document how build system find toolchains on Windows

8 files changed, 83 insertions, 12 deletions

diff --git a/aclocal.m4 b/aclocal.m4
index 8d52319bb9..78ce4ea9f3 100644
--- a/aclocal.m4
+++ b/aclocal.m4
@@ -507,6 +507,7 @@ AC_DEFUN([GET_ARM_ISA],
 # FP_SETTINGS
 # ----------------------------------
 # Set the variables used in the settings file
+# See Note [tooldir: How GHC finds mingw on Windows]
 AC_DEFUN([FP_SETTINGS],
 [
     if test "$windows" = YES -a "$EnableDistroToolchain" = "NO"
diff --git a/compiler/GHC/SysTools/BaseDir.hs b/compiler/GHC/SysTools/BaseDir.hs
index 31077451d7..f47a32e024 100644
--- a/compiler/GHC/SysTools/BaseDir.hs
+++ b/compiler/GHC/SysTools/BaseDir.hs
@@ -34,16 +34,6 @@ import System.FilePath
 import System.Directory (doesDirectoryExist)
 #endif
 
-#if defined(mingw32_HOST_OS)
-# if defined(i386_HOST_ARCH)
-#  define WINDOWS_CCONV stdcall
-# elif defined(x86_64_HOST_ARCH)
-#  define WINDOWS_CCONV ccall
-# else
-#  error Unknown mingw32 arch
-# endif
-#endif
-
 {-
 Note [topdir: How GHC finds its files]
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -70,10 +60,83 @@ Note [tooldir: How GHC finds mingw on Windows]
 GHC has some custom logic on Windows for finding the mingw
 toolchain and perl. Depending on whether GHC is built
 with the make build system or Hadrian, and on whether we're
-running a bindist, we might find the mingw toolchain and perl
+running a bindist, we might find the mingw toolchain
 either under $topdir/../{mingw, perl}/ or
 $topdir/../../{mingw, perl}/.
 
+This story is long and with lots of twist and turns..  But lets talk about how
+the build system finds and wires through the toolchain information.
+
+1) It all starts in configure.ac which has two modes it operates on:
+   a) The default is where `EnableDistroToolchain` is false.  This indicates
+      that we want to use the in-tree bundled toolchains.  In this mode we will
+      download and unpack some custom toolchains into the `inplace/mingw` folder
+      and everything is pointed to that folder.
+   b) The second path is when `EnableDistroToolchain` is true.  This makes the
+      toolchain behave a lot like Linux, in that  the environment is queried for
+      information on the tools we require.
+
+  From configure.ac we export the standard variables to set the paths to the
+  tools for the build system to use.
+
+2) After we have the path to the tools we have to generate the right paths to
+   store in the settings file for ghc to use.  This is done in aclocal.m4.
+   Again we have two modes of operation:
+   a) If not `EnableDistroToolchain` the paths are rewritten to paths using a
+      variable `$tooldir` as we need an absolute path.  $tooldir is filled in by
+      the `expandToolDir` function in this module at GHC startup.
+   b) When `EnableDistroToolchain` then instead of filling in a absolute path
+      we fill in just the program name.  The assumption here is that at runtime
+      the environment GHC is operating on will be the same as the one configure
+      was run in.  This means we expect `gcc, ld, as` etc to be on the PATH.
+
+  From `aclocal.m4` we export a couple of variables starting with `Settings`
+  which will be used to generate the settings file.
+
+3) The next step is to generate the settings file, this is where things diverge
+   based on the build system.  Both Make and Hadrian handle this differently:
+
+make)
+  Make deals with this rather simply.  As an output of configure.ac
+  `config.mk.in` is processed and `config.mk` generated which has the values we
+  set in `aclocal.m4`. This allows the rest of the build system to have access
+  to these and other values determined by configure.
+
+  Based on this file, `includes/ghc.mk` when ran will produce the settings file
+  by echoing the values into a the final file.  Coincidentally this is also
+  where `ghcplatform.h` and `ghcversion.h` generated which contains information
+  about the build platform and sets CPP for use by the entire build.
+
+hadrian)
+  For hadrian the file `cfg/system.config.in` is preprocessed by configure and
+  the output written to `system.config`.  This serves the same purpose as
+  `config.mk` but it rewrites the values that were exported.  As an example
+  `SettingsCCompilerCommand` is rewritten to `settings-c-compiler-command`.
+
+  Next up is `src/Oracles/Settings.hs` which makes from some Haskell ADT to
+  the settings `keys` in the `system.config`.  As an example,
+  `settings-c-compiler-command` is mapped to
+  `SettingsFileSetting_CCompilerCommand`.
+
+  The last part of this is the `generateSettings` in `src/Rules/Generate.hs`
+  which produces the desired settings file out of Hadrian. This is the
+  equivalent to `includes/ghc.mk`.
+
+--
+
+So why do we have these? On Windows there's no such thing as a platform compiler
+and as such we need to provide GCC and binutils.  The easiest way is to bundle
+these with the compiler and wire them up.  This gives you a relocatable
+binball.  This works fine for most users.  However mingw-w64 have a different
+requirement.  They require all packages in the repo to be compiled using the
+same version of the compiler.  So it means when they are rebuilding the world to
+add support for GCC X, they expect all packages to have been compiled with GCC X
+which is a problem since we ship an older GCC version.
+
+GHC is a package in mingw-w64 because there are Haskell packages in the
+repository which of course requires a Haskell compiler.  To help them we
+provide the override which allows GHC to instead of using an inplace compiler to
+play nice with the system compiler instead.
 -}
 
 -- | Expand occurrences of the @$tooldir@ interpolation in a string
@@ -113,7 +176,7 @@ tryFindTopDir Nothing
              Nothing -> getBaseDir
 
 
--- See Note [tooldir: How GHC finds mingw and perl on Windows]
+-- See Note [tooldir: How GHC finds mingw on Windows]
 -- Returns @Nothing@ when not on Windows.
 -- When called on Windows, it either throws an error when the
 -- tooldir can't be located, or returns @Just tooldirpath@.
diff --git a/configure.ac b/configure.ac
index 6a422d7ff9..5e3dfa5470 100644
--- a/configure.ac
+++ b/configure.ac
@@ -435,6 +435,7 @@ set_up_tarballs() {
     fi
 }
 
+# See Note [tooldir: How GHC finds mingw on Windows]
 if test "$HostOS" = "mingw32" -a "$EnableDistroToolchain" = "NO"
 then
     test -d inplace || mkdir inplace
@@ -455,6 +456,7 @@ fi
 
 # We don't want to bundle a MinGW-w64 toolchain
 # So we have to find these individual tools.
+# See Note [tooldir: How GHC finds mingw on Windows]
 if test "$EnableDistroToolchain" = "YES"
 then
     # Ideally should use AC_CHECK_TARGET_TOOL but our triples
diff --git a/hadrian/cfg/system.config.in b/hadrian/cfg/system.config.in
index 3fa53330e1..5abcb96b7e 100644
--- a/hadrian/cfg/system.config.in
+++ b/hadrian/cfg/system.config.in
@@ -129,6 +129,7 @@ conf-merge-objects-args-stage3  = @SettingsMergeObjectsFlags@
 # We are in the process of moving the settings file from being entirely
 # generated by configure, to generated being by the build system. Many of these
 # might become redundant.
+# See Note [tooldir: How GHC finds mingw on Windows]
 
 gcc-extra-via-c-opts = @GccExtraViaCOpts@
 ld-has-no-compact-unwind = @LdHasNoCompactUnwind@
diff --git a/hadrian/src/Oracles/Setting.hs b/hadrian/src/Oracles/Setting.hs
index 1e65ddb52d..06ea13d3f8 100644
--- a/hadrian/src/Oracles/Setting.hs
+++ b/hadrian/src/Oracles/Setting.hs
@@ -184,6 +184,7 @@ settingList key = fmap words $ lookupValueOrError configFile $ case key of
 
 -- | Look up the value of a 'SettingList' in @cfg/system.config@, tracking the
 -- result.
+-- See Note [tooldir: How GHC finds mingw on Windows]
 settingsFileSetting :: SettingsFileSetting -> Action String
 settingsFileSetting key = lookupValueOrError configFile $ case key of
     SettingsFileSetting_CCompilerCommand -> "settings-c-compiler-command"
diff --git a/hadrian/src/Rules/Generate.hs b/hadrian/src/Rules/Generate.hs
index 99a1a9e42f..430946020c 100644
--- a/hadrian/src/Rules/Generate.hs
+++ b/hadrian/src/Rules/Generate.hs
@@ -283,6 +283,7 @@ generateGhcPlatformH = do
         , "#endif /* __GHCPLATFORM_H__ */"
         ]
 
+-- See Note [tooldir: How GHC finds mingw on Windows]
 generateSettings :: Expr String
 generateSettings = do
     ctx <- getContext
diff --git a/includes/ghc.mk b/includes/ghc.mk
index 38addb815e..f4b85d4e2f 100644
--- a/includes/ghc.mk
+++ b/includes/ghc.mk
@@ -207,6 +207,7 @@ $(eval $(call includesHeaderPlatform,1))
 
 # These settings are read by GHC at runtime, so as to not cause spurious
 # rebuilds.
+# See Note [tooldir: How GHC finds mingw on Windows]
 
 includes_SETTINGS = includes/dist/build/settings
 
diff --git a/mk/config.mk.in b/mk/config.mk.in
index 77da4d86bd..11760686ea 100644
--- a/mk/config.mk.in
+++ b/mk/config.mk.in
@@ -475,6 +475,7 @@ endif
 # We are in the process of moving the settings file from being entirely
 # generated by configure, to generated being by the build system. Many of these
 # might become redundant.
+# See Note [tooldir: How GHC finds mingw on Windows]
 
 GccExtraViaCOpts = @GccExtraViaCOpts@
 LdHasFilelist = @LdHasFilelist@