Merge pull request #209870 from amjoseph-nixpkgs/pr/stdenv/external-gcc-bootstrap

3 files changed, 150 insertions, 64 deletions

diff --git a/pkgs/stdenv/linux/bootstrap-tools/scripts/unpack-bootstrap-tools.sh b/pkgs/stdenv/linux/bootstrap-tools/scripts/unpack-bootstrap-tools.sh
index 5b5677eef13..09bf25f5215 100644
--- a/pkgs/stdenv/linux/bootstrap-tools/scripts/unpack-bootstrap-tools.sh
+++ b/pkgs/stdenv/linux/bootstrap-tools/scripts/unpack-bootstrap-tools.sh
@@ -30,6 +30,13 @@ LD_LIBRARY_PATH=$out/lib $LD_BINARY $out/bin/mv $out/lib/libstdc++.* $LIBSTDCXX_
 # use a copy of patchelf.
 LD_LIBRARY_PATH=$out/lib $LD_BINARY $out/bin/cp $out/bin/patchelf .
 
+# Older versions of the bootstrap-files did not compile their
+# patchelf with -static-libgcc, so we have to be very careful not to
+# run patchelf on the same copy of libgcc_s that it links against.
+LD_LIBRARY_PATH=$out/lib $LD_BINARY $out/bin/cp $out/lib/libgcc_s.so.1 .
+LD_LIBRARY_PATH=.:$out/lib:$LIBSTDCXX_SO_DIR $LD_BINARY \
+  ./patchelf --set-rpath $out/lib --force-rpath $out/lib/libgcc_s.so.1
+
 for i in $out/bin/* $out/libexec/gcc/*/*/*; do
     if [ -L "$i" ]; then continue; fi
     if [ -z "${i##*/liblto*}" ]; then continue; fi
diff --git a/pkgs/stdenv/linux/default.nix b/pkgs/stdenv/linux/default.nix
index 3f2d77729ab..ca6d720fe77 100644
--- a/pkgs/stdenv/linux/default.nix
+++ b/pkgs/stdenv/linux/default.nix
@@ -10,13 +10,10 @@
 #
 # Goals of the bootstrap process:
 # 1. final stdenv must not reference any of the bootstrap files.
-# 2. final stdenv must not contain any of the bootstrap files
-#    (the only current violation is libgcc_s.so in glibc).
+# 2. final stdenv must not contain any of the bootstrap files.
 # 3. final stdenv must not contain any of the files directly
 #    generated by the bootstrap code generators (assembler, linker,
-#    compiler). The only current violations are: libgcc_s.so in glibc,
-#    the lib{mpfr,mpc,gmp,isl} which are statically linked
-#    into the final gcc).
+#    compiler).
 #
 # These goals ensure that final packages and final stdenv are built
 # exclusively using nixpkgs package definitions and don't depend
@@ -111,6 +108,21 @@ let
   isBuiltByBootstrapFilesCompiler =
     pkg: isFromNixpkgs pkg && isFromBootstrapFiles pkg.stdenv.cc.cc;
 
+  commonGccOverrides = {
+    # Use a deterministically built compiler
+    # see https://github.com/NixOS/nixpkgs/issues/108475 for context
+    reproducibleBuild = true;
+    profiledCompiler = false;
+
+    # It appears that libcc1 (which is not a g++ plugin; it is a gdb plugin) gets linked against
+    # the libstdc++ from the compiler that *built* g++, not the libstdc++ which was just built.
+    # This causes a reference chain from stdenv to the bootstrapFiles:
+    #
+    #   stdenv -> gcc-lib -> xgcc-lib -> bootstrapFiles
+    #
+    disableGdbPlugin = true;
+  };
+
   commonPreHook =
     ''
       export NIX_ENFORCE_PURITY="''${NIX_ENFORCE_PURITY-1}"
@@ -170,7 +182,7 @@ let
 
         cc = if prevStage.gcc-unwrapped == null
              then null
-             else lib.makeOverridable (import ../../build-support/cc-wrapper) {
+             else (lib.makeOverridable (import ../../build-support/cc-wrapper) {
           name = "${name}-gcc-wrapper";
           nativeTools = false;
           nativeLibc = false;
@@ -184,7 +196,12 @@ let
           inherit lib;
           inherit (prevStage) coreutils gnugrep;
           stdenvNoCC = prevStage.ccWrapperStdenv;
-        };
+        }).overrideAttrs(a: lib.optionalAttrs (prevStage.gcc-unwrapped.passthru.isXgcc or false) {
+          # This affects only `xgcc` (the compiler which compiles the final compiler).
+          postFixup = (a.postFixup or "") + ''
+            echo "--sysroot=${lib.getDev (getLibc prevStage)}" >> $out/nix-support/cc-cflags
+          '';
+        });
 
         overrides = self: super: (overrides self super) // { fetchurl = thisStdenv.fetchurlBoot; };
       };
@@ -226,7 +243,7 @@ in
       ${localSystem.libc} = self.stdenv.mkDerivation {
         pname = "bootstrap-stage0-${localSystem.libc}";
         strictDeps = true;
-        version = "bootstrap";
+        version = "bootstrapFiles";
         enableParallelBuilding = true;
         buildCommand = ''
           mkdir -p $out
@@ -282,7 +299,7 @@ in
       };
       inherit (prevStage)
         ccWrapperStdenv
-        gcc-unwrapped coreutils gnugrep;
+        gcc-unwrapped coreutils gnugrep binutils;
 
       ${localSystem.libc} = getLibc prevStage;
 
@@ -295,6 +312,75 @@ in
     };
   })
 
+  # First rebuild of gcc; this is linked against all sorts of junk
+  # from the bootstrap-files, but we only care about the code that
+  # this compiler *emits*.  The `gcc` binary produced in this stage
+  # is not part of the final stdenv.
+  (prevStage:
+    assert isBuiltByBootstrapFilesCompiler prevStage.binutils-unwrapped;
+    assert            isFromBootstrapFiles prevStage."${localSystem.libc}";
+    assert            isFromBootstrapFiles prevStage.gcc-unwrapped;
+    assert            isFromBootstrapFiles prevStage.coreutils;
+    assert            isFromBootstrapFiles prevStage.gnugrep;
+    stageFun prevStage {
+      name = "bootstrap-stage-xgcc";
+      overrides = final: prev: {
+        inherit (prevStage) ccWrapperStdenv coreutils gnugrep gettext bison texinfo zlib gnum4 perl;
+        patchelf = bootstrapTools;
+        ${localSystem.libc} = getLibc prevStage;
+        gmp      = prev.gmp.override { cxx = false; };
+        gcc-unwrapped =
+          (prev.gcc-unwrapped.override (commonGccOverrides // {
+            # The most logical name for this package would be something like
+            # "gcc-stage1".  Unfortunately "stage" is already reserved for the
+            # layers of stdenv, so using "stage" in the name of this package
+            # would cause massive confusion.
+            #
+            # Gcc calls its "stage1" compiler `xgcc` (--disable-bootstrap results
+            # in `xgcc` being copied to $prefix/bin/gcc).  So we imitate that.
+            #
+            name = "xgcc";
+
+          })).overrideAttrs (a: {
+
+            # This signals to cc-wrapper (as overridden above in this file) to add `--sysroot`
+            # to `$out/nix-support/cc-cflags`.
+            passthru = a.passthru // { isXgcc = true; };
+
+            # Gcc will look for the C library headers in
+            #
+            #    ${with_build_sysroot}${native_system_header_dir}
+            #
+            # The ordinary gcc expression sets `--with-build-sysroot=/` and sets
+            # `native-system-header-dir` to `"${lib.getDev stdenv.cc.libc}/include`.
+            #
+            # Unfortunately the value of "--with-native-system-header-dir=" gets "burned in" to the
+            # compiler, and it is quite difficult to get the compiler to change or ignore it
+            # afterwards.  On the other hand, the `sysroot` is very easy to change; you can just pass
+            # a `--sysroot` flag to `gcc`.
+            #
+            # So we override the expression to remove the default settings for these flags, and
+            # replace them such that the concatenated value will be the same as before, but we split
+            # the value between the two variables differently: `--native-system-header-dir=/include`,
+            # and `--with-build-sysroot=${lib.getDev stdenv.cc.libc}`.
+            #
+            configureFlags = (a.configureFlags or []) ++ [
+              "--with-native-system-header-dir=/include"
+              "--with-build-sysroot=${lib.getDev final.stdenv.cc.libc}"
+            ];
+
+            # This is a separate phase because gcc assembles its phase scripts
+            # in bash instead of nix (we should fix that).
+            preFixupPhases = (a.preFixupPhases or []) ++ [ "preFixupXgccPhase" ];
+
+            # This is needed to prevent "error: cycle detected in build of '...-xgcc-....drv'
+            # in the references of output 'lib' from output 'out'"
+            preFixupXgccPhase = ''
+              find $lib/lib/ -name \*.so\* -exec patchelf --shrink-rpath {} \; || true
+            '';
+          });
+      };
+    })
 
   # 2nd stdenv that contains our own rebuilt binutils and is used for
   # compiling our own Glibc.
@@ -303,7 +389,7 @@ in
     # previous stage1 stdenv:
     assert isBuiltByBootstrapFilesCompiler prevStage.binutils-unwrapped;
     assert            isFromBootstrapFiles prevStage."${localSystem.libc}";
-    assert            isFromBootstrapFiles prevStage.gcc-unwrapped;
+    assert isBuiltByBootstrapFilesCompiler prevStage.gcc-unwrapped;
     assert            isFromBootstrapFiles prevStage.coreutils;
     assert            isFromBootstrapFiles prevStage.gnugrep;
     stageFun prevStage {
@@ -313,7 +399,7 @@ in
       inherit (prevStage)
         ccWrapperStdenv gettext
         gcc-unwrapped coreutils gnugrep
-        perl gnum4 bison;
+        perl gnum4 bison texinfo which;
       dejagnu = super.dejagnu.overrideAttrs (a: { doCheck = false; } );
 
       # We need libidn2 and its dependency libunistring as glibc dependency.
@@ -378,11 +464,12 @@ in
   # binutils and rest of the bootstrap tools, including GCC.
   (prevStage:
     # previous stage2 stdenv:
-    assert isBuiltByBootstrapFilesCompiler prevStage.binutils-unwrapped;
-    assert isBuiltByBootstrapFilesCompiler prevStage.${localSystem.libc};
-    assert            isFromBootstrapFiles prevStage.gcc-unwrapped;
+    assert        isBuiltByNixpkgsCompiler prevStage.binutils-unwrapped;
+    assert        isBuiltByNixpkgsCompiler prevStage.${localSystem.libc};
+    assert isBuiltByBootstrapFilesCompiler prevStage.gcc-unwrapped;
     assert            isFromBootstrapFiles prevStage.coreutils;
     assert            isFromBootstrapFiles prevStage.gnugrep;
+    assert lib.all isBuiltByNixpkgsCompiler (with prevStage; [ gmp isl_0_20 libmpc mpfr ]);
     stageFun prevStage {
     name = "bootstrap-stage3";
 
@@ -390,25 +477,20 @@ in
       inherit (prevStage)
         ccWrapperStdenv
         binutils coreutils gnugrep gettext
-        perl patchelf linuxHeaders gnum4 bison libidn2 libunistring;
+        perl patchelf linuxHeaders gnum4 bison libidn2 libunistring libxcrypt;
+        # We build a special copy of libgmp which doesn't use libstdc++, because
+        # xgcc++'s libstdc++ references the bootstrap-files (which is what
+        # compiles xgcc++).
+        gmp = super.gmp.override { cxx = false; };
+      } // {
       ${localSystem.libc} = getLibc prevStage;
-      gcc-unwrapped =
-        let makeStaticLibrariesAndMark = pkg:
-              lib.makeOverridable (pkg.override { stdenv = self.makeStaticLibraries self.stdenv; })
-                .overrideAttrs (a: { pname = "${a.pname}-stage3"; });
-        in super.gcc-unwrapped.override {
-        # Link GCC statically against GMP etc.  This makes sense because
-        # these builds of the libraries are only used by GCC, so it
-        # reduces the size of the stdenv closure.
-        gmp = makeStaticLibrariesAndMark super.gmp;
-        mpfr = makeStaticLibrariesAndMark super.mpfr;
-        libmpc = makeStaticLibrariesAndMark super.libmpc;
-        isl = makeStaticLibrariesAndMark super.isl_0_20;
-        # Use a deterministically built compiler
-        # see https://github.com/NixOS/nixpkgs/issues/108475 for context
-        reproducibleBuild = true;
-        profiledCompiler = false;
-      };
+      gcc-unwrapped = (super.gcc-unwrapped.override (commonGccOverrides // {
+        inherit (prevStage) which;
+      }
+      )).overrideAttrs (a: {
+        # so we can add them to allowedRequisites below
+        passthru = a.passthru // { inherit (self) gmp mpfr libmpc isl; };
+      });
     };
     extraNativeBuildInputs = [ prevStage.patchelf ] ++
       # Many tarballs come with obsolete config.sub/config.guess that don't recognize aarch64.
@@ -422,18 +504,11 @@ in
   #
   (prevStage:
     # previous stage3 stdenv:
-    assert isBuiltByBootstrapFilesCompiler prevStage.binutils-unwrapped;
-    assert isBuiltByBootstrapFilesCompiler prevStage.${localSystem.libc};
-    assert isBuiltByBootstrapFilesCompiler prevStage.gcc-unwrapped;
-    assert            isFromBootstrapFiles prevStage.coreutils;
-    assert            isFromBootstrapFiles prevStage.gnugrep;
-    # Can assume prevStage.gcc-unwrapped has almost no code from
-    # bootstrapTools as gcc bootstraps internally. The only
-    # exceptions are crt files from glibc built bybootstrapTools
-    # used to link executables and libraries, and the
-    # bootstrapTools-built, statically-linked
-    # lib{mpfr,mpc,gmp,isl}.a which are linked into the final gcc
-    # (see commit cfde88976ba4cddd01b1bb28b40afd12ea93a11d).
+    assert isBuiltByNixpkgsCompiler prevStage.binutils-unwrapped;
+    assert isBuiltByNixpkgsCompiler prevStage.${localSystem.libc};
+    assert isBuiltByNixpkgsCompiler prevStage.gcc-unwrapped;
+    assert     isFromBootstrapFiles prevStage.coreutils;
+    assert     isFromBootstrapFiles prevStage.gnugrep;
     stageFun prevStage {
     name = "bootstrap-stage4";
 
@@ -453,11 +528,6 @@ in
         };
       };
 
-      # force gmp to rebuild so we have the option of dynamically linking
-      # libgmp without creating a reference path from:
-      #   stage5.gcc -> stage4.coreutils -> stage3.glibc -> bootstrap
-      gmp = lib.makeOverridable (super.gmp.override { stdenv = self.stdenv; }).overrideAttrs (a: { pname = "${a.pname}-stage4"; });
-
       # To allow users' overrides inhibit dependencies too heavy for
       # bootstrap, like guile: https://github.com/NixOS/nixpkgs/issues/181188
       gnumake = super.gnumake.override { inBootstrap = true; };
@@ -494,11 +564,11 @@ in
   (prevStage:
     # previous stage4 stdenv; see stage3 comment regarding gcc,
     # which applies here as well.
-    assert        isBuiltByNixpkgsCompiler prevStage.binutils-unwrapped;
-    assert isBuiltByBootstrapFilesCompiler prevStage.${localSystem.libc};
-    assert isBuiltByBootstrapFilesCompiler prevStage.gcc-unwrapped;
-    assert        isBuiltByNixpkgsCompiler prevStage.coreutils;
-    assert        isBuiltByNixpkgsCompiler prevStage.gnugrep;
+    assert isBuiltByNixpkgsCompiler prevStage.binutils-unwrapped;
+    assert isBuiltByNixpkgsCompiler prevStage.${localSystem.libc};
+    assert isBuiltByNixpkgsCompiler prevStage.gcc-unwrapped;
+    assert isBuiltByNixpkgsCompiler prevStage.coreutils;
+    assert isBuiltByNixpkgsCompiler prevStage.gnugrep;
     {
     inherit config overlays;
     stdenv = import ../generic rec {
@@ -546,11 +616,15 @@ in
           )
         # More complicated cases
         ++ (map (x: getOutput x (getLibc prevStage)) [ "out" "dev" "bin" ] )
-        ++  [ /*propagated from .dev*/ linuxHeaders
-            binutils gcc gcc.cc gcc.cc.lib gcc.expand-response-params
+        ++  [ linuxHeaders # propagated from .dev
+            binutils gcc gcc.cc gcc.cc.lib gcc.expand-response-params gcc.cc.libgcc glibc.passthru.libgcc
           ]
-          ++ lib.optionals (!localSystem.isx86 || localSystem.libc == "musl")
-            [ prevStage.updateAutotoolsGnuConfigScriptsHook prevStage.gnu-config ];
+        ++ lib.optionals (!localSystem.isx86 || localSystem.libc == "musl")
+            [ prevStage.updateAutotoolsGnuConfigScriptsHook prevStage.gnu-config ]
+        ++ (with gcc-unwrapped.passthru; [
+          gmp libmpc mpfr isl
+        ])
+      ;
 
       overrides = self: super: {
         inherit (prevStage)
@@ -579,10 +653,10 @@ in
   (prevStage:
     # previous stage5 stdenv; see stage3 comment regarding gcc,
     # which applies here as well.
-    assert        isBuiltByNixpkgsCompiler prevStage.binutils-unwrapped;
-    assert isBuiltByBootstrapFilesCompiler prevStage.${localSystem.libc};
-    assert isBuiltByBootstrapFilesCompiler prevStage.gcc-unwrapped;
-    assert        isBuiltByNixpkgsCompiler prevStage.coreutils;
-    assert        isBuiltByNixpkgsCompiler prevStage.gnugrep;
+    assert isBuiltByNixpkgsCompiler prevStage.binutils-unwrapped;
+    assert isBuiltByNixpkgsCompiler prevStage.${localSystem.libc};
+    assert isBuiltByNixpkgsCompiler prevStage.gcc-unwrapped;
+    assert isBuiltByNixpkgsCompiler prevStage.coreutils;
+    assert isBuiltByNixpkgsCompiler prevStage.gnugrep;
     { inherit (prevStage) config overlays stdenv; })
 ]
diff --git a/pkgs/stdenv/linux/make-bootstrap-tools.nix b/pkgs/stdenv/linux/make-bootstrap-tools.nix
index 3aa7f6a3df5..091130ebf93 100644
--- a/pkgs/stdenv/linux/make-bootstrap-tools.nix
+++ b/pkgs/stdenv/linux/make-bootstrap-tools.nix
@@ -2,6 +2,10 @@
 
 let
   libc = pkgs.stdenv.cc.libc;
+  patchelf = pkgs.patchelf.overrideAttrs(previousAttrs: {
+    NIX_CFLAGS_COMPILE = (previousAttrs.NIX_CFLAGS_COMPILE or []) ++ [ "-static-libgcc" "-static-libstdc++" ];
+    NIX_CFLAGS_LINK = (previousAttrs.NIX_CFLAGS_LINK or []) ++ [ "-static-libgcc" "-static-libstdc++" ];
+  });
 in with pkgs; rec {
 
 
@@ -127,7 +131,7 @@ in with pkgs; rec {
         cp -d ${bootGCC.out}/bin/gcc $out/bin
         cp -d ${bootGCC.out}/bin/cpp $out/bin
         cp -d ${bootGCC.out}/bin/g++ $out/bin
-        cp -d ${bootGCC.lib}/lib/libgcc_s.so* $out/lib
+        cp    ${bootGCC.lib}/lib/libgcc_s.so* $out/lib
         cp -d ${bootGCC.lib}/lib/libstdc++.so* $out/lib
         cp -d ${bootGCC.out}/lib/libssp.a* $out/lib
         cp -d ${bootGCC.out}/lib/libssp_nonshared.a $out/lib
@@ -149,6 +153,7 @@ in with pkgs; rec {
         rm -rf $out/include/c++/*/ext/parallel
 
         cp -d ${gmpxx.out}/lib/libgmp*.so* $out/lib
+        cp -d ${isl.out}/lib/libisl*.so* $out/lib
         cp -d ${mpfr.out}/lib/libmpfr*.so* $out/lib
         cp -d ${libmpc.out}/lib/libmpc*.so* $out/lib
         cp -d ${zlib.out}/lib/libz.so* $out/lib