Move first RISC-V builder over to cross-compiler + execute tests under qemu-system setup

[asb]

This PR is structured as two patches (at the time of writing), though it might make sense to separate out the first one for separate review. I'm posting together to get initial feedback.

My goal is to move clang-riscv-rva20-2stage, clang-riscv-rva23-mrvv-vec-bits-2stage, and clang-riscv-rva23-evl-vec-2stage over so:

• All building is done on the X86 host (first stage straight-forward native build, stage2 a crossbuild using the just-built stage1)
• To execute the tests (ninja check-all) we rely on a llvm-lit wrapper script that spins up a special-purpose VM with the source+build tree copied into it (see here) under qemu-system and runs tests there. This should be much faster than the current approach that does all building under qemu-system.

This PR does this change just for clang-riscv-rva20-2stage for ease of review and so we can get any teething problems sorted out. Although it might be possible to properly configure the cross-build otherwise, I've adding a mechanism for creating and adding options to a toolchains file used by the second stage. This matches CMake recommended practice for cross-building, and after spending a lot of time iterating on various cross-build configs recently I've found it's the option that gives most control and fewer problems (notably, some CMake variables can't be set outside of a toolchain file).

@gkistanova: The underlying worker is currently configured within a RISC-V qemu-system. Once this is ready to merge, I'll bring it over to running under the host meaning I shouldn't need to set up a new password and so on. But let me know if you'd prefer I do set up a new worker with a new name and retire this name.

We'll later want to add the ability to customise the enabled projects/runtimes vs stage1 vs stage2.

For reference, this approximates the logic that the configured builder should follow:

#!/bin/sh

die () {
  printf "%s\n" "$*" >&2
  exit 1
}

mkdir -p build && cd build
mkdir -p stage1 stage1.install stage2
printf "!!!!!!!!!! Building stage 1 !!!!!!!!!!\n"
cd stage1
# TODO: no need to build clang-tools-extra in stage1, but this matches logic in ClangBuilder.py
cmake -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DLLVM_ENABLE_ASSERTIONS=True \
  -DCMAKE_INSTALL_PREFIX=../stage1.install \
  -DLLVM_ENABLE_PROJECTS="clang;clang-tools-extra;lld" \
  -DCMAKE_C_COMPILER=clang \
  -DCMAKE_CXX_COMPILER=clang++ \
  -DLLVM_ENABLE_LLD=True \
  -DLLVM_TARGETS_TO_BUILD=RISCV \
  -DCMAKE_C_COMPILER_LAUNCHER=ccache \
  -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
  ../../llvm || die "Failed to configure stage 1 build"
ninja || die "Failed to build stage 1"
printf "!!!!!!!!!! Installing stage 1 !!!!!!!!!!\n"
ninja install || die "Failed to install stage 1"
printf "!!!!!!!!!! Building stage 2 !!!!!!!!!!\n"
cd ../stage2
cat - <<EOF > stage1-toolchain.cmake
set(CMAKE_SYSTEM_NAME Linux)

set(CMAKE_SYSROOT $HOME/rvsysroot)

set(CMAKE_C_COMPILER $(pwd)/../stage1.install/bin/clang)
set(CMAKE_CXX_COMPILER $(pwd)/../stage1.install/bin/clang++)

set(CMAKE_C_COMPILER_TARGET riscv64-linux-gnu)
set(CMAKE_CXX_COMPILER_TARGET riscv64-linux-gnu)
set(CMAKE_C_FLAGS_INIT '-march=rva20u64')
set(CMAKE_CXX_FLAGS_INIT '-march=rva20u64')

set(CMAKE_LINKER_TYPE LLD)

set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
EOF
cmake -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DLLVM_ENABLE_ASSERTIONS=True \
  -DLLVM_ENABLE_PROJECTS="clang;clang-tools-extra;lld" \
  -DCMAKE_TOOLCHAIN_FILE=$(pwd)/stage1-toolchain.cmake \
  -DLLVM_NATIVE_TOOL_DIR=$(pwd)/../stage1/bin \
  -DLLVM_BUILD_TESTS=True \
  -DLLVM_EXTERNAL_LIT=$HOME/lit-on-qemu-system-rva20.py \
  ../../llvm || die "Failed to configure stage 2 build"
ninja || die "Failed to build stage 2"
printf "!!!!!!!!!! Running check-all on stage 2 build !!!!!!!!!!\n"
ninja check-all || die "check-all failed"

[DavidSpickett]

The obvious alternative is binfmt_misc. We have used this at Linaro to run AArch64 compiler binaries on x86 to build benchmarks. Though we then copied those to real hardware of course, it just let us use spare machines for compiling. Once we had enough Arm hardware we did it all on AArch64.

Splitting the test build vs. test run apart for lit tests is not a viable option however. Benchmark build and run were already discreet steps so it was easy.

Even if you run the tests via qemu-user too, you could argue it's not a "real" system layout. You pass a sysroot but having a proper Linux install is going to be more accurate, even if only by a small amount. It's the "test what you ship" saying but "test where you're going to ship to". And it gives you a VM image you can send to someone if they want to dig into the problem.

There's probably a lot of overhead in ~10,000 qemu-users being set up and torn down as well, even compared to the kernel boot time.

Am I right in thinking that this qemu-system VM is per run?

That clears out any leftover files, which is good, disadvantage is that you can't get at reproducers as easily. Then again, when people want a reproducer from a Linaro bot they have to ask us for it, so only difference here is you would start a VM locally and reproduce it again.

I thought maybe you could cross compile in stage 1 right away, but any bug in the host compiler's risc-v codegen would break that build. So it makes sense to build latest clang as the compiler for stage 2. You have enabled ccache so this mitigates the overhead of stage 1.

We had some issues like this where SVE codegen in our host clang compiler for stage 1 would hit bugs that had already been fixed. Luckily there was a version we could upgrade to but otherwise you're trying to host nightly builds and it's a mess.

[DavidSpickett]

The obvious alternative is binfmt_misc.

Also we cannot test lldb with qemu-user because it doesn't emulate ptrace calls. You can use qemu's own GDB stub as the debug server but lldb -> lldb-server tests don't work.

To be clear, don't add lldb here. I am happy to see this though as interest in RISC-V lldb has been picking up lately and it will need checking eventually.

[asb]

The obvious alternative is binfmt_misc. We have used this at Linaro to run AArch64 compiler binaries on x86 to build benchmarks. Though we then copied those to real hardware of course, it just let us use spare machines for compiling. Once we had enough Arm hardware we did it all on AArch64.

binfmt_misc and qemu-user does work pretty well, and is slightly faster but requires masking out some tests that are incompatible with the approach. My feeling is that for something I hope to move to the non-staging buildmaster, using qemu-system in this way and thus minimising the chance of test failures being an artifact of the emulation process and its interaction with the build system is probably more viable. As you say below, LLDB tests can't run under qemu-user. I'd be a lot happier running libc tests under qemu-system as well rather than using it as a stress test for qemu-user syscall translation! So my reasoning is this approach can work well as a baseline that doesn't require alternate approaches for different sub-projects.

Splitting the test build vs. test run apart for lit tests is not a viable option however. Benchmark build and run were already discreet steps so it was easy.

Even if you run the tests via qemu-user too, you could argue it's not a "real" system layout. You pass a sysroot but having a proper Linux install is going to be more accurate, even if only by a small amount. It's the "test what you ship" saying but "test where you're going to ship to". And it gives you a VM image you can send to someone if they want to dig into the problem.

There's probably a lot of overhead in ~10,000 qemu-users being set up and torn down as well, even compared to the kernel boot time.

qemu-user for separate process typically scales better than running that same set of processes withing qemu-system (sorry I don't have numbers to hand right now). But for the reasons above, and due to concern about qemu-user needing more work in terms of masking specific tasks and the like, I prefer qemu-system.

Am I right in thinking that this qemu-system VM is per run?

Yes, effectively launched as a single-use appliance.

That clears out any leftover files, which is good, disadvantage is that you can't get at reproducers as easily. Then again, when people want a reproducer from a Linaro bot they have to ask us for it, so only difference here is you would start a VM locally and reproduce it again.

Plus a documented cross-compile flow means it's a bit easier to reproduce! I can look at making it easier to reproduce an equivalent VM image, though this will be easier once the next Debian release happens so there's a stable base rather than just using sid at the time the image was generated.

I thought maybe you could cross compile in stage 1 right away, but any bug in the host compiler's risc-v codegen would break that build. So it makes sense to build latest clang as the compiler for stage 2. You have enabled ccache so this mitigates the overhead of stage 1.

My thinking exactly. Especially with the different rva23 vectorisation codegen options, we're using building+testing clang as a bit of a stress test for codegen itself. Plus old enough clang/llvm didn't recognise all the options we're using.

[asb]

FYI I've taken a slight diversion to set up a good flow for testing builder config changes locally, in order to easier land this and other related changes: #289

[asb]

With a proper testing flow in place, I've fixed the issues related to the toolchains file generation as well as bugs in my initial implementation related to buildbot interpolation.

I considered the idea of having the toolchains file as a file in the llvm-zorg repo, but I don't think it would be overly useful (IMHO):

• We need buildbot to inject paths into it, so it needs interpolation
• Configuration for the build is then split between ClangBuilder.py, builders.py, and this toolchain file, so arguably more complex than before.

[DavidSpickett]

As long as it's possible to reproduce without starting a whole master. If we're lucky this download string step records it as an artifact or shows it.

[DavidSpickett]

This script is I think the only thing that bothers me. Perhaps it can live in https://github.com/llvm/llvm-zorg/tree/main/zorg/buildbot/builders and be downloaded as the toolchain file now is?

Though then you need a way to tell the clang builder to download it. Putting it in llvm-project would solve that but as per usual it's finding the appropriate place for it.