diff options
Diffstat (limited to 'compiler/GHC/Driver/Main.hs')
| -rw-r--r-- | compiler/GHC/Driver/Main.hs | 71 |
1 files changed, 60 insertions, 11 deletions
diff --git a/compiler/GHC/Driver/Main.hs b/compiler/GHC/Driver/Main.hs index 07f1e7acda..296a855acf 100644 --- a/compiler/GHC/Driver/Main.hs +++ b/compiler/GHC/Driver/Main.hs @@ -42,6 +42,7 @@ module GHC.Driver.Main , Messager, batchMsg , HscStatus (..) , hscIncrementalCompile + , initModDetails , hscMaybeWriteIface , hscCompileCmmFile @@ -804,16 +805,7 @@ hscIncrementalCompile always_do_basic_recompilation_check m_tc_result -- We didn't need to do any typechecking; the old interface -- file on disk was good enough. Left iface -> do - -- Knot tying! See Note [Knot-tying typecheckIface] - details <- liftIO . fixIO $ \details' -> do - let act hpt = addToHpt hpt (ms_mod_name mod_summary) - (HomeModInfo iface details' Nothing) - let hsc_env' = hscUpdateHPT act hsc_env - -- NB: This result is actually not that useful - -- in one-shot mode, since we're not going to do - -- any further typechecking. It's much more useful - -- in make mode, since this HMI will go into the HPT. - genModDetails hsc_env' iface + details <- liftIO $ initModDetails hsc_env mod_summary iface return (HscUpToDate iface details, hsc_env') -- We finished type checking. (mb_old_hash is the hash of -- the interface that existed on disk; it's possible we had @@ -823,6 +815,64 @@ hscIncrementalCompile always_do_basic_recompilation_check m_tc_result status <- finish mod_summary tc_result mb_old_hash return (status, hsc_env) +-- Knot tying! See Note [Knot-tying typecheckIface] +-- See Note [ModDetails and --make mode] +initModDetails :: HscEnv -> ModSummary -> ModIface -> IO ModDetails +initModDetails hsc_env mod_summary iface = + fixIO $ \details' -> do + let act hpt = addToHpt hpt (ms_mod_name mod_summary) + (HomeModInfo iface details' Nothing) + let hsc_env' = hscUpdateHPT act hsc_env + -- NB: This result is actually not that useful + -- in one-shot mode, since we're not going to do + -- any further typechecking. It's much more useful + -- in make mode, since this HMI will go into the HPT. + genModDetails hsc_env' iface + + +{- +Note [ModDetails and --make mode] +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + +An interface file consists of two parts + +* The `ModIface` which ends up getting written to disk. + The `ModIface` is a completely acyclic tree, which can be serialised + and de-serialised completely straightforwardly. The `ModIface` is + also the structure that is finger-printed for recompilation control. + +* The `ModDetails` which provides a more structured view that is suitable + for usage during compilation. The `ModDetails` is heavily cyclic: + An `Id` contains a `Type`, which mentions a `TyCon` that contains kind + that mentions other `TyCons`; the `Id` also includes an unfolding that + in turn mentions more `Id`s; And so on. + +The `ModIface` can be created from the `ModDetails` and the `ModDetails` from +a `ModIface`. + +During tidying, just before interfaces are written to disk, +the ModDetails is calculated and then converted into a ModIface (see GHC.Iface.Make.mkIface_). +Then when GHC needs to restart typechecking from a certain point it can read the +interface file, and regenerate the ModDetails from the ModIface (see GHC.IfaceToCore.typecheckIface). +The key part about the loading is that the ModDetails is regenerated lazily +from the ModIface, so that there's only a detailed in-memory representation +for declarations which are actually used from the interface. This mode is +also used when reading interface files from external packages. + +In the old --make mode implementation, the interface was written after compiling a module +but the in-memory ModDetails which was used to compute the ModIface was retained. +The result was that --make mode used much more memory than `-c` mode, because a large amount of +information about a module would be kept in the ModDetails but never used. + +The new idea is that even in `--make` mode, when there is an in-memory `ModDetails` +at hand, we re-create the `ModDetails` from the `ModIface`. Doing this means that +we only have to keep the `ModIface` decls in memory and then lazily load +detailed representations if needed. It turns out this makes a really big difference +to memory usage, halving maximum memory used in some cases. + +See !5492 and #13586 +-} + -- Runs the post-typechecking frontend (desugar and simplify). We want to -- generate most of the interface as late as possible. This gets us up-to-date -- and good unfoldings and other info in the interface file. @@ -876,7 +926,6 @@ finish summary tc_result mb_old_hash = do return HscRecomp { hscs_guts = cg_guts, hscs_mod_location = ms_location summary, - hscs_mod_details = details, hscs_partial_iface = partial_iface, hscs_old_iface_hash = mb_old_hash } |