I've attached a patch which has got JIT compilation working (for me at least!) on Android. It turns out that the problem was a bunch of intrinsic __aeabi* functions which reside in libgcc.a rather than libc.so so are not available unless explicitly linked in, so it's rather similar to the StatSymbols hack. I moved the StatSymbols code into ExecutionEngine.cpp rather than JITMemoryManager.cpp since it's required for both the JIT and MCJIT, and deleted the code in RTDyldMemoryManager.cpp which did the same thing for fewer functions. They should now be picked up through the sys::DynamicLibrary::AddSymbol mechanism in both cases. The symbols required for ARM/Android are then added by an identical hack just below. There's one other minor change since setLastModificationAndAccessTime can't be supported on Android; all relevant system calls are missing from the C library. I've therefore changed the code to fail at runtime rather than compile-time here. James -------------- next part -------------- A non-text attachment was scrubbed... Name: llvm-android.patch Type: text/x-patch Size: 9100 bytes Desc: not available URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20131111/06731924/attachment.bin>
On 11 November 2013 01:45, James Lyon <jameslyon0 at gmail.com> wrote:> I've attached a patch which has got JIT compilation working (for me at > least!) on Android. It turns out that the problem was a bunch of intrinsic > __aeabi* functions which reside in libgcc.a rather than libc.so so are not > available unless explicitly linked in, so it's rather similar to the > StatSymbols hack. >Hi James, I may be wrong, but I remember LLVM treating "androideabi" as "aeabi", which is wrong. This may be the source of your problem, and I think we should teach Clang/LLVM to understand the Android ABI in its full form. There's one other minor change since setLastModificationAndAccessTime can't> be supported on Android; all relevant system calls are missing from the C > library. I've therefore changed the code to fail at runtime rather than > compile-time here.This sounds worse... I'd rather we fixed the behaviour than the cause... cheers, --renato -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20131111/78c934f9/attachment.html>
I think __aeabi is ARM EABI not Android EABI, e.g. http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ihi0043d/index.html. The problem is not that the __aeabi functions are missing on Android, they're defined in libgcc.a, the problem is that they don't get linked in unless they're referenced. I don't know how this problem is avoided on ARM/Linux/glibc normally, but something pulls them in there according to my testing. The way to avoid disabling setLastModificationAndAccessTime is to rewrite it to take a path and use utime/utimes instead (incidentally the setLastModificationAndAccessTime function is only used by llvm-ar). I don't know why it was written the way it is now; i.e. I'm not sure what systems have futimens but not futimes and vice-verse. I would guess that the utime function should be universal though. There's a HAVE_UTIME_H test in Config.h but it doesn't seem to be used anywhere. James On 11/11/13 08:30, Renato Golin wrote:> On 11 November 2013 01:45, James Lyon <jameslyon0 at gmail.com > <mailto:jameslyon0 at gmail.com>> wrote: > > I've attached a patch which has got JIT compilation working (for > me at least!) on Android. It turns out that the problem was a > bunch of intrinsic __aeabi* functions which reside in libgcc.a > rather than libc.so so are not available unless explicitly linked > in, so it's rather similar to the StatSymbols hack. > > > Hi James, > > I may be wrong, but I remember LLVM treating "androideabi" as "aeabi", > which is wrong. This may be the source of your problem, and I think we > should teach Clang/LLVM to understand the Android ABI in its full form. > > > There's one other minor change since > setLastModificationAndAccessTime can't be supported on Android; > all relevant system calls are missing from the C library. I've > therefore changed the code to fail at runtime rather than > compile-time here. > > > This sounds worse... I'd rather we fixed the behaviour than the cause... > > cheers, > --renato-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20131111/30415fc4/attachment.html>
I've got a number of problems with this patch. First, we have plans to pry apart the remaining strands connecting JIT with MCJIT, so I don't want to do anything that reconnects them. That is, I'm against moving things from RTDyldMemoryManager into ExecutionEngine. Second, unless I'm reading it wrong, this relies on static constructors. That causes headaches and is against the LLVM coding standards. Third, I don't think sys::DynamicLibrary::AddSymbol() is the right way to go here. I know I have recently suggested using that function, but I was wrong. That function exposes symbols to the entire host program as if they were present in a locally loaded shared library. For some programs that may be acceptable behavior, but it cannot be the default behavior for LLVM. In particular, MCJIT may not even be generating code for local execution. The default RTDyldMemoryManager implementation selectively exposes symbols to execution engines for which it is used, and MCJIT/RuntimeDyld clients performing remote execution can (and should) override this behavior. We probably also need to fix the addGlobalMapping support for MCJIT. -Andy -----Original Message----- From: llvmdev-bounces at cs.uiuc.edu [mailto:llvmdev-bounces at cs.uiuc.edu] On Behalf Of James Lyon Sent: Sunday, November 10, 2013 5:45 PM To: LLVM Dev Subject: [LLVMdev] Android JIT patch I've attached a patch which has got JIT compilation working (for me at least!) on Android. It turns out that the problem was a bunch of intrinsic __aeabi* functions which reside in libgcc.a rather than libc.so so are not available unless explicitly linked in, so it's rather similar to the StatSymbols hack. I moved the StatSymbols code into ExecutionEngine.cpp rather than JITMemoryManager.cpp since it's required for both the JIT and MCJIT, and deleted the code in RTDyldMemoryManager.cpp which did the same thing for fewer functions. They should now be picked up through the sys::DynamicLibrary::AddSymbol mechanism in both cases. The symbols required for ARM/Android are then added by an identical hack just below. There's one other minor change since setLastModificationAndAccessTime can't be supported on Android; all relevant system calls are missing from the C library. I've therefore changed the code to fail at runtime rather than compile-time here. James
On 11/11/13 17:42, Kaylor, Andrew wrote:> I've got a number of problems with this patch. > > First, we have plans to pry apart the remaining strands connecting JIT with MCJIT, so I don't want to do anything that reconnects them. That is, I'm against moving things from RTDyldMemoryManager into ExecutionEngine.The direction of LLVM is not something I'm in a position to comment on. The problem of locating these functions is common to the JIT and MCJIT though (and even the interpreter for Linux/x86).> > Second, unless I'm reading it wrong, this relies on static constructors. That causes headaches and is against the LLVM coding standards.My bad, I just ripped off the existing implementation. I guess the JITMemoryManager part is destined to go away along with the JIT.> > Third, I don't think sys::DynamicLibrary::AddSymbol() is the right way to go here. I know I have recently suggested using that function, but I was wrong. That function exposes symbols to the entire host program as if they were present in a locally loaded shared library. For some programs that may be acceptable behavior, but it cannot be the default behavior for LLVM. In particular, MCJIT may not even be generating code for local execution.Well, that's a bit trickier. Somewhere there needs to be a hook to find these functions (and make the linker include them). For remote execution I suppose that the user would have to load a shared object supporting this functionality into the target address space and then query it for the address of such functions. I'm assuming that the user isn't expected to solve this problem if they're not using remote execution.> > The default RTDyldMemoryManager implementation selectively exposes symbols to execution engines for which it is used, and MCJIT/RuntimeDyld clients performing remote execution can (and should) override this behavior. We probably also need to fix the addGlobalMapping support for MCJIT.The situation right now is actually more complicated than this: since llvm-config --libs MCJIT includes the LLVMJIT library the JITMemoryManager hack is always there (subject to being dropped by ld?) so these symbols are actually exposed globally. As I said, I basically just borrowed from the existing code and added the functions which won't work without glibc on ARM. sys::DynamicLibrary::SearchForAddressOfSymbol only seems to be used by stuff under lib/ExecutionEngine, so making that find functions which are expected to be in libc during normal linking doesn't seem unreasonable. James> > -Andy > > -----Original Message----- > From: llvmdev-bounces at cs.uiuc.edu [mailto:llvmdev-bounces at cs.uiuc.edu] On Behalf Of James Lyon > Sent: Sunday, November 10, 2013 5:45 PM > To: LLVM Dev > Subject: [LLVMdev] Android JIT patch > > I've attached a patch which has got JIT compilation working (for me at > least!) on Android. It turns out that the problem was a bunch of intrinsic __aeabi* functions which reside in libgcc.a rather than libc.so so are not available unless explicitly linked in, so it's rather similar to the StatSymbols hack. > > I moved the StatSymbols code into ExecutionEngine.cpp rather than JITMemoryManager.cpp since it's required for both the JIT and MCJIT, and deleted the code in RTDyldMemoryManager.cpp which did the same thing for fewer functions. They should now be picked up through the sys::DynamicLibrary::AddSymbol mechanism in both cases. The symbols required for ARM/Android are then added by an identical hack just below. > > There's one other minor change since setLastModificationAndAccessTime can't be supported on Android; all relevant system calls are missing from the C library. I've therefore changed the code to fail at runtime rather than compile-time here. > > James