Integrating fine-grained timing constraints into LLVM

This talk covers a problem we are just beginning to tackle of integrating fine-grained timing constraints into LLVM. Code written for real-time tasks often comes with worst case timing constraints on paths within the program (usually between I/O operations). These paths can be quite short or can cross across basic blocks and function boundaries. Unfortunately, the optimizations within LLVM are not aware of these constraints and code motion or control flow optimizations can move calculations into a critical path quite easily. I'll cover the experiences we've had with the optimizations in the compiler and go over the initial ideas we have to tackle the problem.