Currently (GHC 6.10) we report three different things:
warnUnusedModules: import M, where nothing is used from M
warnUnusedImports: import M(f), where f is unused, and M doesn't fall under warnUnusedModules
warnDuplicateImports: import M + import M(f), even when f is used complain about duplicate import of f
The hard bit is to specify what the warning should do.
Consider these examples, where Foo exports x and y, and FooPlus
re-exports all of Foo, plus z:
module X0 where module X1 where import Foo import Foo import Foo( x ) import Foo( x ) bar = x bar = x+y module X2 where module X3 where import Foo( x, y ) import Foo( x, y ) import Foo( x ) import Foo( x ) bar = x bar = x + y module X4 where module X5 where import Foo( x, y ) import Foo( x, y ) as Bar import Foo( x, y ) import Foo( x, y ) bar = x + y bar = x + Bar.y module X6 where module X7 where import Foo( x, y ) as Bar import FooPlus(x,y) import Foo( x, y ) import FooPlus(y,z) bar = Foo.x + Bar.y import FooPlus(z,x) bar = (x,y,z) module X8 import Control.Monad import Control.Monad.State import Control.Monad.Reader -- NB : Control.Monad.State re-exports all of Control.Monad -- so the first decl is actually redundant
Which import is redudant, in each case?
Also: we might warn if you import the same module more than once, and the
imports can be combined (ie they have the same 'qualified' and 'as'
module Y1 where import Foo(x) import Foo(y) bar = (x,y)
Here both are used, but we might want to suggest combining them.
We can at least agree on this:
If the warning suggests that an import can be omitted, and you omit it,
the program should still compile.
It's not worth trying to be too subtle. The 90% case is very simple.
Say that an import-item is either an entire import-all decl (eg import Foo),
or a particular item in an import list (eg import Foo( ..., x, ...)).
The general idea is that for each use of an imported name, we will attribute
that use to one (or possibly more) import-items. Then, any import items with no
uses attributed to them are unused, and are warned about.
For every RdrName in the program text, find all the import-items that brought it into scope. The lookup mechanism on RdrNames already takes account of whether the RdrName was qualified, and which imports have the right qualification etc, so this step is very easy.
Choose one of these, the "chosen import-item", and mark it "used".
Now bleat about any import-items that are unused. For a decl
import Foo(x,y), if both the x and y items are unused, it'd be better
to bleant about the entire decl rather than the individual items.
The import-item choosing step 2 implies that there is a total order on
import-items. We say import-item A dominates import-item B if we choose
A over B. Here is one possible dominance relationship:
(a) import Foo dominates import qualified Foo, regardless of all-or-none.
(b) import Foo dominates import Foo(x).
(c) Otherwise choose the textually first one.
Rationale for (a). Consider
import qualified M -- Import #1 import M( x ) -- Import #2 foo = M.x + x
The unqualified x can only come from import #2. The qualified M.x
could come from either, but bestImport picks import #2, because it is
more likely to be useful in other imports, as indeed it is in this
case (see Trac #5211 (closed) for a concrete example).
The algorithm chooses exactly one import-item in step 2. It would
also be sound to choose more than one if there was a tie, but then completely-duplicate
imports might not be reported.
Note that if we have an import item import Foo (Bar(bar)), then
it's marked as used if either Bar or bar are used. We could have yet finer
resolution and report even unused sub-items.
We should retain the special case of not warning about import Foo (), which implies "instance declarations only".
We want to collect the set of all RdrNames that are mentioned in the
program. We must collect RdrNames not Names:
import Foo( x ) as Bar import Foo( x ) q = (Foo.x, Bar.x)
Here both imports are required, but you can only tell that by
seeing the RdrNames, not by knowing that the name x is used.
I think that all lookups go through either, RnEnv.lookupGreRn_maybe or RnEnv.lookup_sub_bndr.
So in RnEnv.lookupGreRn_maybe, if (gre_prov gre) is (Imported _),
and in RnEnv.lookup_sub_bndr,
put rdr_name in a new
tcg_used_rdrnames :: TcRef (Set RdrName)
in TcGblEnv. All the tcg_used_rdrnames are in scope; if not,
we report an error and do not add it to tcg_used_rdrnames.
Any particular (in-scope) used RdrName is bought into scope by
one or more RdrName.ImportSpec's. You can find these ImportSpecs
in the GRE returned by the lookup.
The unit of "unused import" reporting is one of these ImportSpecs.
Suppose that rn is a used, imported RdrName, and iss is
the [ImportSpecs] that brought it into scope. Then, to a first
approximation all the iss are counted 'used'.
We can compare ImportSpecs for equality by their SrcSpans.
In TcRnDriver.tcRnImports, save import_decls in a new
tcg_rn_rdr_imports :: Maybe [LImportDecl RdrName]
The algorithm for deciding which imports have been used is based around this datatype:
data ImportInfo = ImportInfo SrcSpan SDoc (Maybe ModuleName) -- The effective module name [RdrName] -- The names the import provides Bool -- Has it been used yet? [ImportInfo] -- Child import infos
We convert import declarations into trees of ImportInfos, e.g.
import Foo (a, D(c1, c2))
becomes (only the SDoc and [RdrName] fields are given, as that's the interesting bit)
If a node in the tree is marked as used, then so are all nodes above it. For example, given the tree
a use of "D" marks both the first and third lines as used.
When we come to giving warnings, if a node is unused then we warn about it, and do not descend into the rest of that subtree, as the node we warn about subsumes its children. If the node is marked as used then we descend, looking to see if any of its children are unused.
Here are how some example imports map to trees of ImportInfo, assuming Foo exports a, b, D(c1, c2).
These trees are built by RnNames.mkImportInfo. In RnNames.warnUnusedImportDecls we make two lists of ImportInfos; one list contains all the explicit imports, e.g.
import Foo (a, b)
and the other contains the implicit imports, e.g.
import Fooimport Foo hiding (a, b)
Then RnNames.markUsages is called for each RdrName that was used in the program. The current implementation marks all explicit import as used unless there are no such imports, in which case it marks all implicit imports as used. A small tweak to markUsages would allow it to mark only the first import it finds as used.
As well as the RdrNames used in the source, we also need to mark as used the names that are exported. We first call RnNames.expandExports to expand D(..) into D(c1, c2), and then call RnNames.markExportUsages. Normally this just marks the RdrNames as used in the same way that uses in the module body are handled, but it is also possible for an entire module to be "used", if module Foo is in the export list. In this case RnNames.markModuleUsed does the hard work, marking every module imported with that name as used.