Commit 8f63ba30 authored by Edward Z. Yang's avatar Edward Z. Yang

Compute boot-defined TyCon names from ModIface.

Summary:
Three things in this commit:

    1. Get rid of sb_ids; we are not going to use them
    to avoid infinite unfoldings in hs-boot files.

    2. Compute sb_tcs from ModIface rather than ModDetails.
    This means that the typechecker can look at this field
    without forcing the boot ModDetails, which would be
    bad if the ModDetails is not available yet (due to
    knot tying.)

    3. A big honking comment explaining what is going on
    here.
Signed-off-by: default avatarEdward Z. Yang <ezyang@cs.stanford.edu>

Reviewers: simonpj, austin, bgamari

Subscribers: thomie

Differential Revision: https://phabricator.haskell.org/D2380
parent 9d62f0d1
......@@ -191,7 +191,7 @@ tcHiBootIface hsc_src mod
then do { hpt <- getHpt
; case lookupHpt hpt (moduleName mod) of
Just info | mi_boot (hm_iface info)
-> return (mkSelfBootInfo (hm_details info))
-> mkSelfBootInfo (hm_iface info) (hm_details info)
_ -> return NoSelfBoot }
else do
......@@ -205,7 +205,7 @@ tcHiBootIface hsc_src mod
; case read_result of {
Succeeded (iface, _path) -> do { tc_iface <- typecheckIface iface
; return (mkSelfBootInfo tc_iface) } ;
; mkSelfBootInfo iface tc_iface } ;
Failed err ->
-- There was no hi-boot file. But if there is circularity in
......@@ -237,13 +237,28 @@ tcHiBootIface hsc_src mod
quotes (ppr mod) <> colon) 4 err
mkSelfBootInfo :: ModDetails -> SelfBootInfo
mkSelfBootInfo mds
= SelfBoot { sb_mds = mds
, sb_tcs = mkNameSet (map tyConName (typeEnvTyCons iface_env))
, sb_ids = mkNameSet (map idName (typeEnvIds iface_env)) }
mkSelfBootInfo :: ModIface -> ModDetails -> TcRn SelfBootInfo
mkSelfBootInfo iface mds
= do -- NB: This is computed DIRECTLY from the ModIface rather
-- than from the ModDetails, so that we can query 'sb_tcs'
-- WITHOUT forcing the contents of the interface.
tcs <- mapM (lookupOrig (mi_module iface) . ifName)
. filter isIfaceTyCon
. map snd
$ mi_decls iface
return $ SelfBoot { sb_mds = mds
, sb_tcs = mkNameSet tcs }
where
iface_env = md_types mds
-- | Retuerns @True@ if, when you call 'tcIfaceDecl' on
-- this 'IfaceDecl', an ATyCon would be returned.
-- NB: This code assumes that a TyCon cannot be implicit.
isIfaceTyCon IfaceId{} = False
isIfaceTyCon IfaceData{} = True
isIfaceTyCon IfaceSynonym{} = True
isIfaceTyCon IfaceFamily{} = True
isIfaceTyCon IfaceClass{} = True
isIfaceTyCon IfaceAxiom{} = False
isIfaceTyCon IfacePatSyn{} = False
{-
************************************************************************
......
......@@ -1156,7 +1156,7 @@ Note [Dependency analysis of type, class, and instance decls]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
A TyClGroup represents a strongly connected components of
type/class/instance decls, together with the role annotations for the
type/class declarations. The renamer uses strongyly connected
type/class declarations. The renamer uses strongly connected
comoponent analysis to build these groups. We do this for a number of
reasons:
......@@ -1368,7 +1368,9 @@ getParent rdr_env n
{- Note [Extra dependencies from .hs-boot files]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider the following case:
This is a long story, so buckle in.
**Dependencies via hs-boot files are not obvious.** Consider the following case:
A.hs-boot
module A where
......@@ -1377,24 +1379,61 @@ A.hs-boot
B.hs
module B where
import {-# SOURCE #-} A
type DisguisedA1 = A1
data B1 = B1 DisguisedA1
type B1 = A1
A.hs
module A where
import B
data A2 = A2 A1
data A1 = A1 B1
Here A1 is really recursive (via B1), but we won't see that easily when
doing dependency analysis when compiling A.hs
data A2 = MkA2 B1
data A1 = MkA1 A2
Here A2 is really recursive (via B1), but we won't see that easily when
doing dependency analysis when compiling A.hs. When we look at A2,
we see that its free variables are simply B1, but without (recursively) digging
into the definition of B1 will we see that it actually refers to A1 via an
hs-boot file.
**Recursive declarations, even those broken by an hs-boot file, need to
be type-checked together.** Whenever we refer to a declaration via
an hs-boot file, we must be careful not to force the TyThing too early:
ala Note [Tying the knot] if we force the TyThing before we have
defined it ourselves in the local type environment, GHC will error.
Conservatively, then, it would make sense that we to typecheck A1
and A2 from the previous example together, because the two types are
truly mutually recursive through B1.
If we are being clever, we might observe that while kind-checking
A2, we don't actually need to force the TyThing for A1: B1
independently records its kind, so there is no need to go "deeper".
But then we are in an uncomfortable situation where we have
constructed a TyThing for A2 before we have checked A1, and we
have to be absolutely certain we don't force it too deeply until
we get around to kind checking A1, which could be for a very long
time.
Indeed, with datatype promotion, we may very well need to look
at the type of MkA2 before we have kind-checked A1: consider,
data T = MkT (Proxy 'MkA2)
To promote MkA2, we need to lift its type to the kind level.
We never tested this, but it seems likely A1 would get poked
at this point.
**Here's what we do instead.** So it is expedient for us to
make sure A1 and A2 are kind checked together in a loop.
To ensure that our dependency analysis can catch this,
we add a dependency:
To handle this problem, we add a dependency
- from every local declaration
- to everything that comes from this module's .hs-boot file.
In this case, we'll ad and edges
- from A2 to A1 (but that edge is there already)
- from A1 to A1 (which is new)
- to everything that comes from this module's .hs-boot file
(this is gotten from sb_tcs in the SelfBootInfo).
In this case, we'll add an edges
- from A1 to A2 (but that edge is there already)
- from A2 to A1 (which is new)
Well, not quite *every* declaration. Imagine module A
above had another datatype declaration:
......@@ -1405,7 +1444,15 @@ Even though A3 has a dependency (on Int), all its dependencies are from things
that live on other packages. Since we don't have mutual dependencies across
packages, it is safe not to add the dependencies on the .hs-boot stuff to A2.
Hence function Name.thisPackageImport.
Hence function nameIsHomePackageImport.
Note that this is fairly conservative: it essentially implies that
EVERY type declaration in this modules hs-boot file will be kind-checked
together in one giant loop (and furthermore makes every other type
in the module depend on this loop). This is perhaps less than ideal, because
the larger a recursive group, the less polymorphism available (we
cannot infer a type to be polymorphically instantiated while we
are inferring its kind), but no one has hollered about this (yet!)
-}
addBootDeps :: [(LTyClDecl Name, FreeVars)] -> RnM [(LTyClDecl Name, FreeVars)]
......
......@@ -624,10 +624,10 @@ data SelfBootInfo
= NoSelfBoot -- No corresponding hi-boot file
| SelfBoot
{ sb_mds :: ModDetails -- There was a hi-boot file,
, sb_tcs :: NameSet -- defining these TyCons,
, sb_ids :: NameSet } -- and these Ids
-- We need this info to compute a safe approximation to
-- recursive loops, to avoid infinite inlinings
, sb_tcs :: NameSet } -- defining these TyCons,
-- What is sb_tcs used for? See Note [Extra dependencies from .hs-boot files]
-- in RnSource
{- Note [Tracking unused binding and imports]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment