GitLab

Add comments

--------------------------
	Cleanup in DataCon
	--------------------------

	DO NOT merge with stable

The dataConRepStrictness call used to reuturn a [Demand],
but that's a bit misleading.  In particular, consider a  strict
constructor

	data Foo = MkFoo ![Int]

Then the wrapper MkFoo is strict, but the worker $wMkFoo is not.

	MkFoo x = case x of { DEFAULT -> $wMkFoo x }

Nevertheless, when we pattern-match on $wMkFoo we will surely
find an evaluated component to the data structure, and that is
what dataConRepStrictness reports, and that's how it is used
in Simplify.

Solution: make dataConRepStrictness return [StrictnessMark]
not [Demand]. A small matter really.

oops: hPutStr wasn't flushing a line-buffered handle properly at the
end of a line.

MERGE TO STABLE

Make rotate fns work properly when rotate count is a multiple of the
word size.  This fixes sparc failures in
ghc-regress/numeric/should_run/arith011.  Also fix some
copy-and-paste-o-s.Killed by signal 2.

MERGE TO STABLE

Back out the change to remove Ord as a superclass of Ix; the revised
Haskell 98 report will no longer have this change.

More error-message grammar

merge (ghc-5-02-branch --> HEAD):

Bugfix: there was an implicit assumption that the list of slots passed
to freeStackSlots was already sorted, whereas in fact this wasn't the
case for at least one call.  Now we explicitly sort the list in
freeStackSlots, removing the hidden assumption.

The symptoms of this bug include crashes (perhaps the "AsmCodeGen"
crash), and a tendency to grow the stack a lot when let-no-escapes are
involved (because the bug caused fragmentation of the stack free list,
so we weren't re-using free slots properly).
  1.17.2.1  +3 -2      fptools/ghc/compiler/codeGen/CgStackery.lhs


ASSERT that the list of stack slots we calculate in buildLivenessMask
is sorted, because we rely on that property later.
  1.38.2.1  +5 -6      fptools/ghc/compiler/codeGen/CgBindery.lhs

Redo half-arsed hacks to do with BSS symbols on ELF.  Hopefully get it
right this time.

MERGE TO STABLE

Do the ifdefery properly so it works on Unix.

Updated to new directory structure (no extra-bin), and with all the right
binaries in bin.

Make hsc2hs find template-hsc.h on Windows. This involves putting Main.hs
through cpp; to avoid the string gap problem preprocessing
__GLASGOW_HASKELL__ into its value inside strings, make judicious use of ++
(hope that's OK, Marcin; there seemed very little point in putting it all in
KludgedSystem this time).

Fix comment.

Fix creating HSstd.o.

Don't exclude PrelMain.o from libHSstd.a.

hasktags is now built as part of the standard build process.

--------------------------
	Strictness of blockAsynch
	--------------------------

	MERGE WITH STABLE BRANCH

If we're going to supply strictness info for blockAsynchExceptions#,
it should match its arity (as other State# transformers do).

-----------------------
	More demand-info fixes
	-----------------------

	MERGE WITH STABLE BRANCH

There are a handful of functions in IdInfo that zap the
demand-info stored in an Id.  Alas, they were zapping the
*old* demand-info not the new one.  (The old one is still
there for comparison purposes.)  So we were getting some
spurious warnings and (more seriously) potentially some
incorrect strictness.  Easily fixed though.

Sync the documentation with the extra-bin removal.

oops, revert accidental commit

good riddance

Remove the "extra-bin" subdirectory from $(libexecdir), since there
are too many assumptions in the tree that $(libexecdir) == $(libdir)
(at the moment, binary dists are fairly well broken).

Reuben has promised to track the change in the Windows distribution.

Fixed a bug in TICKY_TICKY profiling.

Previously, the TICK_ENT_DIRECT event was logged before the heap/stack
check was done. As a result, if the check failed, the TICK_ENT_DIRECT
event would be logged a second time, causing TICKY_TICKY to give innacurate
numbers.

This patch shouldn't have any affect on non-ticky compilation.

Also changed the modified bit of code to use "do" notation, and so look a bit
neater.

Fix ungramatical error message

-----------------------------------
	Fix a strictness bug in the simplifier
	-----------------------------------

This one has been there a long time, but hasn't bitten till
now.  We should never float a let that is marked "sure to be
evaluated" out of a let.  It shouldn't happen, and there was
a warning to check, but the warning cried 'wolf' too often, so
we have generally ignored it. But the wolf called for supper,
when compiling spectral/expert with profiling on.

The fix is simple too:
	* use exprIsValue not exprIsCheap as the test
	* move the warning, so it doesn't cry wolf

Documentation with Simplify.simplRhs.

On the way, I'm going to conmmit a change in the same module,
which keeps unfolding info on lambda-bound variables.  This
improves the elimination of cases when the wrapper does the
'seq' -- then the worker gets to know that the arg is evaluated.

Make the binary distribution configure recognize alphaev56-dec-osf5.
MERGE TO STABLE.

Remove ugly special casing - if you want to turn off the
default rule for GHCI_LIBRARY, set DONT_WANT_STD_GHCI_LIB_RULE
to YES in your Makefile prior to including $(TOP)/mk/target.mk

ghc/lib/std and hslibs/win32 both do this now on mingw32.

Make isStrictDmd give the right answer

Make dmdFix not loop forever

Fix compilation with -DDEBUG.

MERGE TO STABLE

Improve the error message for duplicate or invalid binders
in a binding group.

	[Consequence: rnfail034 should not be an expected failure
		      any more]

	MERGE WITH STABLE BRANCH

----------------------------------------
	Make dict funs and default methods
	into LocalIds only at their binding site
	----------------------------------------
        [part of 3 related commits]

There's a long comment about this with MkId.mkDefaultMethodId,
which I reproduce below.

While I was at it, I renamed setIdNoDiscard to setIdLocalExported.
Which is hardly an improvement, I'm afraid.  This renaming touches
	Var.lhs, Id.lhs, SimplCore.lhs
in a trivial way.

	---------------------

Dict funs and default methods are *not* ImplicitIds.  Their definition
involves user-written code, so we can't figure out their strictness etc
based on fixed info, as we can for constructors and record selectors (say).

We build them as GlobalIds, but when in the module where they are
bound, we turn the Id at the *binding site* into an exported LocalId.
This ensures that they are taken to account by free-variable finding
and dependency analysis (e.g. CoreFVs.exprFreeVars).   The simplifier
will propagate the LocalId to all occurrence sites.

Why shouldn't they be bound as GlobalIds?  Because, in particular, if
they are globals, the specialiser floats dict uses above their defns,
which prevents good simplifications happening.  Also the strictness
analyser treats a occurrence of a GlobalId as imported and assumes it
contains strictness in its IdInfo, which isn't true if the thing is
bound in the same module as the occurrence.

It's OK for dfuns to be LocalIds, because we form the instance-env to
pass on to the next module (md_insts) in CoreTidy, afer tidying
and globalising the top-level Ids.

BUT make sure they are *exported* LocalIds (setIdLocalExported) so
that they aren't discarded by the occurrence analyser.

-----------------------------------
	Pin on accurate strictness info for
	record and dictionary selectors
	-----------------------------------
        [part of 3 related commits]

This fixes a long-standing infelicity.  Sometimes selectors aren't
inlined until after strictness analysis, so if we don't have decent
strictness info on them we get bad strictness results.

For record selectors, the unboxing-strict-fields stuff makes it hard
to figurwe out the correct strictness, so we just invoke the demand
analyser to work it out.

------------------------
	Fix the demand analyser
	------------------------

A spiffy new domain for demands, and definitions for lub/both
which are actually monotonic.   Quite a bit of related jiggling
around.

One of the original motivations was to do with functions like:

	sum n []     = n
	sum n (x:xs) = sum (n+x) xs

Even though n is returned boxed from the first case, we don't want
to get strictness
	S(L)V -> T
because that means we pass the box for n, and that is TERRIBLE.
So the new version errs on the side of unboxing, more like the forwards
analyser, and only passes the box if it is *definitely* needed, rather
than if it *may* be needed.

Omit unnecessary import

-------------------
	Newtypes and ccalls
	     [addendum]
	-------------------

	MERGE WITH STABLE BRANCH

I accidentally omitted these two wibbles from my previous commit.
I've added PrelNames.unitTyConKey, and used it in TcType and DsCCall.

-------------------
	Newtypes and ccalls
	-------------------

	MERGE WITH STABLE BRANCH

Yet another bit of newtype-squashing that hadn't been
synced with reality.  In desugaring ccalls, we can still
see newtypes, if they are recursive, and we must generate
appropriate coerces.

Fixes a bug in cg011.

Fix the code to *really* do the following:

On the Alpha we can only handle <= 4 integer arguments with foreign export
dynamic.  Following the example of the corresponding Sparc hack, we detect
when we're being asked to do something we can't and refuse.

MERGE TO STABLE BRANCH

document -fext-core flag and set pointers to ext-core docs/tools (MERGE to STABLE pleeeeeease.)

Import wibbles

Change a couple more Int32s into CTimes, which they are really.