stdenv/linux/default.nix: add gcc rebuild during bootstrap

Before the change our rebuild chain was the followin:

- `bootstrapTools` install
- `binutils` built by `bootstrapTools`
- `glibc` built by `bootstrapTools`
- `gcc` built by `bootstrapTools`

As a result `glibc` contained code generated by `bootstrapTools`'s
`gcc`. And (what worse) copied `libgcc_s.so.1` as is from
`bootstrapTools`'s `gcc`.

Such `libgcc_s.so.1` was not compatible with `nixpkgs` `gcc` at
least on `aarch64` where newer `libgcc` is expected to have new symbols.
As a result linking failed as:

    ld: /build/test.o: in function `foo(int)':
    test.cpp:(.text+0x2c): undefined reference to `__aarch64_ldadd4_acq_rel'
    collect2: error: ld returned 1 exit status

The change rejigs rebuild sequence as:

- `bootstrapTools` install
- `binutils` built by `bootstrapTools`
- `gcc` built by `bootstrapTools`
- `glibc` built by `gcc`
- `gcc` built by `gcc`

As a result `glibc` gets built by fresher `gcc` and embeds a copy
of `libgcc_s.so.1`.

Co-authored-by: Rick van Schijndel <[email protected]>

pkgs/stdenv/linux/default.nix

@@ -9,14 +9,11 @@
 # is used to build all other packages (including the bootstrapFiles).
 #
 # 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).
+# 1. final stdenv must not reference any of the bootstrap packages.
+# 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
@@ -109,12 +106,13 @@ let
     '';
 
 
-  # The bootstrap process proceeds in several steps.
-
-
-  # Create a standard environment by downloading pre-built binaries of
-  # coreutils, GCC, etc.
-
+  # The bootstrap process proceeds in several steps. Our high-level plan
+  # is the following:
+  #
+  # - build binutils with bootstrapTools
+  # - build gcc with bootstrapTools
+  # - build glibc with bootstrapTools + gcc from nixpkgs (embed libgcc_s.so.1)
+  # - build gcc with nixpkgs toolchain
 
   # Download and unpack the bootstrap tools (coreutils, GCC, Glibc, ...).
   bootstrapTools = import (if localSystem.libc == "musl" then ./bootstrap-tools-musl else ./bootstrap-tools) {
@@ -284,16 +282,56 @@ in
     };
   })
 
-
   # 2nd stdenv that contains our own rebuilt binutils and is used for
-  # compiling our own Glibc.
+  # compiling our own gcc.
   #
   # resulting stage2 stdenv:
-  # - coreutils, glibc, gcc: from bootstrapFiles
-  # - binutils: from nixpkgs, built by bootstrapFiles toolchain
+  # - coreutils, glibc: from bootstrapFiles
+  # - binutils, gcc: from nixpkgs, built by bootstrapFiles toolchain
   (prevStage: stageFun prevStage {
     name = "bootstrap-stage2";
 
+    overrides = self: super: {
+      inherit (prevStage)
+        ccWrapperStdenv gettext
+        coreutils gnugrep
+        perl gnum4 bison;
+
+      ${localSystem.libc} = getLibc prevStage;
+
+      gcc-unwrapped =
+        let makeStaticLibrariesAndMark = pkg:
+              lib.makeOverridable (pkg.override { stdenv = self.makeStaticLibraries self.stdenv; })
+                .overrideAttrs (a: { pname = "${a.pname}-stage2"; });
+        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;
+      };
+    };
+
+    # `libtool` comes with obsolete config.sub/config.guess that don't recognize Risc-V.
+    extraNativeBuildInputs =
+      lib.optional (localSystem.isRiscV) prevStage.updateAutotoolsGnuConfigScriptsHook;
+  })
+
+  # 3rd stdenv that contains our own rebuilt binutils and gcc and is
+  # used for compiling our own Glibc.
+  #
+  # resulting stage3 stdenv:
+  # - coreutils: from bootstrapFiles
+  # - glibc, binutils, gcc: from nixpkgs, built by bootstrapFiles toolchain
+  (prevStage: stageFun prevStage {
+    name = "bootstrap-stage3";
+
     overrides = self: super: {
       inherit (prevStage)
         ccWrapperStdenv gettext
@@ -357,15 +395,16 @@ in
   })
 
 
-  # Construct a third stdenv identical to the 2nd, except that this
+  # Construct a fourth stdenv identical to the 3rd, except that this
   # one uses the rebuilt Glibc from stage2.  It still uses the recent
   # binutils and rest of the bootstrap tools, including GCC.
   #
-  # resulting stage3 stdenv:
+  # resulting stage4 stdenv:
   # - coreutils, gcc: from bootstrapFiles
-  # - glibc, binutils: from nixpkgs, built by bootstrapFiles toolchain
+  # - binutils, gcc: from nixpkgs, built by bootstrapFiles toolchain
+  # - glibc: from nixpkgs, built by nixpkgs toolchain
   (prevStage: stageFun prevStage {
-    name = "bootstrap-stage3";
+    name = "bootstrap-stage4";
 
     overrides = self: super: rec {
       inherit (prevStage)
@@ -376,7 +415,7 @@ in
       gcc-unwrapped =
         let makeStaticLibrariesAndMark = pkg:
               lib.makeOverridable (pkg.override { stdenv = self.makeStaticLibraries self.stdenv; })
-                .overrideAttrs (a: { pname = "${a.pname}-stage3"; });
+                .overrideAttrs (a: { pname = "${a.pname}-stage4"; });
         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
@@ -398,21 +437,15 @@ in
   })
 
 
-  # Construct a fourth stdenv that uses the new GCC.  But coreutils is
+  # Construct a fifth stdenv that uses the new GCC.  But coreutils is
   # still from the bootstrap tools.
   #
-  # resulting stage4 stdenv:
+  # resulting stage5 stdenv:
   # - coreutils: from bootstrapFiles
-  # - glibc, binutils: from nixpkgs, built by bootstrapFiles toolchain
-  # - gcc: from nixpkgs, built by bootstrapFiles toolchain. Can assume
-  #        it has almost no code from bootstrapTools as gcc bootstraps
-  #        internally. The only exceptions are crt files from glibc
-  #        built by bootstrapTools 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).
+  # - binutils: from nixpkgs, built by bootstrapFiles toolchain
+  # - glibc, gcc: from nixpkgs, built by nixpkgs toolchain
   (prevStage: stageFun prevStage {
-    name = "bootstrap-stage4";
+    name = "bootstrap-stage5";
 
     overrides = self: super: {
       # Zlib has to be inherited and not rebuilt in this stage,
@@ -433,7 +466,7 @@ 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 = "${a.pname}-stage4"; });
+      gmp = lib.makeOverridable (super.gmp.override { stdenv = self.stdenv; }).overrideAttrs (a: { pname = "${a.pname}-stage5"; });
 
       # To allow users' overrides inhibit dependencies too heavy for
       # bootstrap, like guile: https://github.com/NixOS/nixpkgs/issues/181188
@@ -469,15 +502,7 @@ in
   # binutils built.
   #
   # resulting stage5 (final) stdenv:
-  # - coreutils, binutils: from nixpkgs, built by nixpkgs toolchain
-  # - glibc: from nixpkgs, built by bootstrapFiles toolchain
-  # - gcc: from nixpkgs, built by bootstrapFiles toolchain. Can assume
-  #        it has almost no code from bootstrapTools as gcc bootstraps
-  #        internally. The only exceptions are crt files from glibc
-  #        built by bootstrapTools 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).
+  # - coreutils, binutils, glibc, gcc: from nixpkgs, built by nixpkgs toolchain
   (prevStage: {
     inherit config overlays;
     stdenv = import ../generic rec {