Reid Spencer
2005-Jan-11 04:40 UTC
[Fwd: Re: [LLVMdev] Shared library building problems on Darwin]
Michael, I've implemented a LOADABLE_MODULE feature in the makefiles: http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20050110/023147.html The approach taken is almost what you described below. However, I want to retain the distinction between a "regular" shared library and one that can be dlopened. So, if you specify SHARED_LIBRARY=1 you get a regular shared library intended for dynamic linking (i.e. system automated linking at runtime such as with ld.so) that *may* also be dlopenable on some systems, but isn't gauranteed to be openable. If you want to gaurantee that it can be dlopened then you must also specify LOADABLE_MODULE=1 (in addition to SHARED_LIBRARY=1). Note that the name of the resulting library is $(LIBRARYNAME).la not lib$(LIBRARYNAME).la. This is on purpose because on some systems loadable modules can't be used by the dynamic linker (HP/UX for sure, and probably Darwin). Not having the "lib" prefix means it won't be found by any linker -l options and thus not screw up the linker. The name difference also implies "you must use dlopen/dlsym to access this library". Hopefully this is suitable for you. I've already added LOADABLE_MODULE=1 to the Makefile in lib/Transforms/Hello. Would you please be so kind as to try it on Darwin and see if this solves your "opt -load mymodule.so" problem? Thanks, Reid. -----Forwarded Message-----> From: Michael McCracken <michael.mccracken at gmail.com> > To: Reid Spencer <reid at x10sys.com> > Subject: Re: [LLVMdev] Shared library building problems on Darwin > Date: Mon, 10 Jan 2005 16:54:14 -0800 > > Hi, > > Thanks, that sounds great. > In case it didn't already come up, I might also suggest making > LOADABLE_MODULE=1 just set SHARED_LIBRARY on non-darwinian systems so > you can just always use LOADABLE_MODULE when writing plugins, and > it'll always do the right thing. Then amending the pass-writing > tutorial to just use LOADABLE_MODULE will be OK. > > (This fall-through to shared-library building is identical to what we > just added to SCons, and I think it makes the most sense, although > someone who doesn't have to worry about the difference may not > agree...) > > Also, don't worry about timeliness on my account. Just as long as it's > on the radar, I'm content. > > Thanks, > -mike > > On Mon, 10 Jan 2005 15:25:51 -0800, Reid Spencer <reid at x10sys.com> wrote: > > Michael, > > > > Chris and I have been having a discussion about this topic on the side. > > I won't bore you with the details, but it boils down to this: > > > > 1. libtool (mklib) makes a distinction between "shared library" and > > "loadable module". > > 2. On most systems shared library and loadable module are the same > > thing. That is, the dynamic linker uses dlopen/dlsym on a shared > > object just as a user program might. > > 3. The above isn't true for Darwin > > 4. Our makefiles insist on making "shared library" not loadable modules > > when you specify SHARED_LIBRARY=1 in a makefile. > > > > In most Unixes it doesn't matter but Darwin is "special" so it matters > > and you end up with the scenario you're seeing. The solution is to > > permit both kinds of links to be done (one with -module and one > > without). So, I'm going to add support for LOADABLE_MODULE=1 setting to > > Makefile.rules so we can build both types of libraries. This should > > clear up the problem on Darwin. > > > > Unfortunately, my time is constrained right now. I'll attend to this as > > soon as I'm able. > > > > Thanks for your patience. > > > > Reid. > > > > > > On Thu, 2005-01-06 at 18:39, Michael McCracken wrote: > > > Hi, a while back I wrote that the llvm makefiles didn't create the > > > correct kind of file for use on darwin with -load. > > > > > > Since then, both the shared library and makefile system have been > > > overhauled significantly. > > > So I checked again - as updated from CVS, the current makefiles don't > > > build the right object type on darwin. > > > > > > If you follow the advice of 'Writing an LLVM Pass" tutorial, you will > > > create (say) llvm/lib/Transforms/PMFBuild/ and fill it with the > > > appropriate files. > > > > > > Then, running 'make' in llvm/lib/Transforms/PMFBuild generates a bunch > > > of files in llvm/Debug/lib/: > > > % cd llvm/Debug/lib > > > % ll *PMF* > > > -rw-r--r-- 1 mike staff 385584 Jan 6 18:16 PMFBuild.o > > > -rwxr-xr-x 1 mike staff 1053828 Jan 6 18:16 libPMFBuild.0.0.0.dylib > > > lrwxr-xr-x 1 mike staff 23 Jan 6 18:16 > > > libPMFBuild.0.dylib -> libPMFBuild.0.0.0.dylib > > > -rw-r--r-- 1 mike staff 385600 Jan 6 18:16 libPMFBuild.a > > > lrwxr-xr-x 1 mike staff 23 Jan 6 18:16 > > > libPMFBuild.dylib -> libPMFBuild.0.0.0.dylib > > > -rwxr-xr-x 1 mike staff 921 Jan 6 18:16 libPMFBuild.la > > > > > > I think this means I'm running in OBJDIR==SRCDIR mode, but you tell me. > > > I'm not up to speed on the new makefile system. > > > > > > Note the file types: > > > % file *PMF* > > > PMFBuild.o: Mach-O object ppc > > > libPMFBuild.0.0.0.dylib: Mach-O dynamically linked shared library ppc > > > libPMFBuild.0.dylib: symbolic link to `libPMFBuild.0.0.0.dylib' > > > libPMFBuild.a: current ar archive > > > libPMFBuild.dylib: symbolic link to `libPMFBuild.0.0.0.dylib' > > > libPMFBuild.la: ASCII English text > > > > > > This may look OK, but we see that -load doesn't complain, but fails to > > > actually load it. > > > (Take my word that -pmfbuild would work if not for this, it does later.) > > > > > > So I tried the same hack I found earlier: > > > > > > --- Makefile.rules 3 Jan 2005 17:42:57 -0000 1.287 > > > +++ Makefile.rules 7 Jan 2005 02:29:22 -0000 > > > @@ -326,7 +326,7 @@ > > > LTCompile.CXX = $(LIBTOOL) $(LibTool.Flags) --mode=compile $(Compile.CXX) > > > BCCompile.CXX = $(LLVMGXXWITHPATH) $(CPP.Flags) $(CompileCommonOpts) \ > > > $(CXX.Flags) -c > > > -Link = $(LIBTOOL) $(LibTool.Flags) --mode=link $(CXX) $(CPP.Flags) \ > > > +Link = $(LIBTOOL) $(LibTool.Flags) --mode=link $(CXX) > > > -module $(CPP.Flags) \ > > > $(CompileCommonOpts) $(LD.Flags) $(Strip) > > > Relink = $(LIBTOOL) $(LibTool.Flags) --mode=link $(CXX) $(CPP.Flags) \ > > > $(CompileCommonOpts) > > > > > > Removing the old *PMF* libraries and objects in llvm/Debug/lib/ and > > > re-running make in lib/Transforms/PMFBuild/ gives the following list > > > of files: > > > > > > % ll *PMF* > > > -rw-r--r-- 1 mike staff 385584 Jan 6 18:31 PMFBuild.o > > > -rwxr-xr-x 1 mike staff 1052968 Jan 6 18:31 libPMFBuild.0.0.0.so > > > lrwxr-xr-x 1 mike staff 20 Jan 6 18:31 > > > libPMFBuild.0.so -> libPMFBuild.0.0.0.so > > > -rw-r--r-- 1 mike staff 385600 Jan 6 18:31 libPMFBuild.a > > > -rwxr-xr-x 1 mike staff 910 Jan 6 18:31 libPMFBuild.la > > > lrwxr-xr-x 1 mike staff 20 Jan 6 18:31 libPMFBuild.so > > > -> libPMFBuild.0.0.0.so > > > > > > % file *PMF* > > > PMFBuild.o: Mach-O object ppc > > > libPMFBuild.0.0.0.so: Mach-O bundle ppc > > > libPMFBuild.0.so: symbolic link to `libPMFBuild.0.0.0.so' > > > libPMFBuild.a: current ar archive > > > libPMFBuild.la: ASCII English text > > > libPMFBuild.so: symbolic link to `libPMFBuild.0.0.0.so' > > > > > > Note that the file is now a Mach-O bundle, not a Dynamically Linked > > > Library, and > > > it has the appropriate .so extension (thanks to libtool - Apple > > > actually seems to prefer to build bundles with no extension) > > > > > > This is important, because on darwin, dynamically linked libraries and > > > 'loadable modules' are not the same thing, as they are in object file > > > formats in use elsewhere. In order to be loaded dynamically after the > > > program is already running, the file has to be a "Mach-O Bundle". > > > > > > So with these file types, it works: > > > /Users/mike/Documents/hpcl/LLVM/cfe-src/install/bin/gcc -g > > > tests/daxpy/daxpy.c -o tests/daxpy/daxpy > > > /Users/mike/Documents/hpcl/LLVM/llvm-cvs/Debug/bin//opt -f -stats > > > -debug-pass=Structure -load > > > /Users/mike/Documents/hpcl/LLVM/llvm-cvs/Debug/lib//libPMFBuild.so > > > -buildpmf -o tests/daxpy/daxpy-opt tests/daxpy/daxpy.bc > > > Pass Arguments: -buildpmf -verify > > > Target Data Layout > > > Module Pass Manager > > > Function Pass Manager > > > Build PMF Structures > > > -- Build PMF Structures > > > Immediate Dominators Construction > > > Dominator Set Construction > > > -- Immediate Dominators Construction > > > Module Verifier > > > -- Dominator Set Construction > > > -- Module Verifier > > > Bytecode Writer > > > --Bytecode Writer > > > > > > I'm not sure how best to fit this fix into the LLVM makefiles, but it > > > seems like there should be a way to toss in that -module option for > > > linking shared libraries only on Darwin. However, like I said, I'm not > > > up on the new system, and I just found the Link command and tried my > > > old hack. > > > > > > Hope this helps, > > > -mike > > > > > > >-------------- next part -------------- A non-text attachment was scrubbed... Name: signature.asc Type: application/pgp-signature Size: 189 bytes Desc: This is a digitally signed message part URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20050110/f57a43f6/attachment.sig>
Michael McCracken
2005-Jan-11 07:57 UTC
[Fwd: Re: [LLVMdev] Shared library building problems on Darwin]
Yep, it sounds like a good solution, and it works for me - thanks! -mike On Mon, 10 Jan 2005 20:40:34 -0800, Reid Spencer <reid at x10sys.com> wrote:> Michael, > > I've implemented a LOADABLE_MODULE feature in the makefiles: > > http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20050110/023147.html > > The approach taken is almost what you described below. However, I want > to retain the distinction between a "regular" shared library and one > that can be dlopened. So, if you specify SHARED_LIBRARY=1 you get a > regular shared library intended for dynamic linking (i.e. system > automated linking at runtime such as with ld.so) that *may* also be > dlopenable on some systems, but isn't gauranteed to be openable. If you > want to gaurantee that it can be dlopened then you must also specify > LOADABLE_MODULE=1 (in addition to SHARED_LIBRARY=1). Note that the name > of the resulting library is $(LIBRARYNAME).la not lib$(LIBRARYNAME).la. > This is on purpose because on some systems loadable modules can't be > used by the dynamic linker (HP/UX for sure, and probably Darwin). Not > having the "lib" prefix means it won't be found by any linker -l options > and thus not screw up the linker. The name difference also implies "you > must use dlopen/dlsym to access this library". Hopefully this is > suitable for you. > > I've already added LOADABLE_MODULE=1 to the Makefile in > lib/Transforms/Hello. Would you please be so kind as to try it on Darwin > and see if this solves your "opt -load mymodule.so" problem? > > Thanks, > > Reid. > > -----Forwarded Message----- > > From: Michael McCracken <michael.mccracken at gmail.com> > > To: Reid Spencer <reid at x10sys.com> > > Subject: Re: [LLVMdev] Shared library building problems on Darwin > > Date: Mon, 10 Jan 2005 16:54:14 -0800 > > > > Hi, > > > > Thanks, that sounds great. > > In case it didn't already come up, I might also suggest making > > LOADABLE_MODULE=1 just set SHARED_LIBRARY on non-darwinian systems so > > you can just always use LOADABLE_MODULE when writing plugins, and > > it'll always do the right thing. Then amending the pass-writing > > tutorial to just use LOADABLE_MODULE will be OK. > > > > (This fall-through to shared-library building is identical to what we > > just added to SCons, and I think it makes the most sense, although > > someone who doesn't have to worry about the difference may not > > agree...) > > > > Also, don't worry about timeliness on my account. Just as long as it's > > on the radar, I'm content. > > > > Thanks, > > -mike > > > > On Mon, 10 Jan 2005 15:25:51 -0800, Reid Spencer <reid at x10sys.com> wrote: > > > Michael, > > > > > > Chris and I have been having a discussion about this topic on the side. > > > I won't bore you with the details, but it boils down to this: > > > > > > 1. libtool (mklib) makes a distinction between "shared library" and > > > "loadable module". > > > 2. On most systems shared library and loadable module are the same > > > thing. That is, the dynamic linker uses dlopen/dlsym on a shared > > > object just as a user program might. > > > 3. The above isn't true for Darwin > > > 4. Our makefiles insist on making "shared library" not loadable modules > > > when you specify SHARED_LIBRARY=1 in a makefile. > > > > > > In most Unixes it doesn't matter but Darwin is "special" so it matters > > > and you end up with the scenario you're seeing. The solution is to > > > permit both kinds of links to be done (one with -module and one > > > without). So, I'm going to add support for LOADABLE_MODULE=1 setting to > > > Makefile.rules so we can build both types of libraries. This should > > > clear up the problem on Darwin. > > > > > > Unfortunately, my time is constrained right now. I'll attend to this as > > > soon as I'm able. > > > > > > Thanks for your patience. > > > > > > Reid. > > > > > > > > > On Thu, 2005-01-06 at 18:39, Michael McCracken wrote: > > > > Hi, a while back I wrote that the llvm makefiles didn't create the > > > > correct kind of file for use on darwin with -load. > > > > > > > > Since then, both the shared library and makefile system have been > > > > overhauled significantly. > > > > So I checked again - as updated from CVS, the current makefiles don't > > > > build the right object type on darwin. > > > > > > > > If you follow the advice of 'Writing an LLVM Pass" tutorial, you will > > > > create (say) llvm/lib/Transforms/PMFBuild/ and fill it with the > > > > appropriate files. > > > > > > > > Then, running 'make' in llvm/lib/Transforms/PMFBuild generates a bunch > > > > of files in llvm/Debug/lib/: > > > > % cd llvm/Debug/lib > > > > % ll *PMF* > > > > -rw-r--r-- 1 mike staff 385584 Jan 6 18:16 PMFBuild.o > > > > -rwxr-xr-x 1 mike staff 1053828 Jan 6 18:16 libPMFBuild.0.0.0.dylib > > > > lrwxr-xr-x 1 mike staff 23 Jan 6 18:16 > > > > libPMFBuild.0.dylib -> libPMFBuild.0.0.0.dylib > > > > -rw-r--r-- 1 mike staff 385600 Jan 6 18:16 libPMFBuild.a > > > > lrwxr-xr-x 1 mike staff 23 Jan 6 18:16 > > > > libPMFBuild.dylib -> libPMFBuild.0.0.0.dylib > > > > -rwxr-xr-x 1 mike staff 921 Jan 6 18:16 libPMFBuild.la > > > > > > > > I think this means I'm running in OBJDIR==SRCDIR mode, but you tell me. > > > > I'm not up to speed on the new makefile system. > > > > > > > > Note the file types: > > > > % file *PMF* > > > > PMFBuild.o: Mach-O object ppc > > > > libPMFBuild.0.0.0.dylib: Mach-O dynamically linked shared library ppc > > > > libPMFBuild.0.dylib: symbolic link to `libPMFBuild.0.0.0.dylib' > > > > libPMFBuild.a: current ar archive > > > > libPMFBuild.dylib: symbolic link to `libPMFBuild.0.0.0.dylib' > > > > libPMFBuild.la: ASCII English text > > > > > > > > This may look OK, but we see that -load doesn't complain, but fails to > > > > actually load it. > > > > (Take my word that -pmfbuild would work if not for this, it does later.) > > > > > > > > So I tried the same hack I found earlier: > > > > > > > > --- Makefile.rules 3 Jan 2005 17:42:57 -0000 1.287 > > > > +++ Makefile.rules 7 Jan 2005 02:29:22 -0000 > > > > @@ -326,7 +326,7 @@ > > > > LTCompile.CXX = $(LIBTOOL) $(LibTool.Flags) --mode=compile $(Compile.CXX) > > > > BCCompile.CXX = $(LLVMGXXWITHPATH) $(CPP.Flags) $(CompileCommonOpts) \ > > > > $(CXX.Flags) -c > > > > -Link = $(LIBTOOL) $(LibTool.Flags) --mode=link $(CXX) $(CPP.Flags) \ > > > > +Link = $(LIBTOOL) $(LibTool.Flags) --mode=link $(CXX) > > > > -module $(CPP.Flags) \ > > > > $(CompileCommonOpts) $(LD.Flags) $(Strip) > > > > Relink = $(LIBTOOL) $(LibTool.Flags) --mode=link $(CXX) $(CPP.Flags) \ > > > > $(CompileCommonOpts) > > > > > > > > Removing the old *PMF* libraries and objects in llvm/Debug/lib/ and > > > > re-running make in lib/Transforms/PMFBuild/ gives the following list > > > > of files: > > > > > > > > % ll *PMF* > > > > -rw-r--r-- 1 mike staff 385584 Jan 6 18:31 PMFBuild.o > > > > -rwxr-xr-x 1 mike staff 1052968 Jan 6 18:31 libPMFBuild.0.0.0.so > > > > lrwxr-xr-x 1 mike staff 20 Jan 6 18:31 > > > > libPMFBuild.0.so -> libPMFBuild.0.0.0.so > > > > -rw-r--r-- 1 mike staff 385600 Jan 6 18:31 libPMFBuild.a > > > > -rwxr-xr-x 1 mike staff 910 Jan 6 18:31 libPMFBuild.la > > > > lrwxr-xr-x 1 mike staff 20 Jan 6 18:31 libPMFBuild.so > > > > -> libPMFBuild.0.0.0.so > > > > > > > > % file *PMF* > > > > PMFBuild.o: Mach-O object ppc > > > > libPMFBuild.0.0.0.so: Mach-O bundle ppc > > > > libPMFBuild.0.so: symbolic link to `libPMFBuild.0.0.0.so' > > > > libPMFBuild.a: current ar archive > > > > libPMFBuild.la: ASCII English text > > > > libPMFBuild.so: symbolic link to `libPMFBuild.0.0.0.so' > > > > > > > > Note that the file is now a Mach-O bundle, not a Dynamically Linked > > > > Library, and > > > > it has the appropriate .so extension (thanks to libtool - Apple > > > > actually seems to prefer to build bundles with no extension) > > > > > > > > This is important, because on darwin, dynamically linked libraries and > > > > 'loadable modules' are not the same thing, as they are in object file > > > > formats in use elsewhere. In order to be loaded dynamically after the > > > > program is already running, the file has to be a "Mach-O Bundle". > > > > > > > > So with these file types, it works: > > > > /Users/mike/Documents/hpcl/LLVM/cfe-src/install/bin/gcc -g > > > > tests/daxpy/daxpy.c -o tests/daxpy/daxpy > > > > /Users/mike/Documents/hpcl/LLVM/llvm-cvs/Debug/bin//opt -f -stats > > > > -debug-pass=Structure -load > > > > /Users/mike/Documents/hpcl/LLVM/llvm-cvs/Debug/lib//libPMFBuild.so > > > > -buildpmf -o tests/daxpy/daxpy-opt tests/daxpy/daxpy.bc > > > > Pass Arguments: -buildpmf -verify > > > > Target Data Layout > > > > Module Pass Manager > > > > Function Pass Manager > > > > Build PMF Structures > > > > -- Build PMF Structures > > > > Immediate Dominators Construction > > > > Dominator Set Construction > > > > -- Immediate Dominators Construction > > > > Module Verifier > > > > -- Dominator Set Construction > > > > -- Module Verifier > > > > Bytecode Writer > > > > --Bytecode Writer > > > > > > > > I'm not sure how best to fit this fix into the LLVM makefiles, but it > > > > seems like there should be a way to toss in that -module option for > > > > linking shared libraries only on Darwin. However, like I said, I'm not > > > > up on the new system, and I just found the Link command and tried my > > > > old hack. > > > > > > > > Hope this helps, > > > > -mike > > > > > > > > > > > > > > _______________________________________________ > LLVM Developers mailing list > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu > http://mail.cs.uiuc.edu/mailman/listinfo/llvmdev > > > >-- Michael McCracken UCSD CSE PhD Student San Diego Supercomputer Center http://www.cse.ucsd.edu/~mmccrack/
Bryan Turner
2005-Jan-11 22:33 UTC
[LLVMdev] Race Condition discovery algorithm using LLVM?
Hello, I'm new to LLVM and have spent the last few weeks reading up on the documents & code. I would like to run a project idea by the senior members before I start digging and find out I'm over my head. ;) After reading the Data Structure Analysis paper, it occurred to me that a relatively minor modification to the pass should allow for Race Condition discovery. By this I mean generate warnings when data structures are accessed in a potentially thread un-safe manner such as reading/writing by two threads without synchronization. The algorithm would work something like this: Data is input alongside the Data Structure Analysis pass (perhaps using a separate input file if necessary) which includes a list of Thread Roots, and a list of pairs of lock/unlock functions. During the DSA pass, structures which are accessed outside lock/unlock pairs are marked as being Unsafe. At the end of the pass, the Thread Root functions' DSA graphs are compared. Those structures which are accessed by multiple Thread Roots (and marked Unsafe) are reported as potential race conditions. A separate pass may be needed to generate useful warning information (where in each thread's call stack the potential race condition occurs) such that a programmer can determine if the 'potential' is a real threat or not. My first examination at the code led me to /llvm/Analysis/DataStructure/Local.cpp, which has a function GraphBuilder::visitCallSite(). This appears to be the ideal spot to add a hook for additional DSA processing, such as checking lock/unlock function calls. Also, hooks at the beginning and end of the pass, as well as a hook in the visitor tree for setting flags. My question is; for a first-time project in LLVM, is this going to get me over my head? I have significant application & system design experience, but no compiler experience. Also, do you feel the data generated from such a pass would be useful, or would there be too many false-positives in a typical application? Thanks! --Bryan bryan.turner at pobox.com