Build system
------------

The location of the gcc source and objects should be passed as options to a
configure script rather than being hardwired into the Makefile.

Determination of the target triple should be moved from the Makefile to a
configure script.

The plugin revision is created from the subversion revision.  What if people
are using git etc?  Maybe it should be calculated in a configure script, but
since that might not get run often perhaps the Makefile is the best place.

Target subdirectories should have their own Makefiles, instead of assuming
that there's only one source file and that it's called llvm-target.cpp.

Currently the target directory (eg: i386) is calculated from the target triple
(eg: x86_64-unknown-linux-gnu) using gcc's config.gcc script.  This should be
done from a configure script, rather from the Makefile using the get_arch_dir
wrapper.

The Makefile needs to know the LLVM target name (eg: x86) in order to pass it
to llvm-config so as to get the libraries to link with for target codegen.
This should be calculated in a configure script (presumably by a small program
that uses the LLVM triple facilities to calculate it).  If we know this, then
we might as well rename target directories like i386/ to x86/ instead, and
eliminate use of gcc's config.gcc script.  This also means that LLVM_TARGET_NAME
could be defined from the Makefile rather than being specified in llvm-target.h.
Maybe LLVM_TARGET_INTRINSIC_PREFIX could go too.

Teach the build system that the plugin needs to be rebuilt if any of the bits of
LLVM/gcc it depends on changes.


Optimizations
-------------

After outputting global variables, maybe they can be deleted or marked somehow
(eg: TREE_ASM_WRITTEN) so that GCC does not output them (such output gets sent
to /dev/null, but it would be more efficient to teach GCC to not produce any in
the first place).  Investigate.

Consider having DECL_LLVM and friends store values for trees local to a
function in a separate map that can be cleared once the function has been
emitted.  Determining which trees are local (LABEL_DECL, VAR_DECL but not
static etc) should be pretty quick, so the cost of routing to the right map
should be insignificant, while having a much smaller map for globals could
be a win.

Consider using separate caches for types and globals.

Thanks to gimple being in SSA form, many local variables are not used.
However they are still in the local_decls list, so we output them all,
pointlessly, wasting time and memory.  Consider emitting local variables
on demand instead.


Correctness
-----------

If an ssa name refers to a global (can this happen), the SSANames map might
need to be updated if the target is altered by changeLLVMConstant.

An ssa name can be a complex number, causing the plugin to crash.  Maybe should
consider complex numbers to be scalars rather than aggregates.  Would this get
in the way of sroa?

If the initializer for a static variable contains the address of a label then
we crash.


Features
--------

Output proper debug info rather than throwing most of it away.

Add support for exception handling.