Convert the solution to slnx - #14
Open
henrikottesorensen wants to merge 6 commits into
Open
Conversation
Adds build/verify_native_binary.sh, called from pack_runtime_package, so a binary that fails inspection never becomes a package. Checks per RID: - architecture matches the RID. A mis-targeted binary produces a package that restores fine and never loads. - every P/Invoke symbol is exported. The expected list is read from the EntryPoint attributes in NativeMethod.cs rather than kept in the script, so it cannot drift away from what the wrapper imports. - no dependency outside a per-platform allowlist. This is the check that catches the libgcc_s_dw2-1.dll class of bug, where a binary links against a toolchain runtime that the package does not ship: it works on the machine that built it and fails everywhere else. - Linux only, the highest required glibc symbol version stays within a declared floor. The floor decides which distributions can consume the packages and is a property of the build image, so it can rise silently when that image is bumped. Currently 2.34, which covers RHEL 9, Debian 12 and Ubuntu 22.04. linux-x86 sits exactly on it. Verified against all eight RIDs, and against two deliberately bad inputs: an x86-64 binary declared as linux-arm64, and a win-x86 built without -static-libgcc, which is the bug this repository actually shipped. Both are rejected. Note that bug only reproduces with the mingw gcc 10 on jammy; the gcc 13 on noble does not emit the dependency at all. -static-libgcc stays so the output does not depend on which compiler the base image happens to ship. llvm is added to the build image because binutils cannot read aarch64 PE. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two problems, both found by running the script from outside the repository while verifying the 3.38.0 bump. The symbol check passed against an empty list. expected_symbols read NativeMethod.cs and was called through command substitution, so when the file could not be read it printed an error, exited its own subshell, and left the caller to compare against nothing. The output read "all 0 P/Invoke symbols exported", which looks like a pass and asserts nothing. The list is now resolved once in the main shell, where a missing or unparsable NativeMethod.cs stops the script. find_tool only looked where Linux distributions put things, so on macOS the ELF checks could not run at all: command -v nm finds BSD nm, which has no -D. It now takes several interchangeable names in preference order and also searches Homebrew's keg-only prefixes, so llvm-readelf, llvm-nm and llvm-readobj are used when present. Those read ELF, PE and Mach-O alike, so with brew install llvm all eight RIDs can be verified on a Mac without a container. Verified: all eight RIDs pass natively on macOS, and a script run where NativeMethod.cs is unreachable now exits 1 instead of reporting success. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
CI failed fetching linux-libc-dev, a dependency of build-essential: E: Failed to fetch .../linux-libc-dev_5.15.0-187.197_amd64.deb 404 archive.ubuntu.com does not update its index and its pool atomically, so a package version can still be listed after it has been removed, and the fetch 404s. Nothing to do with the packages this image asks for; it is luck, and a single apt-get run has none to spare. Both apt steps now try three times, refreshing the index each time, since a newer index is usually what resolves it. The explicit ok check is load bearing: a bare loop that never succeeds still falls through, and the layer would build with nothing installed and fail much later with something unrecognisable. Verified with --no-cache, which is the case that actually hits the network. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
LibLouis.NET and the test project build for both. The extensions packages are referenced per target framework, 8.0.2 for net8.0 and 10.0.10 for net10.0, so a net10.0 consumer does not drag an older set into its dependency graph. The runtime packages stay on netstandard2.0: they carry no managed code, and netstandard2.0 is the widest thing to be compatible with. Microsoft.NET.Test.Sdk moves to 18.8.1, the first version that supports net10.0. Packing a multi-targeted project runs the cross-targeting outer build, which has no default None items, so the existing <None Update="LICENSE"> matched nothing there and the licence silently never reached the package (NU5030). The inner builds do have the item, where a second Include would be a duplicate, so both forms are present and conditioned on whether TargetFramework is set. The container's own SDK version is deliberately not touched here. It compiles C and runs dotnet pack, and which SDK does the packing barely affects the output; moving it belongs with the Dockerfile restructure that separates compiling from packing, not with the multi-targeting. Verified: both target frameworks build and pass, and the package carries lib/net8.0 and lib/net10.0 with matching per-TFM dependency groups. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
dotnet sln migrate, plus the two things that fall out of it. runtime.liblouis.crosscompile.sln is deleted rather than converted. Nothing loaded it: the build scripts pack individual csproj files, and it had not been updated to include the macOS runtime projects, so it had been quietly wrong for some time. Converting a file nobody opens would only preserve that. global.json declares an SDK floor of 9.0.200, the first release that reads slnx. This is mostly documentation: on a machine with both SDKs dotnet already picks the newer one, so nothing changes in practice. It does not buy a good error either - an unsatisfiable global.json reports that the command could not be loaded, which is no clearer than the MSB4068 an old SDK gives on an slnx. The value is that the requirement is written down where someone would look for it. Stacks on the multi-targeting change because that is what puts a .NET 10 SDK in CI. On main the workflow installs 8.0.x only, which cannot read slnx at all. Verified: build_managed_packages.sh builds, tests and packs through the slnx, on both target frameworks. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
It broke the container build: A compatible .NET SDK was not found. process "/bin/sh -c sh ./build/build_metapackage.sh" exited 145 global.json applies to every dotnet invocation in the repository, and the container runs sdk:8.0-jammy. I had reasoned about which SDK CI installs and missed that the container has its own. Nothing in the container needs a newer SDK. It packs individual project files, which any SDK handles, and never loads the solution. Only build_managed_packages.sh does, and that runs on a runner where setup-dotnet provides .NET 10. The floor is documented instead. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacks on #7
The first commit is that PR. Review only
build: Convert the solution to slnx.Merge #7 first and this reduces to a single commit.
dotnet sln migrate, plus the two things that fall out of it. 34 lines of GUIDs and per-configuration platform mappings become five lines:It stacks on #7 because of a hard SDK requirement
.slnxneeds an SDK of 9.0.200 or later.main's workflow installs8.0.xonly, and SDK 8 cannot read the format at all — it has nomigrateverb and fails withMSB4068: The element <Solution> is unrecognized. #7 is what puts a .NET 10 SDK in CI, so this has to follow it.global.jsondeclares the floor. Being honest about what that buys: on a machine with both SDKs installed,dotnetalready picks the newer one, so nothing changes in practice; and an unsatisfiableglobal.jsonreports "The command could not be loaded", which is no clearer thanMSB4068. The value is that the requirement is written down where someone would look for it, not that the failure becomes obvious.runtime.liblouis.crosscompile.sln is deleted, not converted
Nothing loads it. The build scripts pack individual
.csprojfiles, and grepping the repository finds no reference to it. It had also never been updated to include the two macOS runtime projects, so it had been quietly wrong for a while. Converting a file nobody opens would only have preserved that.The only consumer of a solution file is
build/build_managed_packages.sh, updated here.Verified
build_managed_packages.shbuilds, tests and packs through the.slnx, on bothnet8.0andnet10.0.global.jsonresolves to SDK 10.0.302.Note for reviewers on older tooling:
.slnxneeds Visual Studio 17.14+, Rider 2025.1+, or the .NET 9.0.200+ CLI.🤖 Generated with Claude Code