Alexey Samsonov
2014-Aug-19 23:40 UTC
[LLVMdev] [RFC] Adding functions for unaligned load/store to Support for JIT/RuntimeDyld
Hi, I've noticed that certain parts of LLVM (RuntimeDyld, JIT) use unaligned memory accesses to load/store pointers or just 32- or 64-bit integers. Technically, this is undefined behavior, and UBSan reports errors when this happens. Even if we believe we're running some x86-specific code (e.g. RuntimeDyldELF::resolveX86_64Relocation) what matters for unaligned acceses in source code is the host architecture, not the target one. What do you think of adding T unaligned_load(const void *addr) void unaligned_store(void *addr, T value); to support headers? We can then provide the default memcpy implementations of these methods and, optionally, straightforward implementations for x86 hosts (and disable UBSan alignment checks for the latter). -- Alexey Samsonov vonosmas at gmail.com -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140819/67b5bf2f/attachment.html>
Lang Hames
2014-Aug-20 19:04 UTC
[LLVMdev] [RFC] Adding functions for unaligned load/store to Support for JIT/RuntimeDyld
Hi Alexey, That sounds good to me. RuntimeDyldMachO already has writeBytesUnaligned for this reason, though it's use is somewhat patchy. We should also make better use of the facilities provided in llvm/Support/Endian.h - it has alignment and endianness aware definitions for common primitive types. Cheers, Lang. On Tue, Aug 19, 2014 at 4:40 PM, Alexey Samsonov <vonosmas at gmail.com> wrote:> Hi, > > I've noticed that certain parts of LLVM (RuntimeDyld, JIT) use unaligned > memory accesses to load/store pointers or just 32- or 64-bit integers. > Technically, this is undefined behavior, and UBSan reports errors when this > happens. Even if we believe we're running some x86-specific code (e.g. > RuntimeDyldELF::resolveX86_64Relocation) what matters for unaligned acceses > in source code is the host architecture, not the target one. > > What do you think of adding > > T unaligned_load(const void *addr) > void unaligned_store(void *addr, T value); > > to support headers? We can then provide the default memcpy implementations > of these methods and, optionally, straightforward implementations for x86 > hosts (and disable UBSan alignment checks for the latter). > > -- > Alexey Samsonov > vonosmas at gmail.com > > _______________________________________________ > LLVM Developers mailing list > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140820/20bd136e/attachment.html>
Alexey Samsonov
2014-Aug-20 23:03 UTC
[LLVMdev] [RFC] Adding functions for unaligned load/store to Support for JIT/RuntimeDyld
Hi Lang, On Wed, Aug 20, 2014 at 12:04 PM, Lang Hames <lhames at gmail.com> wrote:> Hi Alexey, > > That sounds good to me. RuntimeDyldMachO already has writeBytesUnaligned > for this reason, though it's use is somewhat patchy. We should also make > better use of the facilities provided in llvm/Support/Endian.h - it has > alignment and endianness aware definitions for common primitive types. >Cool, thanks. Looks like Endian.h already has read/write methods with configurable alignment. I'll try to plug them into ExecutionEngine then.> > Cheers, > Lang. > > > On Tue, Aug 19, 2014 at 4:40 PM, Alexey Samsonov <vonosmas at gmail.com> > wrote: > >> Hi, >> >> I've noticed that certain parts of LLVM (RuntimeDyld, JIT) use unaligned >> memory accesses to load/store pointers or just 32- or 64-bit integers. >> Technically, this is undefined behavior, and UBSan reports errors when this >> happens. Even if we believe we're running some x86-specific code (e.g. >> RuntimeDyldELF::resolveX86_64Relocation) what matters for unaligned acceses >> in source code is the host architecture, not the target one. >> >> What do you think of adding >> >> T unaligned_load(const void *addr) >> void unaligned_store(void *addr, T value); >> >> to support headers? We can then provide the default memcpy >> implementations of these methods and, optionally, straightforward >> implementations for x86 hosts (and disable UBSan alignment checks for the >> latter). >> >> -- >> Alexey Samsonov >> vonosmas at gmail.com >> >> _______________________________________________ >> LLVM Developers mailing list >> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu >> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev >> >> >-- Alexey Samsonov vonosmas at gmail.com -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140820/354bc548/attachment.html>