GitLab

This replaces some complicated locking schemes with message-passing
in the implementation of throwTo. The benefits are

 - previously it was impossible to guarantee that a throwTo from
   a thread running on one CPU to a thread running on another CPU
   would be noticed, and we had to rely on the GC to pick up these
   forgotten exceptions. This no longer happens.

 - the locking regime is simpler (though the code is about the same
   size)

 - threads can be unblocked from a blocked_exceptions queue without
   having to traverse the whole queue now.  It's a rare case, but
   replaces an O(n) operation with an O(1).

 - generally we move in the direction of sharing less between
   Capabilities (aka HECs), which will become important with other
   changes we have planned.

Also in this patch I replaced several STM-specific closure types with
a generic MUT_PRIM closure type, which allowed a lot of code in the GC
and other places to go away, hence the line-count reduction.  The
message-passing changes resulted in about a net zero line-count
difference.

We were getting a bogus claim that a newtype "data constructor" was
unused.  The fix is easy, although I had to add a field to the constructor
TcEnv.NewTypeDerived

See Note [Newtype deriving and unused constructors] in TcDeriv

While I was looking at the rule binders generated in DsBinds for specialise pragmas,
I also looked at Specialise.  It too was "cloning" the dictionary binders including
their IdInfo. In this case they should not have any, but its seems better to make
them completely fresh rather than substitute in existing (albeit non-existent) IdInfo.

When DsBinds deals with a SPECIALISE pragma, it makes up the binders
for a RULE. These binders should be very vanilla: no IdInfo of any sort.
But the way it was before they could have DFun pragmas, which led to
Bad Joss downstream.  (Actually to cause a downstream *error* was itself
a bug, which I've fixed, but it's clearly wrong for them to be DFuns!)

isClosedUnfolding should say False for DFUnUnfolding!

We were getting "INLINEALWAYS" printed out here and there.
Now there are always brackets around the activation, thus "INLINE[ALWAYS]"

The idea is that this leaves Tasks and OSThread in one-to-one
correspondence.  The part of a Task that represents a call into
Haskell from C is split into a separate struct InCall, pointed to by
the Task and the TSO bound to it.  A given OSThread/Task thus always
uses the same mutex and condition variable, rather than getting a new
one for each callback.  Conceptually it is simpler, although there are
more types and indirections in a few places now.

This improves callback performance by removing some of the locks that
we had to take when making in-calls.  Now we also keep the current Task
in a thread-local variable if supported by the OS and gcc (currently
only Linux).

This fixes occasional failures of ffi002(threaded1) on a loaded
machine.

  (and similarly for PS and PDF)

The previous setup nuked the rules for making the documentation when
BUILD_DOCBOOK_HTML=NO.  This meant that "make html" didn't work.

There isn't any reason to nuke the rules (so far as Simon and I know).

preInlineUnconditionally was, in effect, nuking an INLINE pragma, with
very bad effect on runtime in this program.  Fortunately the fix is
very simple.

See Note [InlineRule and preInlineUnconditionally] in SimplUtils.

This one was bigger than I anticipated!  The problem was that were
were gathering the binders from a pattern before renaming -- but with
record wild-cards we don't know what variables are bound by C {..}
until after the renamer has filled in the "..".

So this patch does the following

* Change all the collect-X-Binders functions in HsUtils so that
  they expect to only be called *after* renaming.  That means they
  don't need to return [Located id] but just [id].  Which turned out
  to be a very worthwhile simplification all by itself.

* Refactor the renamer, and in ptic RnExpr.rnStmt, so that it
  doesn't need to use collectLStmtsBinders on pre-renamed Stmts.

* This in turn required me to understand how GroupStmt and
  TransformStmts were renamed.  Quite fiddly. I rewrote most of it;
  result is much shorter.

* In doing so I flattened HsExpr.GroupByClause into its parent
  GroupStmt, with trivial knock-on effects in other files.

Blargh.

a) Even with -dsuppress-uniques, don't suppress them when outputing
   code, else the assembler falls over bleating

b) Do suppress uniques in names generated by TH.  It's a bit grungy
   to do this: see Note [Suppressing uniques in OccNames].  But
   it's only needed for test de-wobblification so the grunge isn't
   really important.

This makes things more consistent, and avoids a strange "Prelude is
not loaded" error when there is a Prelude.hs on the search path.

Now not only is if/then a bracketting pair, but then/else is too

If enabled, these accept more layout, but give warnings

It used to break on
    let {x = 'a'} in x
as the 'in' token would keep closing contexts looking for an implicit
'let' layout.

This just ensures that an unsafe coercion is as localised as possible.
For example, instead of
    UnsafeCo (Int -> t1) (Int -> t2)
use 
    Int -> UnsafeCo t1 t2

If we can't derive a type, but it's a reasonable possibility that
newtype deriving would do the job, suggest it. 

A little refactoring too, moving non_iso_class to top level,
and putting it with std_class_via_iso.

If we find ourselves making up an AnyTyCon of kind '??', say, 
then default it to liftedTypeKind.  And similarly for any sub-kind
of LiftedTypeKind.

This is just a tidy-up.

The panic message in ifaceTyConName on IfaceAnyTc 
called the pretty-printer, called ifaceTyConName again,
which caused an infinite regress.  Stupid.

See 
 * Note [The Name of an IfaceAnyTc], and 
 * The Outputable instance for IfaceTyCon