04,
     osrEntryBytecodeIndex=0) at 
#28 0x65b4d983 in JSC::DFG::tryCompileFunction (exec=0x6343bf48, 
codeBlock=0x43ce1308, jitCode=..., jitCodeWithArityCheck=..., 
#29 0x659ee25b in JSC::jitCompileFunctionIfAppropriate (exec=0x6343bf48, 
codeBlock=..., jitCode=..., jitCodeWithArityCheck=..., 
jitType=JSC::JITCode::DFGJIT, bytecodeIndex=0,
     effort=JSC::JITCompilationCanFail) at 
#30 0x659ee4a4 in JSC::prepareFunctionForExecution (exec=0x6343bf48, 
codeBlock=..., jitCode=..., jitCodeWithArityCheck=..., 
jitType=JSC::JITCode::DFGJIT, bytecodeIndex=0,
     kind=JSC::CodeForCall) at 
#31 0x659ebe91 in JSC::FunctionExecutable::compileForCallInternal 
(this=0x5f52bad8, exec=0x6343bf48, scope=0x61d5dbf0, 
jitType=JSC::JITCode::DFGJIT, bytecodeIndex=0)
#32 0x659eb659 in JSC::FunctionExecutable::compileOptimizedForCall 
(this=0x5f52bad8, exec=0x6343bf48, scope=0x61d5dbf0, bytecodeIndex=0)
#33 0x65ae4398 in JSC::FunctionExecutable::compileOptimizedFor 
(this=0x5f52bad8, exec=0x6343bf48, scope=0x61d5dbf0, bytecodeIndex=0, 
#34 0x65ae05cf in JSC::FunctionCodeBlock::compileOptimized 
(this=0x4b40f818, exec=0x6343bf48, scope=0x61d5dbf0, bytecodeIndex=0)
#35 0x659a406d in JSC::cti_optimize (args=0x68a9fcc0) at 
---Type <return> to continue, or q <return> to quit---
#36 0x659a0ce1 in JSC::tryCacheGetByID (callFrame=0x0, 
codeBlock=0xfffffffb, returnAddress=..., baseValue=..., 
propertyName=..., slot=..., stubInfo=0x0)
#37 0x6343bad8 in ?? ()
#38 0x65994da7 in JSC::JITCode::execute (this=0x5f5237ac, 
stack=0x70df9cf0, callFrame=0x6343a058, vm=0x70df3df8)
     at ../../../../src/main/native/Source/JavaScriptCore/jit/JITCode.h:135
#39 0x659922bc in JSC::Interpreter::executeCall (this=0x70df9ce8, 
callFrame=0x6253f994, function=0x5eedf618, callType=JSC::CallTypeJS, 
callData=..., thisValue=..., args=...)
#40 0x659dc14c in JSC::call (exec=0x6253f994, functionObject=..., 
callType=JSC::CallTypeJS, callData=..., thisValue=..., args=...)
#41 0x64a34c05 in WebCore::JSMainThreadExecState::call (exec=0x6253f994, 
functionObject=..., callType=JSC::CallTypeJS, callData=..., 
thisValue=..., args=...)
#42 0x64a7e0cb in WebCore::ScheduledAction::executeFunctionInContext 
(this=0x524f9388, globalObject=0x6253f838, thisValue=..., 
#43 0x64a7e312 in WebCore::ScheduledAction::execute (this=0x524f9388, 
document=0x62c589f0) at 
#44 0x64a7de62 in WebCore::ScheduledAction::execute (this=0x524f9388, 
context=0x62c58a50) at 
#45 0x651930df in WebCore::DOMTimer::fired (this=0x442283f8) at 
#46 0x65360730 in WebCore::ThreadTimers::sharedTimerFiredInternal 
(this=0x70dd3a50) at 
#47 0x65360663 in WebCore::ThreadTimers::sharedTimerFired () at 
#48 0x65303abe in Java_com_sun_webkit_Timer_twkFireTimerEvent 
(env=0x6890552c, clazz=0x68aa00c0)

...from first look at relevant sources it seems that the crash is a 
result of not being able to allocate 64 KiB block with 64 KiB alignment 
which means that a 128 KiB block is attempted to be mmap-ed and it 
fails. The WebKit Javascript engine's answer to not being able to 
allocate memory is simply a CRASH()? It seems so. Compared to HOTSPOT 
this is very primitive, but such is the reality of this fast-grown code 
unfortunately which is 4-times as big as the HOTSPOT. Anybody venturing 
into using WebKit to build and run complex GUI-s should be told that 
(and offered Java + JavaFX  instead as alternative ;-)...

Regards, Peter

On 05/28/2014 05:17 PM, David Hill wrote:
> On 5/28/14, May 28, 9:47 AM, Peter Levart wrote:
>> Hi Again,
>> The idea that comes to my mind is the following: would it be possible 
>> to cross-compile the openjfx on the 64 bit Linux using 64 bit tools, 
>> but targeted at 32 bit Linux? What would have to be changed in build 
>> environment to accomplish that? I imagine the environment would have 
>> to have access to headers and development libraries of the 32 bit 
>> Linux system - that's no problem - I would just mount the root of a 
>> 32 bit Ubuntu to some directory, but then I would have to force all 
>> tools to use that path instead of the running system one. And the gcc 
>> would have to be given some cross-compiling options too, I guess...
> In theory this is easy :-)
> We already have a cross compile mechanism in gradle, used for the 
> Linux arm and Android builds. (see buildSrc/armv* for example).
> You would start with one of the build files as a template - 
> linux.gradle, sed/copy it to linux32.gradle  with something like:
> cd buildSrc
> sed -s 's/LINUX/LINUX32/' linux.gradle > linux32.gradle
> After that, it would be a matter of tweaking the cc and ld flags 
> contained in the file, to tell gcc to be 32bit and to use the right 
> libraries. Looks like adding -m32 to commonFlags is all it should need:
> // A set of common parameters to use for both compiling and linking
> def commonFlags = [
>         "-m32",  // <<<<< 32 bit output
>         "-fno-strict-aliasing", "-fPIC", "-fno-omit-frame-pointer", // 
> optimization flags
>         "-W", "-Wall", "-Wno-unused", "-Wno-parentheses", 
> "-Werror=implicit-function-declaration"] // warning flags
> and tweak where the libraries are installed:
> // Libraries end up in the sdk/rt/lib/$OS_ARCH directory for Linux
> LINUX32.libDest = "lib/i386"
> LINUX32.arch = "i386"  # << this is an addition not needed in 
> linux.gradle
> You specify the crossbuild like this:
>     gradle  -PUSE_DEPEND=false -PBUILD_NATIVES=true 
> But...
>    looks like we need to disable
>         LINUX32.compileFXPackager = false;
> because build.gradle has some stuff hardcoded in it (would need a jira 
> for that if people care).
> And - you need the i386 development packages installed. I had already 
> added this to get my cross compilers to work,
>     apt-get install libstdc++6:i386
> but needed to add other to get more of the build to complete and ran 
> into apt-get telling me all my i386 deps were broken. Perhaps it would 
> work better if I was on a newer distro (I am on 12.04)
> has the list of packages, in theory, you just need to add :i386 to 
> each one.
> Since I can't get the i386 packages, at this point you are on your own 
> :-)
> Dave
>> Peter
>> On 05/28/2014 03:39 PM, Peter Levart wrote:
>>> Hi,
>>> I'm new to this list and I have searched the archives, but haven't 
>>> found a discussion about this. If I have missed it, please just 
>>> direct me to the archived thread.
>>> I'm trying to debug a crash in a native part of WebKit-JavaFX 
>>> bindings ( I have 
>>> only been able to reproduce it on 32 bit platforms (Windows and 
>>> Linux so far). I'm trying to build the WebKit natives with debugging 
>>> symbols, so that I can pin-point the location of the crash in code 
>>> using gdb. I have managed to do this successfully on 64 bit Linux, 
>>> producing 1.7 GB large But on 32 bit Linux, where 
>>> only I can reproduce the crash, the build fails with linker error: 
>>> "memory exhausted" when trying to assemble the from 
>>> object files. So far I have not been successful in my attempts to 
>>> circumvent this build error. I fave tried the following:
>>> - added '--no-keep-memory' to linker options. The man page says 
>>> about it:
>>>            ld normally optimizes for speed over memory usage by 
>>> caching the symbol tables of input files in memory.  This option 
>>> tells ld to instead optimize for memory usage, by
>>>            rereading the symbol tables as necessary.  This may be 
>>> required if ld runs out of memory space while linking a large 
>>> executable.
>>> - booted the 64bit Ubuntu system and chrooted into a 32 bit 
>>> environment. Some say that with 64 bit kernel, each 32 bit 
>>> user-space process has more address space than with 32 bit kernel.
>>> - both of the above
>>> My question to the list is: How do you guys produce 32 bit 
>>> with debugging symbols? Have you been able to? Do 
>>> you have any special tricks in your sleeves?
>>> Regards,
>>> Peter

More information about the openjfx-dev mailing list