help-octave
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: LLVM/JIT on MacOS X


From: Ben Abbott
Subject: Re: LLVM/JIT on MacOS X
Date: Thu, 12 Jun 2014 13:36:01 -0400

On Jun 12, 2014, at 11:36 AM, Stefan Mahr <address@hidden> wrote:

>>>>>>>>>>> 
>>>>>>>>>>> Ben Abbott wrote: 
>>>>>>>>>>>> 
>>>>>>>>>>>> Carlo,
>>>>>>>>>>>> 
>>>>>>>>>>>> Have you been successful getting JIT to work on Mac OS X?  When I 
>>>>>>>>>>>> try, I encounter the error below.
>>>>>>>>>>>> 
>>>>>>>>>>>> Ben
>>>>>>>>>>>> 
>>>>>>>>>>>> [...]
>>>>>>>>>>>> 
>>>>>>>>>>>> Undefined symbols for architecture x86_64:
>>>>>>>>>>>> "llvm::verifyModule(llvm::Module const&, 
>>>>>>>>>>>> llvm::VerifierFailureAction, std::basic_string<char, 
>>>>>>>>>>>> std::char_traits<char>, std::allocator<char> >*)", referenced from:
>>>>>>>>>>>> tree_jit::optimize(llvm::Function*)      in 
>>>>>>>>>>>> libcorefcn.a(corefcn_libcorefcn_la-pt-jit.o)
>>>>>>>>>>>> "llvm::raw_fd_ostream::raw_fd_ostream(char const*, 
>>>>>>>>>>>> std::basic_string<char, std::char_traits<char>, 
>>>>>>>>>>>> std::allocator<char> >&, unsigned int)", referenced from:
>>>>>>>>>>>> tree_jit::optimize(llvm::Function*)      in 
>>>>>>>>>>>> libcorefcn.a(corefcn_libcorefcn_la-pt-jit.o)
>>>>>>>>>>>> "llvm::ExecutionEngine::createJIT(llvm::Module*, 
>>>>>>>>>>>> std::basic_string<char, std::char_traits<char>, 
>>>>>>>>>>>> std::allocator<char> >*, llvm::JITMemoryManager*, 
>>>>>>>>>>>> llvm::CodeGenOpt::Level, bool, llvm::Reloc::Model, 
>>>>>>>>>>>> llvm::CodeModel::Model)", referenced from:
>>>>>>>>>>>> tree_jit::initialize()      in 
>>>>>>>>>>>> libcorefcn.a(corefcn_libcorefcn_la-pt-jit.o)
>>>>>>>>>>>> ld: symbol(s) not found for architecture x86_64
>>>>>>>>>>>> 
>>>>>>>>>>> 
>>>>>>>>>>> Hi Ben,
>>>>>>>>>>> 
>>>>>>>>>>> your error message sounds like a incompatibility with newer LLVM 
>>>>>>>>>>> versions and should be fixed with my attached patch in this bug 
>>>>>>>>>>> report: http://savannah.gnu.org/bugs/?41061
>>>>>>>>>>> 
>>>>>>>>>>> As always, a rebased changeset to a more recent repository state 
>>>>>>>>>>> can be found here: http://inversethought.com/hg/octave-nkf/nkf-ready
>>>>>>>>>>> 
>>>>>>>>>>> Could you try the patch?
>>>>>>>>>>> 
>>>>>>>>>>> Stefan
>>>>>>>>>> 
>>>>>>>>>> Patching my default branch with the changeset from bug 41061 ...
>>>>>>>>>> 
>>>>>>>>>> $ patch -p1 < ~/Downloads/llvm-3.4-3.5pre.patch 
>>>>>>>>>> patching file configure.ac
>>>>>>>>>> patching file libinterp/corefcn/jit-typeinfo.cc
>>>>>>>>>> patching file libinterp/corefcn/jit-util.h
>>>>>>>>>> patching file libinterp/corefcn/pt-jit.cc
>>>>>>>>>> Hunk #2 succeeded at 2064 (offset -8 lines).
>>>>>>>>>> Hunk #3 succeeded at 2187 (offset -8 lines).
>>>>>>>>>> patching file libinterp/corefcn/pt-jit.h
>>>>>>>>>> patching file m4/acinclude.m4
>>>>>>>>>> $ make -j4
>>>>>>>>>> <snip>
>>>>>>>>>> </snip>
>>>>>>>>>> Undefined symbols for architecture x86_64:
>>>>>>>>>> "llvm::verifyModule(llvm::Module const&, 
>>>>>>>>>> llvm::VerifierFailureAction, std::basic_string<char, 
>>>>>>>>>> std::char_traits<char>, std::allocator<char> >*)", referenced from:
>>>>>>>>>>  tree_jit::optimize(llvm::Function*)      in 
>>>>>>>>>> libcorefcn.a(corefcn_libcorefcn_la-pt-jit.o)
>>>>>>>>>> "llvm::raw_fd_ostream::raw_fd_ostream(char const*, 
>>>>>>>>>> std::basic_string<char, std::char_traits<char>, std::allocator<char> 
>>>>>>>>>> >&, unsigned int)", referenced from:
>>>>>>>>>>  tree_jit::optimize(llvm::Function*)      in 
>>>>>>>>>> libcorefcn.a(corefcn_libcorefcn_la-pt-jit.o)
>>>>>>>>>> "llvm::ExecutionEngine::createJIT(llvm::Module*, 
>>>>>>>>>> std::basic_string<char, std::char_traits<char>, std::allocator<char> 
>>>>>>>>>> >*, llvm::JITMemoryManager*, llvm::CodeGenOpt::Level, bool, 
>>>>>>>>>> llvm::Reloc::Model, llvm::CodeModel::Model)", referenced from:
>>>>>>>>>>  tree_jit::initialize()      in 
>>>>>>>>>> libcorefcn.a(corefcn_libcorefcn_la-pt-jit.o)
>>>>>>>>>> ld: symbol(s) not found for architecture x86_64
>>>>>>>>>> 
>>>>>>>>>> Ben
>>>>>>>>> 
>>>>>>>>> bootstrap and configure is needed before starting make.
>>>>>>>> 
>>>>>>>> Configure ran (automatically) after the patch was applied.  Just to be 
>>>>>>>> sure, I ran bootstrap & configure manually, and still obtained the 
>>>>>>>> same result.
>>>>>>>> 
>>>>>>>> Ben
>>>>>>>> 
>>>>>>> 
>>>>>>> I'm really confused that compiling works, but linking fails. This 
>>>>>>> usually happens in the past (for me) if there are multiple versions of 
>>>>>>> LLVM installed (Debian/Ubuntu) or some older headerfiles were not 
>>>>>>> deleted during update (seen with MXE).
>>>>>>> 
>>>>>>> E.g. for linking error with llvm::verifyModule. The header file for my 
>>>>>>> LLVM-3.3 installation is in <llvm/Analysis/Verifier.h>. For llvm 3.5 
>>>>>>> it's in <llvm/IR/Verifier.h> with different declaration. If compiling 
>>>>>>> with 3.3 headers and linking to 3.5 lib I would expect the same error 
>>>>>>> as you get. Same for llvm::raw_fd_ostream::raw_fd_ostream, since the 
>>>>>>> last argument is not an integer type anymore in newer LLVM versions. 
>>>>>>> However, this doesn't explain a link error for 
>>>>>>> llvm::ExecutionEngine::createJIT. I see no difference in declarions 
>>>>>>> between LLVM versions 3.3 to 3.5.
>>>>>>> 
>>>>>>> To summarise: If you only have LLVM 3.3 installed, I have no idea. 
>>>>>>> Sorry.
>>>>>>> 
>>>>>>> Stefan
>>>>>> 
>>>>>> hmmm, it does appear there is more than one llvm present.
>>>>>> 
>>>>>> $ port installed *llvm*
>>>>>> The following ports are currently installed:
>>>>>> llvm-3.3 @3.3_1
>>>>>> llvm-3.3 @3.3_4 (active)
>>>>>> llvm-gcc42 @2336.11_1 (active)
>>>>>> llvm_select @0.2_0
>>>>>> llvm_select @1.0_0 (active)
>>>>>> 
>>>>>> I've deactivate llvm-gcc42
>>>>>> 
>>>>>> $ sudo port deactivate llvm-gcc42
>>>>>> Password:
>>>>>> --->  Deactivating llvm-gcc42 @2336.11_1
>>>>>> --->  Cleaning llvm-gcc42
>>>>>> 
>>>>>> And ran configure/make again. Unfortunately, I'm still seeing the same 
>>>>>> error.
>>>>>> 
>>>>>> Looking at "config.log", there is one problem with llvm ....
>>>>>> 
>>>>>> configure:15039: checking llvm/IR/Function.h presence
>>>>>> configure:15039: /opt/local/bin/g++-mp-4.7 -E -D__STDC_CONSTANT_MACROS 
>>>>>> -D__STDC_LIMIT_MACROS -isystem /opt/local/libexec/llvm-3.3/include 
>>>>>> -D_THREAD_SAFE -I/opt/local/include conftest.cpp
>>>>>> configure:15039: $? = 0
>>>>>> configure:15039: result: yes
>>>>>> configure:15039: checking for llvm/IR/Function.h
>>>>>> configure:15039: result: yes
>>>>>> configure:15057: checking llvm/Support/IRBuilder.h usability
>>>>>> configure:15057: /opt/local/bin/g++-mp-4.7 -c  -pipe -O2 -m64 
>>>>>> -D_THREAD_SAFE -pthread -D__STDC_CONSTANT_MACROS -D__STDC_LIMIT_MACROS 
>>>>>> -isystem /opt/local/libexec/llvm-3.3/include -D_THREAD_SAFE 
>>>>>> -I/opt/local/include conftest.cpp >&5
>>>>>> conftest.cpp:84:36: fatal error: llvm/Support/IRBuilder.h: No such file 
>>>>>> or directory
>>>>>> 
>>>>>> Ben
>>>>> 
>>>>> The output of config.log looks OK. configure checks for presence of
>>>>> llvm/Support/IRBuilder.h, but with LLVM 3.3 the header file is located
>>>>> in llvm/IR/IRBuilder.h.
>>>> 
>>>> hmmm ... I checked config.log a second time ... looks like I examined the 
>>>> wrong file previously. 
>>>> 
>>>> configure:15075: checking llvm/Target/TargetData.h usability
>>>> configure:15075: /opt/local/bin/g++-mp-4.7 -c  -pipe -O2 -m64 
>>>> -D_THREAD_SAFE -pthread -D__STDC_CONSTANT_MACROS -D__STDC_LIMIT_MACROS 
>>>> -isystem /opt/local/libexec/llvm-3.3/include -D_THREAD_SAFE 
>>>> -I/opt/local/include conftest.cpp >&5
>>>> conftest.cpp:85:36: fatal error: llvm/Target/TargetData.h: No such file or 
>>>> directory
>>>> 
>>>> I checked my install and have confirmed that TargetData.h does not exist.
>>>> 
>>>> Ben
>>> 
>>> No, same as before. For LLVM 3.3 the correct header file is 
>>> llvm/IR/Datalayout.h, so the configure check is absolutely right.
>> 
>> Do you mean that I don't need llvm/Target/TargetData.h because I have 
>> llvm/IR/Datalayout.h?
>> 
>> Ben
>> 
> 
> Yes, unfortunately the LLVM guys love to move and rename files and change the 
> API.
> 
> The include for llvm/IR/Datalayout.h is in pt-jit.cc:
> 
> #ifdef HAVE_LLVM_IR_DATALAYOUT_H
> #include <llvm/IR/DataLayout.h>
> #elif defined(HAVE_LLVM_DATALAYOUT_H)
> #include <llvm/DataLayout.h>
> #else
> #include <llvm/Target/TargetData.h>
> #endif
> 
> So you should have only one of this files in your include path.
> 
> Did you already check if MacOS provides it's own version of the llvm library, 
> independant from the one of macports?

Yes. MacOS does have its own llvm.

$ ls -l /usr/bin/ll*
-rwxr-xr-x  1 root  wheel  14224 Jun  2 23:52 /usr/bin/lldb
lrwxr-xr-x  1 root  wheel      7 Oct 22  2013 /usr/bin/llvm-g++ -> clang++
lrwxr-xr-x  1 root  wheel      5 Oct 22  2013 /usr/bin/llvm-gcc -> clang

But, I assume the llcm libs are part of the clang libs.

Ben




reply via email to

[Prev in Thread] Current Thread [Next in Thread]