Implement support for object caching through pkgimages (#47184)

This implements caching native code in "package images" (pkgimages).
We now write two serialization files, one ending in `*.ji` and the other with
the platform dynamic library extension (e.g., `*.so`). The `*.ji` contains
"extended" header information (include the source-code dump for Revise),
whereas the dynamic library includes the Julia objects, including LLVM-generated
native code.

Native code is compiled once during precompilation and then a second time
to build a "clean" module. When we find an edge to an external function (already
cached in anloaded pkgimage), we emit a global variable which we will patch during
loading with the address of the function to call. This allows us to leverage the
standard multiversioning capabilities.

Co-authored-by: Tim Holy <tim.holy@gmail.com>
Co-authored-by: Kristoffer Carlsson <kristoffer.carlsson@chalmers.se>
Co-authored-by: Mosè Giordano <giordano@users.noreply.github.com>
Co-authored-by: Ian Butterworth <i.r.butterworth@gmail.com>
Co-authored-by: Max Horn <max@quendi.de>
Co-authored-by: Michael Schlottke-Lakemper <michael@sloede.com>
Co-authored-by: Alex Ames <alexander.m.ames@gmail.com>

Makefile

@@ -262,13 +262,21 @@ ifeq ($(OS),WINNT)
 	-$(INSTALL_M) $(wildcard $(build_bindir)/*.dll) $(DESTDIR)$(bindir)/
 ifeq ($(JULIA_BUILD_MODE),release)
 	-$(INSTALL_M) $(build_libdir)/libjulia.dll.a $(DESTDIR)$(libdir)/
+	-$(INSTALL_M) $(build_libdir)/libjulia-internal.dll.a $(DESTDIR)$(libdir)/
 else ifeq ($(JULIA_BUILD_MODE),debug)
 	-$(INSTALL_M) $(build_libdir)/libjulia-debug.dll.a $(DESTDIR)$(libdir)/
+	-$(INSTALL_M) $(build_libdir)/libjulia-internal-debug.dll.a $(DESTDIR)$(libdir)/
 endif
 
 	# We have a single exception; we want 7z.dll to live in libexec, not bin, so that 7z.exe can find it.
 	-mv $(DESTDIR)$(bindir)/7z.dll $(DESTDIR)$(libexecdir)/
 	-$(INSTALL_M) $(build_bindir)/libopenlibm.dll.a $(DESTDIR)$(libdir)/
+	-$(INSTALL_M) $(build_libdir)/libssp.dll.a $(DESTDIR)$(libdir)/
+	# The rest are compiler dependencies, as an example memcpy is exported by msvcrt
+	# These are files from mingw32 and required for creating shared libraries like our caches.
+	-$(INSTALL_M) $(build_libdir)/libgcc_s.a $(DESTDIR)$(libdir)/
+	-$(INSTALL_M) $(build_libdir)/libgcc.a $(DESTDIR)$(libdir)/
+	-$(INSTALL_M) $(build_libdir)/libmsvcrt.a $(DESTDIR)$(libdir)/
 else
 
 # Copy over .dSYM directories directly for Darwin
@@ -333,6 +341,11 @@ else ifeq ($(JULIA_BUILD_MODE),debug)
 	$(INSTALL_M) $(build_private_libdir)/sys-debug.$(SHLIB_EXT) $(DESTDIR)$(private_libdir)
 endif
 
+	# Cache stdlibs
+	@$(call PRINT_JULIA, $(call spawn,$(JULIA_EXECUTABLE)) --startup-file=no $(JULIAHOME)/contrib/cache_stdlibs.jl)
+	# CI uses `--check-bounds=yes` which impacts the cache flags
+	@$(call PRINT_JULIA, $(call spawn,$(JULIA_EXECUTABLE)) --startup-file=no --check-bounds=yes $(JULIAHOME)/contrib/cache_stdlibs.jl)
+
 	# Copy in all .jl sources as well
 	mkdir -p $(DESTDIR)$(datarootdir)/julia/base $(DESTDIR)$(datarootdir)/julia/test
 	cp -R -L $(JULIAHOME)/base/* $(DESTDIR)$(datarootdir)/julia/base

base/Base.jl

@@ -418,6 +418,7 @@ include("threadcall.jl")
 include("uuid.jl")
 include("pkgid.jl")
 include("toml_parser.jl")
+include("linking.jl")
 include("loading.jl")
 
 # misc useful functions & macros

base/Makefile

@@ -81,6 +81,10 @@ ifeq ($(DARWIN_FRAMEWORK), 1)
 	@echo "const DARWIN_FRAMEWORK_NAME = \"$(FRAMEWORK_NAME)\"" >> $@
 else
 	@echo "const DARWIN_FRAMEWORK = false" >> $@
+endif
+ifeq ($(OS), Darwin)
+	@echo "const MACOS_PRODUCT_VERSION = \"$(shell sw_vers -productVersion)\"" >> $@
+	@echo "const MACOS_PLATFORM_VERSION = \"$(shell xcrun --show-sdk-version)\"" >> $@
 endif
 	@echo "const BUILD_TRIPLET = \"$(BB_TRIPLET_LIBGFORTRAN_CXXABI)\"" >> $@
 

base/linking.jl

@@ -0,0 +1,147 @@
+# This file is a part of Julia. License is MIT: https://julialang.org/license
+module Linking
+
+import Base.Libc: Libdl
+
+# inlined LLD_jll
+# These get calculated in __init__()
+const PATH = Ref("")
+const LIBPATH = Ref("")
+const PATH_list = String[]
+const LIBPATH_list = String[]
+const lld_path = Ref{String}()
+const lld_exe = Sys.iswindows() ? "lld.exe" : "lld"
+
+if Sys.iswindows()
+    const LIBPATH_env = "PATH"
+    const LIBPATH_default = ""
+    const pathsep = ';'
+elseif Sys.isapple()
+    const LIBPATH_env = "DYLD_FALLBACK_LIBRARY_PATH"
+    const LIBPATH_default = "~/lib:/usr/local/lib:/lib:/usr/lib"
+    const pathsep = ':'
+else
+    const LIBPATH_env = "LD_LIBRARY_PATH"
+    const LIBPATH_default = ""
+    const pathsep = ':'
+end
+
+function adjust_ENV!(env::Dict, PATH::String, LIBPATH::String, adjust_PATH::Bool, adjust_LIBPATH::Bool)
+    if adjust_LIBPATH
+        LIBPATH_base = get(env, LIBPATH_env, expanduser(LIBPATH_default))
+        if !isempty(LIBPATH_base)
+            env[LIBPATH_env] = string(LIBPATH, pathsep, LIBPATH_base)
+        else
+            env[LIBPATH_env] = LIBPATH
+        end
+    end
+    if adjust_PATH && (LIBPATH_env != "PATH" || !adjust_LIBPATH)
+        if !isempty(get(env, "PATH", ""))
+            env["PATH"] = string(PATH, pathsep, env["PATH"])
+        else
+            env["PATH"] = PATH
+        end
+    end
+    return env
+end
+
+function __init_lld_path()
+    # Prefer our own bundled lld, but if we don't have one, pick it up off of the PATH
+    # If this is an in-tree build, `lld` will live in `tools`.  Otherwise, it'll be in `libexec`
+    for bundled_lld_path in (joinpath(Sys.BINDIR, Base.LIBEXECDIR, lld_exe),
+                             joinpath(Sys.BINDIR, "..", "tools", lld_exe),
+                             joinpath(Sys.BINDIR, lld_exe))
+        if isfile(bundled_lld_path)
+            lld_path[] = abspath(bundled_lld_path)
+            return
+        end
+    end
+    lld_path[] = something(Sys.which(lld_exe), lld_exe)
+    return
+end
+
+const VERBOSE = Ref{Bool}(false)
+
+function __init__()
+    VERBOSE[] = parse(Bool, get(ENV, "JULIA_VERBOSE_LINKING", "false"))
+
+    __init_lld_path()
+    PATH[] = dirname(lld_path[])
+    if Sys.iswindows()
+        # On windows, the dynamic libraries (.dll) are in Sys.BINDIR ("usr\\bin")
+        append!(LIBPATH_list, [abspath(Sys.BINDIR, Base.LIBDIR, "julia"), Sys.BINDIR])
+    else
+        append!(LIBPATH_list, [abspath(Sys.BINDIR, Base.LIBDIR, "julia"), abspath(Sys.BINDIR, Base.LIBDIR)])
+    end
+    LIBPATH[] = join(LIBPATH_list, pathsep)
+    return
+end
+
+function lld(; adjust_PATH::Bool = true, adjust_LIBPATH::Bool = true)
+    env = adjust_ENV!(copy(ENV), PATH[], LIBPATH[], adjust_PATH, adjust_LIBPATH)
+    return Cmd(Cmd([lld_path[]]); env)
+end
+
+function ld()
+    default_args = ``
+    @static if Sys.iswindows()
+        # LLD supports mingw style linking
+        flavor = "gnu"
+        m = Sys.ARCH == :x86_64 ? "i386pep" : "i386pe"
+        default_args = `-m $m -Bdynamic --enable-auto-image-base --allow-multiple-definition`
+    elseif Sys.isapple()
+        flavor = "darwin"
+        arch = Sys.ARCH == :aarch64 ? :arm64 : Sys.ARCH
+        default_args = `-arch $arch -undefined dynamic_lookup -platform_version macos $(Base.MACOS_PRODUCT_VERSION) $(Base.MACOS_PLATFORM_VERSION)`
+    else
+        flavor = "gnu"
+    end
+
+    `$(lld()) -flavor $flavor $default_args`
+end
+
+const WHOLE_ARCHIVE = if Sys.isapple()
+    "-all_load"
+else
+    "--whole-archive"
+end
+
+const NO_WHOLE_ARCHIVE = if Sys.isapple()
+    ""
+else
+    "--no-whole-archive"
+end
+
+const SHARED = if Sys.isapple()
+    "-dylib"
+else
+    "-shared"
+end
+
+is_debug() = ccall(:jl_is_debugbuild, Cint, ()) == 1
+libdir() = abspath(Sys.BINDIR, Base.LIBDIR)
+private_libdir() = abspath(Sys.BINDIR, Base.PRIVATE_LIBDIR)
+if Sys.iswindows()
+    shlibdir() = Sys.BINDIR
+else
+    shlibdir() = libdir()
+end
+
+function link_image_cmd(path, out)
+    LIBDIR = "-L$(libdir())"
+    PRIVATE_LIBDIR = "-L$(private_libdir())"
+    SHLIBDIR = "-L$(shlibdir())"
+    LIBS = is_debug() ? ("-ljulia-debug", "-ljulia-internal-debug") : ("-ljulia", "-ljulia-internal")
+    @static if Sys.iswindows()
+        LIBS = (LIBS..., "-lopenlibm", "-lssp", "-lgcc_s", "-lgcc", "-lmsvcrt")
+    end
+
+    V = VERBOSE[] ? "--verbose" : ""
+    `$(ld()) $V $SHARED -o $out $WHOLE_ARCHIVE $path $NO_WHOLE_ARCHIVE $LIBDIR $PRIVATE_LIBDIR $SHLIBDIR $LIBS`
+end
+
+function link_image(path, out, internal_stderr::IO = stderr, internal_stdout::IO = stdout)
+    run(link_image_cmd(path, out), Base.DevNull(), stderr, stdout)
+end
+
+end # module Linking