GitLab

Merge backend-hacking-branch onto HEAD.  Yay!

wibble

avoid string gaps

Take the timestamp of an interpreted linkable from the timestamp of
the source file.  This works better when the local clock is out of
sync with the filesystem, and it's just as accurate.

Add location information to :i command

Correction to the construction of .hi filenames in dependency
generation.

Fixes sourceforge bug #978543

Merge to STABLE

Tiny perf improvement

Renamed Opt_WarnMisc to Opt_WarnDodgyImports

Improve error reporting for static flags

Documentation typos

Use IfaceDeprecs synonym

Move the definition of rawSystem into a separate file which we
#include in the places it is needed.  This is slightly better than
copying the code, since we now need it in three places
(ghc/utils/runghc is the 3rd).

wibble

DoEval is a compilation manager mode too.

MERGE TO STABLE

----------------------------------------
     1. Make primOpIsCheap do something sensible
     2. Make the state hack work better
     ----------------------------------------

1.  In March 2001, we changed primOpIsCheap to
	primOpIsCheap op = False
thereby making *no* primops seem cheap.  But this killed eta
expansion on case (x ==# y) of True -> \s -> ...
which is bad.  In particular a loop like
  doLoop n = loop 0
    where
      loop i | i == n    = return ()
             | otherwise = bar i >> loop (i+1)
allocated a closure every time round because it didn't eta expand.

The problem that made us set primOpIsCheap to False was
		let x = a +# b *# c in x +# x
where we don't want to inline x. But primopIsCheap doesn't control
that (it's exprIsDupable that does) so the problem doesn't occur
even if primOpIsCheap sometimes says 'True'.  I think that perhaps
something changed since March 2001.


2.  Consider this
	case x of
	  True  -> \(s:RealWorld) -> e
	  False -> foo
where foo has arity 1.  If we are using the "state hack" we want to
eta expand here.  This commit fixes arityType in the Var case (for foo)
to take account of foo's type.

Also add -fno-state-hack to the static flags, to allow the state hack to
be switched off.

Add :k(ind) command to ghci to find the kind of a type.
This works very, very similarly to :t(ype)

Extend the "newtype deriving" feature a little bit more
(at the request of Wolfgang Jeltsch)

Here's the example:
    class C a b
    instance C [a] Char
    newtype T = T Char deriving( C [a] )

Perfectly sensible, and no reason it should not work.
Fixing this required me to generalise the abstract syntax of
a 'deriving' item, hence the non-local effects.

Add a flag -fno-state-hack, which switches off the "state hack".

It's claims that every function over realWorldStatePrimTy is a
one-shot function.  This is pretty true in practice, and makes a big
difference.  For example, consider
	a `thenST` \ r -> ...E...
The early full laziness pass, if it doesn't know that r is one-shot
will pull out E (let's say it doesn't mention r) to give
	let lvl = E in a `thenST` \ r -> ...lvl...
When `thenST` gets inlined, we end up with
	let lvl = E in \s -> case a s of (r, s') -> ...lvl...
and we don't re-inline E.

-i on its own didn't delete the list of search paths as advertised.

Remove all known hacks in rawSystem:

  - no splitting of the program name using toArgs

  - no avoiding translate when the string already appears to be quoted

  - no avoiding translate for the command name

We now keep separate program name & args for various SysTools
programs: gcc, as, ld, mkdll.

MERGE TO STABLE

Remove -static flag from non-darwin PowerPC builds
(this was a relic from the long-dead AIX support)

Merge to STABLE.

Fix #ifdef'ed flags for FreeBSD.

comments

Add a -fno-full-laziness flag

Combined build tags get '_' between them

Add missing case to allowedWith

Experimental support for RTS-only "ways"

HEADS UP!  This changes the way that the threaded RTS is used, and
also the use of debugging RTSs:

  - We always build threaded and debugging variants of the RTS now.
    The --enable-threaded-rts configure option is ignored (and will
    be removed at some point).

  - New option:  -debug     enables the debugging RTS

  - New option:  -threaded  enables the threaded RTS.  When the threaded
    RTS is stable enough, we might make it the default.

The new options just cause a different variant of the RTS to be linked
in, and they cause one or two extra options to be enabled too.  The
implementation is via the usual ways machinery in the compiler, except
that these ways are labelled as RTS-only, and so don't require
rebuilding all the libraries too.

All of this means we can ship threaded and debugging RTSs with GHC, so
that users don't need to fetch and build a GHC source tree to use
them.

I'd like to get this functionality into 6.2.1 if possible, so please
test (I'm willing to stretch the definition of "interface change" to
accomodate this, since having a threaded RTS available without having
to build GHC will be a big win for the Visual Studio project).

---------------------------------------
  	 Record dependency on Template Haskell package
	  ---------------------------------------

An unforseen consequence of making the Template Haskell package separate
is that we need to record dependency on the package, even if no TH module
is imported.  So we carry round (another) mutable variable tcg_th_used in
the tyepchecker monad, and zap it when $(...) and [| ... |] are used.

I did a little tidy-up and documentation in ListSetOps too

New version of translate for mingw32, which correctly (allegedly)
reverses the command-line translation done by the standard C runtime
on Windows.

Make trace message more accurate

Fix build

---------------------------------------
  	 Fix the WinME/98/95 double-backslash bug
	  ---------------------------------------

		Merge to stable branch, at least
		once we are sure it's right

Backslashes should not be escaped in command-line arguments for rawSystem,
on Win32.  This only actually causes failures on WinME/98/95, and we can't
test that since we don't have it.  But this fix seems right regardless.

There are extensive new comments in libraries/base/System/Cmd.hs which
describe the issues.

error message sorting wibble

Comments only

Fix error message printing

Some small steps in the direction of making GHC useable as a library:

  - The ErrMsg type is now richer: we keep the location info and the
    PrintUnqualified separate until the message is printed out, and
    messages have a short summary and "extra info", where the extra
    info is used for things like the context info in the typechecker
    (stuff that you don't normally want to see in a more visual setting,
    where the context is obvious because you're looking at the code).

  - hscMain now takes an extra argument of type (Messages -> IO ()),
    which says what to do with the error messages.  In normal usage,
    we just pass ErrUtils.printErrorsAndWarnings, but eg. a development
    environment will want to do something different.  The direction we
    need to head in is for hscMain to *never* do any output to
    stdout/stderr except via abstractions like this.

Split Template Haskell out to its own package and update docs and tests.

Wibbles to exporting types abstractly

Add (missing?) -pgmdll option.  Remove duplicate -pgmP option.

Fix bug in genOutputFilenameFunc: the -keep-blah-files options weren't
being considered for the final phase.  Upshot: -keep-hc-files wasn't
working in --make mode (see driver060 & driver061 tests).

MERGE TO STABLE