Setup dependencies that match the GCC ABI of the compiler (#604)

* Add `shards` as keyword argument to the runners

* Add functions to query preferred ABIs for the given compiler shard

* Setup dependencies that match the GCC ABI of the compiler

* Move here definition of Pkg.BinaryPlatforms.gcc_versions

* Define custom structs for compilers builds

* Simplify logic of selection of GCC builds

* Changes suggested by Elliot

* Make sure the compiler wrappers have the requested C++ string ABI

Author: giordano

src/AutoBuild.jl

@@ -629,25 +629,41 @@ function autobuild(dir::AbstractString,
         build_path = joinpath(dir, "build", triplet(platform))
         mkpath(build_path)
 
+        shards = choose_shards(platform; extract_kwargs(kwargs, (:preferred_gcc_version,:preferred_llvm_version,:bootstrap_list,:compilers))...)
+        # We want to get dependencies that have exactly the same GCC ABI as the
+        # choosen compiler, otherwise we risk, e.g., to build in an environment
+        # with libgfortran3 a dependency built with libgfortran5.
+        # `concrete_platform` is needed only to setup the dependencies and the
+        # runner.  We _don't_ want the platform passed to `audit()` or
+        # `package()` to be more specific than it is.
+        concrete_platform = platform
+        gccboostrap_shard_idx = findfirst(x -> x.name == "GCCBootstrap" && x.target.arch == platform.arch && x.target.libc == platform.libc, shards)
+        if !isnothing(gccboostrap_shard_idx)
+            libgfortran_version = preferred_libgfortran_version(platform, shards[gccboostrap_shard_idx])
+            cxxstring_abi = preferred_cxxstring_abi(platform, shards[gccboostrap_shard_idx])
+            concrete_platform = replace_cxxstring_abi(replace_libgfortran_version(platform, libgfortran_version), cxxstring_abi)
+        end
+
         prefix = setup_workspace(
             build_path,
             src_paths,
             src_hashes;
             verbose=verbose,
         )
-        artifact_paths = setup_dependencies(prefix, dependencies, platform)
+        artifact_paths = setup_dependencies(prefix, dependencies, concrete_platform)
 
         # Create a runner to work inside this workspace with the nonce built-in
         ur = preferred_runner()(
             prefix.path;
             cwd = "/workspace/srcdir",
-            platform = platform,
+            platform = concrete_platform,
             verbose = verbose,
             workspaces = [
                 joinpath(prefix, "metadir") => "/meta",
             ],
             compiler_wrapper_dir = joinpath(prefix, "compiler_wrappers"),
             src_name = src_name,
+            shards = shards,
             extract_kwargs(kwargs, (:preferred_gcc_version,:preferred_llvm_version,:compilers,:allow_unsafe_flags))...,
         )
 

src/DockerRunner.jl

@@ -72,6 +72,7 @@ function DockerRunner(workspace_root::String;
                       verbose::Bool = false,
                       compiler_wrapper_path::String = mktempdir(),
                       src_name::AbstractString = "",
+                      shards = nothing,
                       kwargs...)
     global use_ccache
 
@@ -97,8 +98,10 @@ function DockerRunner(workspace_root::String;
         push!(workspaces, "binarybuilder_ccache" => "/root/.ccache")
     end
 
-    # Choose the shards we're going to mount
-    shards = choose_shards(platform; extract_kwargs(kwargs, (:preferred_gcc_version,:preferred_llvm_version,:bootstrap_list,:compilers))...)
+    if isnothing(shards)
+        # Choose the shards we're going to mount
+        shards = choose_shards(platform; extract_kwargs(kwargs, (:preferred_gcc_version,:preferred_llvm_version,:bootstrap_list,:compilers))...)
+    end
 
     # Import docker image
     import_docker_image(shards[1], workspace_root; verbose=verbose)

src/Rootfs.jl

@@ -319,15 +319,78 @@ function select_closest_version(preferred::VersionNumber, versions::Vector{Versi
     return versions[closest_idx]
 end
 
-const available_gcc_builds = [v"4.8.5", v"5.2.0", v"6.1.0", v"7.1.0", v"8.1.0", v"9.1.0"]
-const available_llvm_builds = [v"6.0.1", v"7.1.0", v"8.0.1", v"9.0.1"]
+abstract type CompilerBuild end
+
+struct GCCBuild <: CompilerBuild
+    version::VersionNumber
+    abi::CompilerABI
+end
+GCCBuild(v::VersionNumber) = GCCBuild(v, CompilerABI())
+
+struct LLVMBuild <: CompilerBuild
+    version::VersionNumber
+    abi::CompilerABI
+end
+LLVMBuild(v::VersionNumber) = LLVMBuild(v, CompilerABI())
+
+getversion(c::CompilerBuild) = c.version
+getabi(c::CompilerBuild) = c.abi
+
+const available_gcc_builds = [GCCBuild(v"4.8.5", CompilerABI(libgfortran_version = v"3", libstdcxx_version = v"3.4.19", cxxstring_abi = :cxx03)),
+                              GCCBuild(v"5.2.0", CompilerABI(libgfortran_version = v"3", libstdcxx_version = v"3.4.21", cxxstring_abi = :cxx11)),
+                              GCCBuild(v"6.1.0", CompilerABI(libgfortran_version = v"3", libstdcxx_version = v"3.4.22", cxxstring_abi = :cxx11)),
+                              GCCBuild(v"7.1.0", CompilerABI(libgfortran_version = v"4", libstdcxx_version = v"3.4.23", cxxstring_abi = :cxx11)),
+                              GCCBuild(v"8.1.0", CompilerABI(libgfortran_version = v"5", libstdcxx_version = v"3.4.25", cxxstring_abi = :cxx11)),
+                              GCCBuild(v"9.1.0", CompilerABI(libgfortran_version = v"5", libstdcxx_version = v"3.4.26", cxxstring_abi = :cxx11))]
+const available_llvm_builds = [LLVMBuild(v"6.0.1"),
+                               LLVMBuild(v"7.1.0"),
+                               LLVMBuild(v"8.0.1"),
+                               LLVMBuild(v"9.0.1")]
+
+"""
+    gcc_version(cabi::CompilerABI, GCC_builds::Vector{GCCBuild})
+
+Returns the closest matching GCC version number for the given CompilerABI
+representing a particular platofrm, from the given set of options.  If no match
+is found, returns an empty list.  This method assumes that `cabi` represents a
+platform that binaries will be run on, and thus versions are always rounded
+down; e.g. if the platform supports a `libstdc++` version that corresponds to
+`GCC 5.1.0`, but the only GCC versions available to be picked from are `4.8.5`
+and `5.2.0`, it will return `4.8.5`, as binaries compiled with that version
+will run on this platform, whereas binaries compiled with `5.2.0` may not.
+"""
+function gcc_version(cabi::CompilerABI, GCC_builds::Vector{GCCBuild})
+    # First, filter by libgfortran version.
+    if cabi.libgfortran_version !== nothing
+        GCC_builds = filter(b -> getabi(b).libgfortran_version == cabi.libgfortran_version, GCC_builds)
+    end
+
+    # Next, filter by libstdc++ GLIBCXX symbol version.  Note that this
+    # mapping is conservative; it is often the case that we return a
+    # version that is slightly lower than what is actually installed on
+    # a system.  See https://gcc.gnu.org/onlinedocs/libstdc++/manual/abi.html
+    # for the whole list, as well as many other interesting factoids.
+    if cabi.libstdcxx_version !== nothing
+        GCC_builds = filter(b -> getabi(b).libstdcxx_version <= cabi.libstdcxx_version, GCC_builds)
+    end
+
+    # Finally, enforce cxxstring_abi guidelines.  It is possible to build
+    # :cxx03 binaries on GCC 5+, (although increasingly rare) so the only
+    # filtering we do is that if the platform is explicitly :cxx11, we
+    # disallow running on < GCC 5.
+    if cabi.cxxstring_abi !== nothing && cabi.cxxstring_abi === :cxx11
+        GCC_builds = filter(b -> getversion(b) >= v"5", GCC_builds)
+    end
+
+    return getversion.(GCC_builds)
+end
 
 function select_gcc_version(p::Platform,
-             GCC_builds::Vector{VersionNumber} = available_gcc_builds,
-             preferred_gcc_version::VersionNumber = GCC_builds[1],
+             GCC_builds::Vector{GCCBuild} = available_gcc_builds,
+             preferred_gcc_version::VersionNumber = getversion(GCC_builds[1]),
          )
     # Determine which GCC build we're going to match with this CompilerABI:
-    GCC_builds = Pkg.BinaryPlatforms.gcc_version(compiler_abi(p), GCC_builds)
+    GCC_builds = gcc_version(compiler_abi(p), GCC_builds)
 
     if isempty(GCC_builds)
         error("Impossible CompilerABI constraints $(cabi)!")
@@ -350,19 +413,22 @@ function choose_shards(p::Platform;
             compilers::Vector{Symbol} = [:c],
             rootfs_build::VersionNumber=v"2020.01.07",
             ps_build::VersionNumber=v"2020.01.15",
-            GCC_builds::Vector{VersionNumber}=available_gcc_builds,
-            LLVM_builds::Vector{VersionNumber}=available_llvm_builds,
+            GCC_builds::Vector{GCCBuild}=available_gcc_builds,
+            LLVM_builds::Vector{LLVMBuild}=available_llvm_builds,
             Rust_build::VersionNumber=v"1.18.3",
             Go_build::VersionNumber=v"1.13",
             archive_type::Symbol = (use_squashfs ? :squashfs : :unpacked),
             bootstrap_list::Vector{Symbol} = bootstrap_list,
-            # We prefer the oldest GCC version by default
-            preferred_gcc_version::VersionNumber = GCC_builds[1],
-            preferred_llvm_version::VersionNumber = LLVM_builds[end],
+            # Because GCC has lots of compatibility issues, we always default to
+            # the earliest version possible.
+            preferred_gcc_version::VersionNumber = getversion(GCC_builds[1]),
+            # Because LLVM doesn't have compatibilty issues, we always default
+            # to the newest version possible.
+            preferred_llvm_version::VersionNumber = getversion(LLVM_builds[end]),
         )
 
     GCC_build = select_gcc_version(p, GCC_builds, preferred_gcc_version)
-    LLVM_build = select_closest_version(preferred_llvm_version, LLVM_builds)
+    LLVM_build = select_closest_version(preferred_llvm_version, getversion.(LLVM_builds))
     # Our host platform is x86_64-linux-musl
     host_platform = Linux(:x86_64; libc=:musl)
 
@@ -531,6 +597,58 @@ function expand_cxxstring_abis(ps::Vector{<:Platform})
     return Platform[p for p in Iterators.flatten(expand_cxxstring_abis.(ps))]
 end
 
+"""
+    preferred_libgfortran_version(platform::Platform, shard::CompilerShard)
+
+Return the libgfortran version preferred by the given platform or GCCBootstrap shard.
+"""
+function preferred_libgfortran_version(platform::Platform, shard::CompilerShard)
+    # Some input validation
+    if shard.name != "GCCBootstrap"
+        error("Shard must be `GCCBootstrap`")
+    end
+    if shard.target.arch != platform.arch || shard.target.libc != platform.libc
+        error("Incompatible platform and shard target")
+    end
+
+    if compiler_abi(platform).libgfortran_version != nothing
+        # Here we can't use `shard.target` because the shard always has the
+        # target as ABI-agnostic, thus we have also to ask for the platform.
+        return compiler_abi(platform).libgfortran_version
+    elseif shard.version < v"7"
+        return v"3"
+    elseif v"7" <= shard.version < v"8"
+        return v"4"
+    else
+        return v"5"
+    end
+end
+
+"""
+    preferred_cxxstring_abi(platform::Platform, shard::CompilerShard)
+
+Return the C++ string ABI preferred by the given platform or GCCBootstrap shard.
+"""
+function preferred_cxxstring_abi(platform::Platform, shard::CompilerShard)
+    # Some input validation
+    if shard.name != "GCCBootstrap"
+        error("Shard must be `GCCBootstrap`")
+    end
+    if shard.target.arch != platform.arch || shard.target.libc != platform.libc
+        error("Incompatible platform and shard target")
+    end
+
+    if compiler_abi(platform).cxxstring_abi != nothing
+        # Here we can't use `shard.target` because the shard always has the
+        # target as ABI-agnostic, thus we have also to ask for the platform.
+        return compiler_abi(platform).cxxstring_abi
+    elseif shard.version < v"5"
+        return :cxx03
+    else
+        return :cxx11
+    end
+end
+
 """
     download_all_shards(; verbose::Bool=false)
 

src/Runner.jl

@@ -33,7 +33,11 @@ exeext(p::Union{Linux,FreeBSD,MacOS}) = ""
 exeext(p::Platform) = error("Unknown exeext for platform $(p)")
 
 """
-    generate_compiler_wrappers(p::Platform, bin_path::AbstractString)
+    generate_compiler_wrappers!(platform::Platform; bin_path::AbstractString,
+                                host_platform::Platform = Linux(:x86_64; libc=:musl),
+                                rust_platform::Platform = Linux(:x86_64; libc=:glibc),
+                                compilers::Vector{Symbol} = [:c],
+                                allow_unsafe_flags::Bool = false)
 
 We generate a set of compiler wrapper scripts within our build environment to force all
 build systems to honor the necessary sets of compiler flags to build for our systems.
@@ -61,6 +65,32 @@ function generate_compiler_wrappers!(platform::Platform; bin_path::AbstractStrin
     host_target = aatriplet(host_platform)
     rust_target = aatriplet(rust_platform)
 
+
+    # If the ABI-agnostic triplets of the target and the host platform are the
+    # same, then we have to be very careful: we can't have distinct wrappers, so
+    # we have to be sure that their ABIs are consistent, in particular that
+    # we're correctly writing the wrappers for the target platform.  In
+    # particular, what we care about with regard to the wrappers is the C++
+    # string ABI.  We have the following situations:
+    #   * they are equal: this is fine
+    #   * they're different and the host has a prefernce for the C++ string ABI:
+    #     we can't deal with this situation as the host wrappers will be always
+    #     overwritten, then error out
+    #   * in all other cases we don't do anything here but later below we'll let
+    #     the host wrappers to be overwritten by the wrappers for the target
+    if target == host_target
+        target_cxxabi = compiler_abi(platform).cxxstring_abi
+        host_cxxabi   = compiler_abi(host_platform).cxxstring_abi
+        if target_cxxabi !== host_cxxabi
+            if host_cxxabi !== nothing
+                # This is a very unlikely situation as ideally the host
+                # shouldn't have particular preferences for the ABI, thus in
+                # practice we shouldn't never reach this.
+                error("Incompatible C++ string ABIs between the host and target platforms")
+            end
+        end
+    end
+
     # If we should use ccache, prepend this to every compiler invocation
     ccache = use_ccache ? "ccache" : ""
 
@@ -73,7 +103,7 @@ function generate_compiler_wrappers!(platform::Platform; bin_path::AbstractStrin
                      unsafe_flags = String[])
         write(io, """
         #!/bin/bash
-        # This compiler wrapper script brought into existence by `generate_compiler_wrappers()`
+        # This compiler wrapper script brought into existence by `generate_compiler_wrappers!()`
 
         if [ "x\${SUPER_VERBOSE}" = "x" ]; then
             vrun() { "\$@"; }
@@ -308,8 +338,11 @@ function generate_compiler_wrappers!(platform::Platform; bin_path::AbstractStrin
         chmod(joinpath(bin_path, fname), 0o775)
     end
 
-    ## Generate compiler wrappers for both our target and our host
-    for p in unique((platform, host_platform))
+    ## Generate compiler wrappers for both our host and our target.  In
+    ## particular, we write the wrapper for the target after those for the host,
+    ## to override host-specific ABI in case this is incompatible with that of
+    ## the target
+    for p in unique((host_platform, platform))
         t = aatriplet(p)
 
         # Generate `:c` compilers

src/UserNSRunner.jl

@@ -27,6 +27,7 @@ function UserNSRunner(workspace_root::String;
                       verbose::Bool = false,
                       compiler_wrapper_path::String = mktempdir(),
                       src_name::AbstractString = "",
+                      shards = nothing,
                       kwargs...)
     global use_ccache, use_squashfs, runner_override
 
@@ -55,8 +56,10 @@ function UserNSRunner(workspace_root::String;
         push!(workspaces, ccache_dir() => "/root/.ccache")
     end
 
-    # Choose the shards we're going to mount
-    shards = choose_shards(platform; extract_kwargs(kwargs, (:preferred_gcc_version,:preferred_llvm_version,:bootstrap_list,:compilers))...)
+    if isnothing(shards)
+        # Choose the shards we're going to mount
+        shards = choose_shards(platform; extract_kwargs(kwargs, (:preferred_gcc_version,:preferred_llvm_version,:bootstrap_list,:compilers))...)
+    end
 
     # Construct sandbox command to look at the location it'll be mounted under
     mpath = mount_path(shards[1], workspace_root)

src/wizard/deploy.jl

@@ -43,11 +43,11 @@ function print_build_tarballs(io::IO, state::WizardState)
         push!(kwargs_vec, "compilers = [$(join(map(x -> ":$(x)", state.compilers), ", "))]")
     end
     # Default GCC version is the oldest one
-    if state.preferred_gcc_version != available_gcc_builds[1]
+    if state.preferred_gcc_version != getversion(available_gcc_builds[1])
         push!(kwargs_vec, "preferred_gcc_version = v\"$(state.preferred_gcc_version)\"")
     end
     # Default LLVM version is the latest one
-    if state.preferred_llvm_version != available_llvm_builds[end]
+    if state.preferred_llvm_version != getversion(available_llvm_builds[end])
         push!(kwargs_vec, "preferred_llvm_version = v\"$(state.preferred_llvm_version)\"")
     end
     kwargs = ""

src/wizard/obtain_source.jl

@@ -404,13 +404,13 @@ function step2(state::WizardState)
     get_name_and_version(state)
     if yn_prompt(state, "Do you want to customize the set of compilers?", :n) == :y
         get_compilers(state)
-        get_preferred_version(state, "GCC", available_gcc_builds)
-        get_preferred_version(state, "LLVM", available_llvm_builds)
+        get_preferred_version(state, "GCC", getversion.(available_gcc_builds))
+        get_preferred_version(state, "LLVM", getversion.(available_llvm_builds))
     else
         state.compilers = [:c]
         # Default GCC version is the oldest one
-        state.preferred_gcc_version = available_gcc_builds[1]
+        state.preferred_gcc_version = getversion(available_gcc_builds[1])
         # Default LLVM version is the latest one
-        state.preferred_llvm_version = available_llvm_builds[end]
+        state.preferred_llvm_version = getversion(available_llvm_builds[end])
     end
 end