Lots of refactoring. In particular I have now combined
TansformStmt and GroupStmt into a single constructor TransStmt.
This gives lots of useful code sharing.

This is the work of Nils Schweinsberg <mail@n-sch.de>

It adds the language extension -XMonadComprehensions, which
generalises list comprehension syntax [ e | x <- xs] to work over
arbitrary monads.

This allows tools like Linux's perf events to display symbol names for
CMM primops.

Signed-off-by: Edward Z. Yang <ezyang@mit.edu>

We achieve this by splitting up instruction selection for case
switches into two parts: the actual code generation, and the
generation of the accompanying jump table.  With this scheme,
the jump fixup code can modify the contents of the jump table
stored within the JMP_TBL (or BCTL) instruction, before the
actual data section is created.

SPARC and PPC patches are untested; they might not work!

Signed-off-by: Edward Z. Yang <ezyang@mit.edu>

Signed-off-by: Edward Z. Yang <ezyang@mit.edu>

This patch fixes silliness where the SP pointer is continually
bumped up and down.

Signed-off-by: Edward Z. Yang <ezyang@mit.edu>

Run a CSE pass after vectorisation. This cleans up common sub-expressions introduced by the vectoriser. It has no effect unless -fvectorise is enabled.

Signed-off-by: Edward Z. Yang <ezyang@mit.edu>

This makes the following statements valid C--, which uses the
block ID directly rather than generating an anonymous block:

    if (expr) goto foo;
    switch expr {
        case 0: goto foo;
        case 1: goto bar;
    }

Signed-off-by: Edward Z. Yang <ezyang@mit.edu>

We were using the Typeable.hs macros, but for no good reason as
far as I can tell.

I also made the -ffloat-all-lams flag not take an argument, as it just
ignored it.

Makes it consistent with the existing "C compiler flags"
field.

This reverts commit b1f453e1.
It caused dynHelloWorld(dyn) to fail.

They were hardcoded in Config.hs

This allows tools like Linux's perf events to display symbol names for
CMM primops.

The pgm* fields are already in Settings

Stops us having to temporarily have a panic in the DynFlags.
We still need a panic in the DynFlags used for the top-level
error handler, though.

This just a refactoring, removing dead code

Some out of date code in TcErrors was suppressing an error
report -- so some type-incorrect code leaked out and confused
a later bit of the compiler.