[LLVMdev] Theoretical grounds of IR.

[Mahadevan R]

Hello all,

I've been trying to figure out some "questions" related to LLVM IR
(not specifically LLVM's, but LLVM's IR does appear to be the best on
this planet!) since some time. It'd be great if some one could give
some pointers.

Given a specific target, and a set of rules that define what is
optimal code for that target, can it be proven that the LLVM IR
provides the necessary and sufficient abstractions from which optimal
code can be generated? Is there a theoretical basis for the
constitution of the IR?

If I "raise" optimized assembly language code into LLVM IR, is it
possible (thoeretically, provably) to generate equal, or more, optimal
assembly code from it?

Wouldn't this involve some sort of formal representation of the
target's architecture and optimization rules? Is there any literature
related to this?

Thanks a lot in advance!
-Mahadevan.