Build on PPC64LE with installed sysroot 2.28 fails #20335
Comments
MementoRC
added
the
bug
|
Update: Patching signal.h to comment out There seems to be an collision with the definitions of the registers The CMake cross-compilation failed due to a silly mistake, the zig-wasm2 used in However, even though the emulation works the compilation sets the HOST_TARGET to linux-x64 |
|
Progressing on the cross-build till: I had to patch the zig2.c with args being in the clobber list Somehow the compiler_rt.c has zig_f64 defined as a double instead of a long double and have several read 16bit from 8bits in memcpy() |
|
The following instruction in building zig2: |
|
This is solved with |
|
This is the PR trying to cross-cmake x86_64 -> ppc64le: https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=965648&view=logs&j=8c139a40-380b-5518-bf6b-edcb18cbf319&t=e71f8d0d-13be-528e-b5e7-bfc20e0db699&l=1001 It fails to make |
|
Does it fail? From the log it looks like it timed out at |
|
It does. The relocation errors is quite large so it is redirected to a file (not sure why it did not fail on the runner, it fails locally) I don't understand the relocation errors (currently stuck there): |
|
Perhaps a question that can be answered by an LLD developer. Does LLD perhaps lack implementation of some relocations for ppc64le? Or perhaps clang++ is emitting some strange relocations for that target, for some reason. This target builds fine statically with zig-bootstrap, so maybe LLD is lacking a feature related to dynamic linking, or clang++ is lacking a feature related to dynamic linking. |
|
So with zig-bootstrap (the 0.13.0 zig linux-x64 cross-targeting ppc64le): |
|
The 31 and 60 relocs are valid, but they seem to not be implemented in LLD https://refspecs.linuxfoundation.org/ELF/ppc64/PPC-elf64abi.html#RELOC https://github.com/llvm/llvm-project/blob/main/lld/ELF/Relocations.h#L102-L106 |
|
The target |
|
Mold does implement these relocs, maybe you could workaround by using it? |
|
Zig does not integrate with mold since the plan is to provide our own linking functionality for this use case. |
|
Very cool, I kept banging my head on a brick wall it seems. I don't follow very well, but are we saying that I can only create a statically built zig for ppc64le? |
|
Yes, that is the current state of affairs, until Zig's linker surpasses LLD's functionality and adds these relocations. |
|
Cool, I will try. Based upon your previous comment, I should focus on a CMake build, right. I do not have ppc64le, so I used a patched CMakeLists.txt with an emulator, do you see any issues with that as opposed to a native cmake build on Travis runner? They are quite a pain to use while debugging the recipe I will open a different issue with the Windows build that is also not working at my end (suspecting something I don't understand again!) |
|
For ppc64le specifically, I think your best bet is using https://github.com/ziglang/zig-bootstrap to cross-compile. Be sure to read the README and use additional environment variables to utilize the logical cores of the host. This build process will take a long time because it will
However, it is definitely working, because it is how the CI builds are produced. |
|
Right, we would prefer to use the tools already available for host and target within the conda environment. |
|
When working on the I don't understand, I provide the sysroot and the libc file, and the target (I tried adding the |
|
you passed |
|
I recently gained access to the GCC compiler farm and I wanted test compatibility on a Power10 machine. Unfortunately I encountered many build errors that have limited me to using the 0.13 release via First, I tried building LLVM+Clang+LLD+Zig from source. The llvm portion went fine, but like @MementoRC mentioned at the top I kept running into conflicting type errors - more than sigaction. BTW, I also experienced the same conflicts on a Loongarch64 server which shows that it isnt ppc being special. I then tried using zig-bootstrap, but ran into the same errors. Since I couldn't edit the system headers, I tried using the latest master build. Running Then I spent a lot of time doing bootstrap builds on an x86_64 server, but got to the same results:
So, to build anything I have to run |
|
Chatting with @alexrp led me to the same conclusion re: signals, but instead of patching out diff --git a/stage1/zig.h b/stage1/zig.h
index 05529feec7..38f01ca020 100644
--- a/stage1/zig.h
+++ b/stage1/zig.h
@@ -271,6 +271,8 @@ typedef char bool;
#define zig_breakpoint() __asm__ volatile("bkpt #0");
#elif defined(__aarch64__)
#define zig_breakpoint() __asm__ volatile("brk #0");
+#elif defined(__powerpc__)
+#define zig_breakpoint() __asm__ volatile("twge r2,r2");
#else
#include <signal.h>
#if defined(SIGTRAP)With this, I get the same asm clobber errors mentioned above: @MementoRC can you share the changes you made to the clobber lists? I have no idea how inline asm works. |
|
Sure, you can find the patches under: https://github.com/conda-forge/zig-feedstock/pull/30/files#diff-6c2d6aaa2d4b5f7aafb076ae683ca83a14d62b89abc619d56b3f586e60335c79 The last issue I encountered was with the relocation that are not supported by LLVM, but I need to build dynamically for the conda-forge recipe, this may not be an issue for a static build |
|
Just leaving a link to my comment over here: conda-forge/zig-feedstock#44 (comment) TL;DR: LLD doesn't support some relocations that GCC/binutils use in practice. Self-hosted linker will hopefully solve this. |
Zig Version
0.13.0
Steps to Reproduce and Observed Behavior
cmake buildfails with:[17/19] Building C object CMakeFiles/zig2.dir/zig2.c.oThis is under a conda environment that installs sysroot for GLIBC 2.28 (which seems needed for zig). Is there a way to patch the source so that the built zig2.c complies with the
/powerpc64le-conda-linux-gnu/sysroot/usr/include/signal.h?Or an alternative method to build zig for ppc64le?
Using the distributed binary fails on the Travis runner without any error (objdump -h works but nm fails)
Cross-compiling with CMake fails because
qemu-ppc64le-staticfails similarly when trying to execute/usr/bin/qemu-ppc64le-static -s 32097152 -d guest_errors -p 65536 zig-wasm2cCross-compiling with x86_64 binary distro also fails (I'd need to dig the loogs). But basically, I seem unable to build zig for PPC64LE
This is part of an effort to bring zig to the conda framework. Currently, x86_64, aarch64 and osx-64 builds with the recipe
Thank you
Expected Behavior
completes the CMake build and generates a functional zig (the recipe tests the behavior.zig in an isolated and emulated environment
The text was updated successfully, but these errors were encountered: