MkIface.writeIfaceFile doesn't check GhcMode anymore. All it does
is what the name say: write an interface to disk.
I've refactored HscMain so the logic is easier to manage. That means
we can avoid running the simplifier when typechecking (: And best of
all, HscMain doesn't use GhcMode at all, anymore!

The new HscMain intro looks like this:

It's the task of the compilation proper to compile Haskell, hs-boot and
core files to either byte-code, hard-code (C, asm, Java, ect) or to
nothing at all (the module is still parsed and type-checked. This
feature is mostly used by IDE's and the likes).
Compilation can happen in either 'one-shot', 'batch', 'nothing',
or 'interactive' mode. 'One-shot' mode targets hard-code, 'batch' mode
targets hard-code, 'nothing' mode targets nothing and 'interactive' mode
targets byte-code.
The modes are kept separate because of their different types and meanings.
In 'one-shot' mode, we're only compiling a single file and can therefore
discard the new ModIface and ModDetails. This is also the reason it only
targets hard-code; compiling to byte-code or nothing doesn't make sense
when we discard the result.
'Batch' mode is like 'one-shot' except that we keep the resulting ModIface
and ModDetails. 'Batch' mode doesn't target byte-code since that require
us to return the newly compiled byte-code.
'Nothing' mode has exactly the same type as 'batch' mode but they're still
kept separate. This is because compiling to nothing is fairly special: We
don't output any interface files, we don't run the simplifier and we don't
generate any code.
'Interactive' mode is similar to 'batch' mode except that we return the
compiled byte-code together with the ModIface and ModDetails.

In the rare event that a .o file contains more than one flavour of a
[non]lazy pointers sections, resolve all of them, not just one.

I've changed the name to 'getGhcMode'. If someone changes
it back, please write an explanation above it.

None of the new code is in use yet.

The current Haskell compiler (HscMain.hscMain) isn't as typed
and as hack-free as we'd like. Here's a list of the things it
does wrong:
  * In one shot mode, it returns the new interface as _|_,
    when recompilation isn't required. It's then up to the
    users of hscMain to keep their hands off the result.
  * (Maybe ModIface) is passed around when it's known that it's
    a Just. Hey, we got a type-system, let's use it.
  * In one shot mode, the backend is returning _|_ for the
    new interface. This is done to prevent space leaks since
    we know that the result of a one shot compilation is never
    used. Again, it's up to the users of hscMain to keep their
    hands off the result.
  * It is allowed to compile a hs-boot file to bytecode even
    though that doesn't make sense (it always returns
    Nothing::Maybe CompiledByteCode).
  * Logic and grunt work is completely mixed. The frontend
    and backend keeps checking what kind of input they're handling.
    This makes it very hard to get an idea of what the functions
    actually do.
  * Extra work is performed when using a null code generator.


The new code refactors out the frontends (Haskell, Core), the
backends (Haskell, boot) and the code generators (one-shot, make,
nothing, interactive) and allows them to be combined in typesafe ways.
A one-shot compilation doesn't return new interfaces at all so we
don't need the _|_ space-leak hack. In 'make' mode (when not
targeting bytecode) the result doesn't contain
Nothing::Maybe CompiledByteCode. In interactive mode, the result
is always a CompiledByteCode. The code gens are completely separate
so compiling to Nothing doesn't perform any extra work.

DriverPipeline needs a bit of work before it can use the new
API.

When the volatile regs attached to a CmmCall is Nothing, it means
"save everything", not "save nothing".

After a long hunt I discovered that the reason that GHC.Enum.eftIntFB
was being marked as a loop-breaker was the bizare behaviour of exprFreeVars,
which returned not only the free variables of an expression but also the
free variables of RULES attached to variables occuring in the expression!

This was clearly deliberate (the comment was CoreFVs rev 1.1 in 1999) but
I've removed it; I've left the comment with further notes in case there
turns out to be a Deep Reason.

This turned out to be a lot easier than I thought.  Just moving a few
bits of -split-objs support from the build system into the compiler
was enough.  The only thing that Cabal needs to do in order to support
-split-objs now is to pass the names of the split objects rather than
the monolithic ones to 'ar'.

The NCG cannot be used in an unregisterised compiler, so there's no
point in including it.

Sadly the above patch wasn't right, because it fouls
up pre/postInlineUnconditionally.  This patch puts
things back as they were functionally, but with slightly
tidied-up code.

contributed by Neil Mitchell <ndmitchell@gmail.com>, with docs by me.

this is so that the stage1 compiler has proper support for Unicode.
Should fix these errors:

  lexical error in string/character literal at character '\8759'

when building the stage2 compiler.