Add guaranteed TCO

[qdbplang]

Currently, we compile with clang -O2. Unfortunately, while most(all?) of the time clang will optimize tail calls, it is not a guarantee. We should alter the produced LLVM IR to force TCO.

[dghosef]

One potential way is to do:

• emit llvm ir(if o0, make sure we provide the -Xclang -disable-O0-optnone option
• Manually modify the llvm ir to use the tailcc calling convention for methods. Essentially
  • Every generated method should use the tailcc calling convention
  • Make all the invoke_<n> be force inlined and all their invocations of code be tailcalled w/ tailcc
• Run opt with --passes=tailcallelim

Something that will potentially simplify things is if we put everything into one file.

We should implement this with qdbp :)

[dghosef]

Mark all invoke fns as optnone and always inline

Insert inline asm as annotation. Mark all definitions as tailcc

Run tail call elimination pass and remove optnone annotations

[dghosef]

You can just tailcc mark all metbods and method calls. No need for inline asm