ref: 3f71c019bc3fd13360c650cd69e5e211660db158 Carth/src d---------
Delay StartNotDefined error until after typechecking
Make Pair a builtin datatype
Fix fun-match globals not allowed to be recursive

;; This was and should be allowed
(define (foo x) (foo x))

;; This was and should be disallowed
(define foo foo)

;; This was disallowed, but should be allowed
(define foo
    ((case x
           (foo (deref x))))))
Disallow []. Add "case" keywords to match-expression to read better
Add basic import system. Basically C-style include
Refactor Codegen. Separate module for ABI-related stuff

Also had to separate a module for the Gen-monad, as Haskell can't have
cyclical imports.
Allow integer and boolean literals in patterns

Could exploit the existing Match.PCon constructor to represent not
just datatypes, but these other enumerables as well. Didn't have to
make many additions/modifications at all -- quite nice!
Verify LLVM module before compiling to catch errors

Of course, these are internal compiler errors, but they're better to
catch in `verify` for debugging purposes. Previously, LLVM would just
segfault when processing a malformed module.
Fix incorrect from-label used in phi-node of decision switch
Fix bools not truncated from i8 to i1 for condbr
Improve gen of str lits

Instead of just defining a global var for the raw byte array and
create a Str struct at every "call"site, define 2 global vars -- one
for the raw bytes, and one for the Str structs defined in terms of it.
Scale datatype tag size by n of variants. 1 variant => no tag

Types with only 1 variant have no tag. Types with <=2^n variants have
n-bit tag.
Fix struct member padding set to alignment when remainder = 0
Fix some harmless warnings
Get rid of extractvalueFromNamed, use lookupDataType in extractvalue
Fix wrong size of bools. Was i1, C ABI says i8.

Not sure exactly what happened before, but in practice, bools
often/always had the opposite value.
Impl sizeof ourselves instead of relying on messy getTypeAllocSize

The problem with the earlier approach was that we had to pull in the
IO-related EncodeAST monad in our Gen stack, and the implementation to
call getTypeAllocSize was kind of messy, as llvm-hs didn't expose it
other than in an Internal module.

EncodeAST in Gen (and Gen') prevented use of `mfix`, which I needed.

Also note that we had to change Map.Strict to Map, as, of course, we
can't have recursive bindings in the strict Map.
Move some Pretty instances from module Codegen to Misc
Pass `byval` when passing by reference

So instead of

  declare @foo(%Foo* %x)

  extern fn foo(x: &Foo) { ... }

we do

  declare @foo(%Foo* byval %x)

  extern fn foo(x: Foo) { ... }

It just seemed like a nice thing to do, especially now that
we (mostly?) conform to C calling convention / System V ABI.
Make passByRef more exhaustive and correct

I think it's pretty close to System V ABI spec now.

Sadly, access to env of datatypes is needed to do lookup for
NamedTypeReference, and this had something of a chain reaction. Not
only passByRef became Gen', but also toLlvmType! This had some effect
throughout the module.