Sébastien Marie <semarie-open...@latrappe.fr> writes: > Hi, > > I am currently working to port rustc to openbsd (amd64).
Awesome! > > I have wrote a compatibility layer for openbsd (using some part of > http://www.ntecs.de/blog/2014/07/29/rust-ported-to-dragonflybsd/ and > https://github.com/qbit/rust-cross-openbsd), and I am able to > cross-compile a rustc openbsd-binary from linux. My fork was a pretty sad attempt at getting it ported.. I stopped with the lack of segmented stacks. That said Dave Huseby made it much further: https://github.com/dhuseby/rust-cross-openbsd ! IIRC he ran into another issue with linking.. and is currently working on https://github.com/dhuseby/rust-cross-bitrig . He would probably be the person to talk to about furthering the OpenBSD port! > > I have done lot of rewrite for have a more valid openbsd support in > rustc. For explain to not-rustc-dev-aware persons: rustc don't used > header files (.h). It needs to rewriting them in rustc-language, and use > that for FFI calls. It is methods definition, types definition (size, > struct members...). > > And as openbsd isn't same as freebsd or dragonfly, just adding > conditional-compilation at same places of freebsd/dragonfly isn't > enought. > > > The rustc code is here: https://github.com/semarie/rust/tree/openbsd, > and for compilation I mainly take inspiration from > http://github.jfet.org/Rust_cross_bootstrapping.html. > > My build framework is somehow fragile: it use shared filesystem between > a linux (debian wheezy) and an openbsd (openbsd-current, somewhere at > starting 5.7-BETA). For openbsd target, the objects files are build on > linux, and linking (using remote-command) on openbsd (modulo some parts > based on C code, whole compiled under openbsd, like LLVM, compiler-rt, > libbacktrace...). > > As result, the rustc openbsd-binary is mainly fonctionnal: I am able to > build simples rustc programs (like hello-world.hs, or guess.rs [from > rustc guide]). > > So I started to build rustc it-self, fully under openbsd. The current > status is the build isn't complete because of random panic or coredump > (more details below). > > The current goal is to have a working rustc binary compiled from > openbsd, and next check it with the rustc testing framework. > > Some workarounds/problems: > - I compile without jemalloc (the jemalloc "make check" crash > systematically. The issue wasn't investigated as jemalloc seems to be > optional) > > - I don't implement 'load_self' function which live under > 'libstd/sys/unix/os.rs'. The function should return the filepath of > the current running program (and it isn't trivial as there are > possibles races). > > As result, rustc *must* have a --sysroot argument to work, else it > would panic (and so I hack the build process to pass the argument). > Note this build hack isn't in my github repository. > >> mk/main.mk >> -EXTRAFLAGS_STAGE$(1) = $$(RUSTFLAGS_STAGE$(1)) >> +EXTRAFLAGS_STAGE$(1) = --sysroot $$(HROOT$(1)_H_$(3)) >> $$(RUSTFLAGS_STAGE$(1)) > > - under openbsd, it seems rustc couldn't use ports/llvm. > > ports/llvm annonce it-self as 3.5, but directory structure is > something between 3.4.2 et 3.5. and even if I hack the #include in > src/rustllvm, I have compilation errors... > > As result, I currently use the llvm version provided with rustc, and > so need to compile the embeded llvm version before compile rustc. > > It is ok for now, but when rustc would work enought to be packaged, > it should be investigated... > > - for coredump, there are double-free and coredump that occurs > generally in thread code. I think the openbsd-layer of rustc isn't > fully correct. > > - sometimes, panic randomly occurs (rustc detect itself a problem and > abort). Generally it is in checking of args for FFI ('\0' char in > middle of CString for example). So could be related to > openbsd-specific code in rustc. > > - sometimes, rustc enter in infinite-loop (but it stuck in a wait, not > a busy loop). It is generally in spawing... so it may be a race in > waiting the process to ending ? Seems related to spawning too... > > > In conclusion, I am able to build rustc until processing > librustc_llvm.rlib (librustc_llvm.so is ok). > > And I am currently blocking by thread-related coredump (or > infinite-waiting). > > For how I understand the build system of rustc, building > src/librustc_llvm need to heavy use of spawning, for construct the rlib > file (an object archive, if I understand well), as it take objects files > from LLVM libraries (so a lot of libraries). So I hit theses bugs each > time (but never at same place...) I try to build it. > > So I am in a linuxBuild-openbsdBuild-correctOpenbsdRustcSupport loop... > (and a full linuxBuild is something like 90min compile time). > > Review of the rustc openbsd specific code are welcome... or any others > ideas :) > > Thanks reading.
