DXIL-style typed pointer handling.

[maleadt]

Like #5, but similar to how the DXILWriter in LLVM proper currently works. I had high hopes for this approach, because it doesn't mutate the module, doesn't need to inject additional intrinsics or metadata. However after hacking on it for a day or two it seems like:

• this requires much more (simple, but conflict prone) changes in the bitcode writer, making it difficult to maintain this across LLVM versions;
• the approach is more fragile, as it keeps typed pointers around much more, which easily results in type mismatches in the serialized IR (see example below);
• the approach doesn't seem to handle pointers all that well, presumably because it's been developed with shader inputs (HLSL) in mind.

---

An example of the fragility:

define i64 @kernel(ptr %0) {
  %box = alloca i64, i32 0, align 8
  %ifelse = select i1 false, ptr %box, ptr %0
  %value = load i64, ptr %ifelse, align 4
  ret i64 %value
}

The pointer analysis pass deduces that %box is i64*, but %box falls back to i8*, resulting in a crash when deserializing the select with two differently-typed pointer inputs