lib/tests/triples.nix: add test suite and fix our parser so it passes

lib/systems/default.nix

@@ -52,7 +52,7 @@ rec {
       parsed = parse.mkSystemFromString (if args ? config then args.config else args.system);
       # Either of these can be losslessly-extracted from `parsed` iff parsing succeeds.
       system = parse.doubleFromSystem final.parsed;
-      config = parse.tripleFromSystem final.parsed;
+      config = parse.tripleFromSystemLossy final.parsed;
       # Determine whether we can execute binaries built for the provided platform.
       canExecute = platform:
         final.isAndroid == platform.isAndroid &&
@@ -325,7 +325,7 @@ rec {
             }.${cpu.name} or cpu.name;
             vendor_ = final.rust.platform.vendor;
           in rust.config
-            or "${cpu_}-${vendor_}-${kernel.name}${lib.optionalString (abi.name != "unknown") "-${abi.name}"}";
+            or "${cpu_}-${vendor_}-${kernel.name}${lib.optionalString (abi.name != "unknown" && abi.name != "") "-${abi.name}"}";
 
           # The name of the rust target if it is standard, or the json file
           # containing the custom target spec.

lib/systems/doubles.nix

@@ -19,7 +19,7 @@ let
     "aarch64-genode" "i686-genode" "x86_64-genode"
 
     # illumos
-    "x86_64-solaris"
+    "x86_64-solaris2"
 
     # JS
     "javascript-ghcjs"
@@ -43,7 +43,9 @@ let
     "aarch64_be-none" "aarch64-none" "arm-none" "armv6l-none" "avr-none" "i686-none"
     "microblaze-none" "microblazeel-none" "mips-none" "mips64-none" "msp430-none" "or1k-none" "m68k-none"
     "powerpc-none" "powerpcle-none" "riscv32-none" "riscv64-none" "rx-none"
-    "s390-none" "s390x-none" "vc4-none" "x86_64-none"
+    "s390-none" "s390x-none"
+    "vc4-elf"  # nonstandard triple, requires many hacks
+    "x86_64-none"
 
     # OpenBSD
     "i686-openbsd" "x86_64-openbsd"

lib/systems/inspect.nix

@@ -73,7 +73,7 @@ rec {
     isMacOS        = { kernel = kernels.macos; };
     isiOS          = { kernel = kernels.ios; };
     isLinux        = { kernel = kernels.linux; };
-    isSunOS        = { kernel = kernels.solaris; };
+    isSunOS        = { kernel = kernels.solaris2; };
     isFreeBSD      = { kernel = { name = "freebsd"; }; };
     isNetBSD       = { kernel = kernels.netbsd; };
     isOpenBSD      = { kernel = kernels.openbsd; };

lib/systems/parse.nix

@@ -230,13 +230,19 @@ rec {
   vendors = setTypes types.openVendor {
     apple = {};
     pc = {};
+    ibm = {};
     knuth = {};
+    xilinx = {};
+    solo5 = {};
 
     # Actually matters, unlocking some MinGW-w64-specific options in GCC. See
     # bottom of https://sourceforge.net/p/mingw-w64/wiki2/Unicode%20apps/
     w64 = {};
 
-    none = {};
+    # gnu-config distinguishes between `foo-bar-baz` and
+    # `foo-unknown-bar-baz`; we use `""` to represent the
+    # former case.
+    "" = { name = ""; };
     unknown = {};
   };
 
@@ -287,6 +293,9 @@ rec {
 
   types.kernel = enum (attrValues kernels);
 
+  # TODO (@amjoseph): we should eliminate execFormats or move it to
+  # abis; executable format is a property of the ABI not the kernel.
+  # This will also clear up the netbsd kludge in tripleFromSystem.
   kernels = with execFormats; with kernelFamilies; setTypes types.openKernel {
     # TODO(@Ericson2314): Don't want to mass-rebuild yet to keeping 'darwin' as
     # the normalized name for macOS.
@@ -300,8 +309,9 @@ rec {
     linux    = { execFormat = elf;     families = { }; };
     netbsd   = { execFormat = elf;     families = { inherit bsd; }; };
     none     = { execFormat = unknown; families = { }; };
+    ""       = { execFormat = unknown; families = { }; };
     openbsd  = { execFormat = elf;     families = { inherit bsd; }; };
-    solaris  = { execFormat = elf;     families = { }; };
+    solaris2 = { execFormat = elf;     families = { }; name = "solaris"; version = 2; };
     wasi     = { execFormat = wasm;    families = { }; };
     redox    = { execFormat = elf;     families = { }; };
     windows  = { execFormat = pe;      families = { }; };
@@ -324,23 +334,23 @@ rec {
     merge = mergeOneOption;
   };
 
-  types.abi = enum (attrValues abis);
+  types.abi = enum (attrValues (builtins.removeAttrs abis ["unknown"]));
 
-  abis = setTypes types.openAbi {
-    cygnus       = {};
-    msvc         = {};
+  abis = setTypes types.openAbi rec {
+    cygnus       = { kernels = [ "windows" ]; };
+    msvc         = { kernels = [ "windows" ]; };
 
     # Note: eabi is specific to ARM and PowerPC.
     # On PowerPC, this corresponds to PPCEABI.
     # On ARM, this corresponds to ARMEABI.
-    eabi         = { float = "soft"; };
-    eabihf       = { float = "hard"; };
+    eabi         = { float = "soft"; kernels = builtins.attrNames kernels; };
+    eabihf       = { float = "hard"; kernels = builtins.attrNames kernels; };
 
     # Other architectures should use ELF in embedded situations.
-    elf          = {};
+    elf          = { kernels = [ "none" "" ]; };
 
-    androideabi  = {};
-    android      = {
+    androideabi  = { inherit (android) kernels; };
+    android      = { kernels = [ "linux" ];
       assertions = [
         { assertion = platform: !platform.isAarch32;
           message = ''
@@ -350,9 +360,9 @@ rec {
       ];
     };
 
-    gnueabi      = { float = "soft"; };
-    gnueabihf    = { float = "hard"; };
-    gnu          = {
+    gnueabi      = { float = "soft"; inherit (gnu) kernels; };
+    gnueabihf    = { float = "hard"; inherit (gnu) kernels; };
+    gnu          = { kernels = [ "freebsd12" "freebsd13" "linux" "windows" ];
       assertions = [
         { assertion = platform: !platform.isAarch32;
           message = ''
@@ -366,27 +376,34 @@ rec {
         }
       ];
     };
-    gnuabi64     = { abi = "64"; };
-    muslabi64    = { abi = "64"; };
+    gnuabi64     = { abi = "64"; inherit (gnu) kernels; };
+    muslabi64    = { abi = "64"; kernels = [ "linux" ]; };
 
     # NOTE: abi=n32 requires a 64-bit MIPS chip!  That is not a typo.
     # It is basically the 64-bit abi with 32-bit pointers.  Details:
     # https://www.linux-mips.org/pub/linux/mips/doc/ABI/MIPS-N32-ABI-Handbook.pdf
-    gnuabin32    = { abi = "n32"; };
-    muslabin32   = { abi = "n32"; };
+    gnuabin32    = { abi = "n32"; inherit (gnu) kernels; };
+    muslabin32   = { abi = "n32"; inherit (musl) kernels; };
 
-    gnuabielfv2  = { abi = "elfv2"; };
-    gnuabielfv1  = { abi = "elfv1"; };
+    gnuabielfv2  = { abi = "elfv2"; kernels = [ "linux" ]; };
+    gnuabielfv1  = { abi = "elfv1"; kernels = [ "linux" ]; };
 
-    musleabi     = { float = "soft"; };
-    musleabihf   = { float = "hard"; };
-    musl         = {};
+    musleabi     = { float = "soft"; inherit (musl) kernels; };
+    musleabihf   = { float = "hard"; inherit (musl) kernels; };
+    musl         = { kernels = [ "linux" ]; };
 
-    uclibceabi   = { float = "soft"; };
-    uclibceabihf = { float = "hard"; };
-    uclibc       = {};
+    uclibceabi   = { float = "soft"; inherit (uclibc) kernels; };
+    uclibceabihf = { float = "hard"; inherit (uclibc) kernels; };
+    uclibc       = { kernels = [ "linux" ]; };
 
-    unknown = {};
+    "" = { kernels = [ "darwin" "freebsd12" "freebsd13" "freebsd" "genode" "solaris2" "solaris" "ghcjs" "mmixware" "netbsd" "none" "" "openbsd" "redox" "wasi" ]; };
+
+    # in gnu-config triples, this abi is actually the empty string "" rather than "-unknown"
+    unknown =
+      # Added 2023-08-18.  When removing, please also remove the
+      # `removeAttrs ["unknown"]` in lib.tests.triples and
+      # lib.systems.types.abi
+      abis."" // { assertions = lib.warn "please use abis.\"\" instead; \"unknown\" is not a valid ABI." [ ]; };
   };
 
   ################################################################################
@@ -406,11 +423,40 @@ rec {
 
   mkSystem = components:
     assert types.parsedPlatform.check components;
+    assert with components;
+      (kernel == kernels.ghcjs) != (cpu == cpuTypes.javascript)
+      -> throw ''
+        cpu "javascript" and kernel "ghcjs" are valid only with each other;
+        you attempted to use cpu "${components.cpu.name}" with kernel
+        "${components.kernel.name}"
+      '';
+    assert with components;
+      !(lib.elem kernel.name (abi.kernels or [ kernel.name ])) &&
+      !(isVc4 components)
+       -> throw ''
+         kernel ${components.kernel.name} does not allow abi ${components.abi.name}
+       '';
+    assert with components;
+      !(isVc4 components -> (with components; vendor.name == "" && kernel.name == "" && abi.name == "elf"))
+       -> throw ''
+         Broadcom VC4 CPUs may be used only in the nonstandard triple "vc4-elf".
+         You tried to create "${cpu.name}-${vendor.name}-${kernel.name}-${abi.name}".
+       '';
+    assert with components;
+      !(isJavaScript components ->
+        (vendor.name == "unknown" || vendor.name == "") &&
+        kernel == kernels.ghcjs &&
+        abi == abis."")
+      -> throw ''
+      The special "javascript" cpu may be used only in the nonstandard triple
+      "javascript-unknown-ghcjs".  You tried to create
+      "${cpu.name}-${vendor.name}-${kernel.name}-${abi.name}"
+      '';
     setType "system" components;
 
   mkSkeletonFromList = l: {
     "1" = if elemAt l 0 == "avr"
-      then { cpu = elemAt l 0; kernel = "none"; abi = "unknown"; }
+      then { cpu = elemAt l 0; kernel = "none"; abi = ""; }
       else throw "Target specification with 1 components is ambiguous";
     "2" = # We only do 2-part hacks for things Nix already supports
       if elemAt l 1 == "cygwin"
@@ -422,20 +468,29 @@ rec {
       else if elemAt l 1 == "windows"
         then { cpu = elemAt l 0;                      kernel = "windows";  abi = "msvc";     }
       else if (elemAt l 1) == "elf"
-        then { cpu = elemAt l 0; vendor = "unknown";  kernel = "none";     abi = elemAt l 1; }
+        then { cpu = elemAt l 0;                      kernel = "";         abi = elemAt l 1; }
       else   { cpu = elemAt l 0;                      kernel = elemAt l 1;                   };
     "3" =
+      # cpu-vendor-""-abi
+      if elemAt l 1 == "unknown" && elemAt l 2 == "elf"
+      then {
+        cpu    = elemAt l 0;
+        vendor = elemAt l 1;
+        kernel = "";
+        abi    = elemAt l 2;
+      }
       # cpu-kernel-environment
-      if elemAt l 1 == "linux" ||
-         elem (elemAt l 2) ["eabi" "eabihf" "elf" "gnu"]
+      else if elemAt l 1 == "linux" ||
+              elemAt l 1 == "none" ||
+              elem (elemAt l 2) ["eabi" "eabihf" "elf" "gnu"]
       then {
         cpu    = elemAt l 0;
         kernel = elemAt l 1;
         abi    = elemAt l 2;
-        vendor = "unknown";
       }
       # cpu-vendor-os
       else if elemAt l 1 == "apple" ||
+              elemAt l 1 == "solo5" ||
               elem (elemAt l 2) [ "wasi" "redox" "mmixware" "ghcjs" "mingw32" ] ||
               hasPrefix "freebsd" (elemAt l 2) ||
               hasPrefix "netbsd" (elemAt l 2) ||
@@ -454,39 +509,61 @@ rec {
 
   # This should revert the job done by config.guess from the gcc compiler.
   mkSystemFromSkeleton = { cpu
-                         , # Optional, but fallback too complex for here.
-                           # Inferred below instead.
-                           vendor ? assert false; null
+                         , vendor ? ""
                          , kernel
                          , # Also inferred below
-                           abi    ? assert false; null
+                           abi    ? null
                          } @ args: let
     getCpu    = name: cpuTypes.${name} or (throw "Unknown CPU type: ${name}");
     getVendor = name:  vendors.${name} or (throw "Unknown vendor: ${name}");
     getKernel = name:  kernels.${name} or (throw "Unknown kernel: ${name}");
     getAbi    = name:     abis.${name} or (throw "Unknown ABI: ${name}");
 
+    vendor_ = getVendor vendor;
+
+    # These the three CPUs for which gnu-config will sometimes
+    # clobber an explicitly-specified `-unknown` vendor with
+    # something else.
+    clobberedVendor =
+      if      isS390 parsed        then vendors.ibm
+      else if isMicroBlaze parsed  then vendors.xilinx
+      else if isMmix parsed        then vendors.knuth
+      else if isVc4 parsed         then vendors.""     # hack for nonstandard triple
+      else vendors.unknown;
+
     parsed = {
       cpu = getCpu args.cpu;
       vendor =
-        /**/ if args ? vendor    then getVendor args.vendor
-        else if isDarwin  parsed then vendors.apple
-        else if isWindows parsed then vendors.pc
-        else                     vendors.unknown;
+          if vendor_ == vendors.unknown then clobberedVendor
+          else if vendor_ != vendors."" then vendor_
+          else if (isx86_64 parsed || isi686 parsed) &&
+                  parsed.abi == abis.gnu then
+                    vendors.unknown  # nixpkgs-specific behavior
+          else if (isx86_64 parsed || isAarch64 parsed) &&
+                  isLittleEndian parsed &&
+                  parsed.kernel == kernels.darwin &&
+                  parsed.abi == abis."" then
+                    vendors.apple    # nixpkgs-specific behavior
+          else if isx86 parsed && isLinux parsed then vendors.pc
+          else if ((abi=="eabi" || abi=="eabihf")) && !(isLinux parsed || isNetBSD parsed) then vendor_
+          else if isNone parsed && vendor_ == vendors."" && abi == "elf" then vendors.""
+          else if isx86 parsed then vendors.pc
+          else clobberedVendor;
+
       kernel = if hasPrefix "darwin" args.kernel      then getKernel "darwin"
                else if hasPrefix "netbsd" args.kernel then getKernel "netbsd"
                else                                   getKernel (removeAbiSuffix args.kernel);
       abi =
-        /**/ if args ? abi       then getAbi args.abi
+        /**/ if abi != null      then getAbi abi
         else if isLinux parsed || isWindows parsed then
           if isAarch32 parsed then
-            if lib.versionAtLeast (parsed.cpu.version or "0") "6"
+            if parsed.vendor == vendors.apple then abis.gnu
+            else if isLinux parsed then abis.gnu
+            else if lib.versionAtLeast (parsed.cpu.version or "0") "6"
             then abis.gnueabihf
             else abis.gnueabi
-          # Default ppc64 BE to ELFv2
-          else if isPower64 parsed && isBigEndian parsed then abis.gnuabielfv2
           else abis.gnu
-        else                     abis.unknown;
+        else                     abis."";
     };
 
   in mkSystem parsed;
@@ -496,18 +573,56 @@ rec {
   kernelName = kernel:
     kernel.name + toString (kernel.version or "");
 
-  doubleFromSystem = { cpu, kernel, abi, ... }:
+  doubleFromSystem = { cpu, kernel, abi, ... }@platform:
     /**/ if abi == abis.cygnus       then "${cpu.name}-cygwin"
     else if kernel.families ? darwin then "${cpu.name}-darwin"
+    else if isVc4 platform           then "${cpu.name}-${abi.name}"  # hack for nonstandard triple
     else "${cpu.name}-${kernelName kernel}";
 
+  # Convert a system into a string triple
   tripleFromSystem = { cpu, vendor, kernel, abi, ... } @ sys: assert isSystem sys; let
     optExecFormat =
       lib.optionalString (kernel.name == "netbsd" &&
                           gnuNetBSDDefaultExecFormat cpu != kernel.execFormat)
         kernel.execFormat.name;
-    optAbi = lib.optionalString (abi != abis.unknown) "-${abi.name}";
-  in "${cpu.name}-${vendor.name}-${kernelName kernel}${optExecFormat}${optAbi}";
+    in
+      tripleFromSkeleton {
+        cpu = cpu.name;
+        vendor = vendor.name;
+        kernel = kernelName kernel;
+        abi = abi.name;
+      };
+
+  # Convert a system into a string triple, erasing the distinction
+  # between vendors."" and vendors.unknown -- i.e. `mips-linux-gnu`
+  # and `mips-unknown-linux-gnu` both become
+  # `"mips-unknown-linux-gnu"`.
+  tripleFromSystemLossy = { cpu, vendor, kernel, abi, ... } @ sys: assert isSystem sys;
+    tripleFromSystem (sys // lib.optionalAttrs (sys.vendor == vendors."") {
+      vendor = vendors.unknown;
+    });
+
+  tripleFromSkeleton = { cpu, vendor, kernel, optExecFormat?"", abi }: let
+    optAbi = lib.optionalString (abi != "") "-${abi}";
+    optVendor = lib.optionalString (vendor != "") "-${vendor}";
+    optKernel = lib.optionalString (kernel != "") "-${kernel}";
+  in
+    # gnu-config considers "mingw32" and "cygwin" to be kernels.
+    # This is obviously bogus, which is why nixpkgs has historically
+    # parsed them differently.  However for regression testing
+    # reasons (see lib/tests/triples.nix) we need to replicate this
+    # quirk when unparsing in order to round-trip correctly.
+    if      abi == "cygnus"     then "${cpu}${optVendor}-cygwin"
+    else if kernel == "windows" then "${cpu}${optVendor}-mingw32"
+    else "${cpu}${optVendor}${optKernel}${optExecFormat}${optAbi}";
+
+  # To "canonicalize" a triple is to parse it and then unparse (turn
+  # back into a string) it.
+  canonicalize = triple:
+    lib.pipe triple [
+      mkSystemFromString   # parse
+      tripleFromSystem     # unparse
+    ];
 
   ################################################################################