[LLVMdev] dyld: lazy symbol binding failed: fast lazy bind offset out of range
baldrick at free.fr
Tue Oct 23 00:41:46 PDT 2012
> I looks like the code is calling dlopen() on LLVMPolly.so and it or something it links against has an initializer. The initialer is run before dlopen() returns and the crash is in the initializer. The message:
> dyld: fast lazy bind offset out of range (53437, max=7640) in image /sw/lib/gcc4.7/libexec/gcc/x86_64-apple-darwin12.2.0/4.7.2/cc1
> means the initializer called something which made its way into the cc1 binary, probably it called some weak symbol which resolved in the main executable cc1.
this reminds me of something that came up recently (PR13981): both polly and
gcc are dynamically linked against libcloog (and usually against different
versions). When dlopen loads polly it resolves the libcloog symbols using
the version of libcloog that was loaded when gcc (the main binary) was started
up, causing all kinds of mayhem on linux. Maybe this could cause the kind of
issue you are seeing, especially if the two libcloog library versions differ?
> In gdb you can set:
> (gdb) set env DYLD_PRINT_INITIALIZERS
> and dyld will print out the address of each initializer before calling it. You can then run again and set a break point at that address and step through it.
> Also, the lazy bind info for LLVMPolly.so was a little odd in that all symbols were marked "flat-namespace". On MacOSX we normally record in which dylib each symbol should be found. The use of flat-namespace tells dyld to look everywhere for those symbols. This can cause the wrong symbol to be used at runtime. Was LLVMPolly.so build with -flat_namespace or -undefined dynamic_lookup?
> On Oct 22, 2012, at 8:07 AM, Jack Howarth wrote:
>> On Mon, Oct 22, 2012 at 09:48:59AM -0400, Jack Howarth wrote:
>>> On Sun, Oct 21, 2012 at 06:00:36PM -0700, Nick Kledzik wrote:
>>>> Some binary has an initializer which dyld is calling. Somehow the initializer gets to:
>>>> #4 0x0000000100f3b2c0 in Json::Value::maxUInt ()
>>>> which is calling a function in another dylib for the first time. When you call a function in another dylib, you actually jump through a (lazy) pointer. The pointer initially points to a helper which loads an index parameter specifying which function to bind, then jumps into dyld. Dyld is erroring because the index is out of range.
>>>> 1) The binary is corrupt. You can check that by running:
>>>> dyldinfo -lazy_bind to see the lazy pointers
>>> I attached dragonegg_so_lazy_bind.txt and LLVMPolly_so_lazy_bind.txt to http://llvm.org/bugs/show_bug.cgi?id=14140.
>>> Is there anything significant in those lists of lazy pointers relative to the failure we are seeing? FYI, I see the same
>>> failure for both x86_64-apple-darwin11 and x86_64-apple-darwin12 with current llvm/polly/dragonegg svn using Xcode 4.5.1's
>>>> 2) Something else is wrong with the initializer and it happens to be jumping to the helper (that jumps into dyld), but the index parameter is garbage. You'll need to step though this in the debugger.
>>>> From the backtrace quoted below, can you suggest a logical place to set the breakpoint as the starting point to step through this code?
>>> Thanks in advance.
>> I have attached a gdb walk log (LLVMPolly_load_walk.log) to http://llvm.org/bugs/show_bug.cgi?id=14140 which
>> steps from llvm-3.2/lib/Support/CommandLine.cpp, line 259. The tail end of the log is as follows...
>> llvm::sys::DynamicLibrary::getPermanentLibrary (filename=0x144002188 "/sw/opt/llvm-3.2/lib/LLVMPolly.so", errMsg=0x7fff5fbfe700) at /sw/src/fink.build/llvm32-3.2-0/llvm-3.2/lib/Support/DynamicLibrary.cpp:77
>> 77 void *handle = dlopen(filename, RTLD_LAZY|RTLD_GLOBAL);
>> Reading symbols for shared libraries ... done
>> dyld: lazy symbol binding failed: fast lazy bind offset out of range (53437, max=7640) in image /sw/lib/gcc4.7/libexec/gcc/x86_64-apple-darwin12.2.0/4.7.2/cc1
>> dyld: fast lazy bind offset out of range (53437, max=7640) in image /sw/lib/gcc4.7/libexec/gcc/x86_64-apple-darwin12.2.0/4.7.2/cc1
>> Let me know if there is something else I can provide to help debug this nested plugin loading issue on darwin.
>> ps This is for an llvm/polly debug cmake build with assertions on as well as dragonegg built with -g.
>>>> On Oct 21, 2012, at 9:27 AM, Jack Howarth wrote
>>>>> While enhancing the fink llvm32 packaging to support the polly tool, I ran into
>>>>> a dyld failure when the dragonegg plugin tries to load LLVMPolly.so plugin...
>>>>> dyld: lazy symbol binding failed: fast lazy bind offset out of range (39257, max=7640) in image /sw/lib/gcc4.7/libexec/gcc/x86_64-apple-darwin11.4.2/4.7.2/cc1
>>>>> dyld: fast lazy bind offset out of range (39257, max=7640) in image /sw/lib/gcc4.7/libexec/gcc/x86_64-apple-darwin11.4.2/4.7.2/cc1
>>>>> Trace/BPT trap
>>>>> which backtraces as...
>>>>> (gdb) bt
>>>>> #0 0x00007fff5fc0106d in __dyld_dyld_fatal_error ()
>>>>> #1 0x00007fff5fc048c8 in __dyld__ZN4dyld4haltEPKc ()
>>>>> #2 0x00007fff5fc04948 in __dyld__ZN4dyld18fastBindLazySymbolEPP11ImageLoaderm ()
>>>>> #3 0x00007fff90af7716 in dyld_stub_binder_ ()
>>>>> #4 0x0000000100f3b2c0 in Json::Value::maxUInt ()
>>>>> #5 0x00007fff5fc0fda6 in __dyld__ZN16ImageLoaderMachO18doModInitFunctionsERKN11ImageLoader11LinkContextE ()
>>>>> #6 0x00007fff5fc0faf2 in __dyld__ZN16ImageLoaderMachO16doInitializationERKN11ImageLoader11LinkContextE ()
>>>>> #7 0x00007fff5fc0d2e4 in __dyld__ZN11ImageLoader23recursiveInitializationERKNS_11LinkContextEjRNS_21InitializerTimingListE ()
>>>>> #8 0x00007fff5fc0e0b7 in __dyld__ZN11ImageLoader15runInitializersERKNS_11LinkContextERNS_21InitializerTimingListE ()
>>>>> #9 0x00007fff5fc031b9 in __dyld__ZN4dyld15runInitializersEP11ImageLoader ()
>>>>> #10 0x00007fff5fc09657 in __dyld_dlopen ()
>>>>> #11 0x00007fff90af795b in dlopen ()
>>>>> #12 0x0000000142f3888e in llvm::sys::DynamicLibrary::getPermanentLibrary ()
>>>>> #13 0x0000000142f256bd in llvm::PluginLoader::operator= ()
>>>>> #14 0x000000014280fbda in llvm::cl::opt<llvm::PluginLoader, false, llvm::cl::parser<std::string> >::handleOccurrence ()
>>>>> #15 0x0000000142f1730c in CommaSeparateAndAddOccurence ()
>>>>> #16 0x0000000142f1395f in ProvideOption ()
>>>>> #17 0x0000000142f1224b in llvm::cl::ParseCommandLineOptions ()
>>>>> #18 0x000000014280d3ff in InitializeBackend ()
>>>>> #19 0x000000014280c191 in llvm_emit_globals ()
>>>>> This will be difficult to testcase since the FSF gcc 4.7.2 compiler is involved but I was hoping that you might be able
>>>>> to make some guesses about the origin of the problem from the current polly and dragonegg source code. There doesn't
>>>>> seem to be any reports of this error type (but then again few plugins, dragonegg, indirectly call additional plugins,
>>>>> LLVMPolly. Thanks in advance for any hints.
>>> LLVM Developers mailing list
>>> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu
More information about the llvm-dev