Running the ACCEPT passes currently requires invoking opt or whatever with the -O1 optimization level, which wastes time since we don't need all those other optimizations at this point. It should instead be possible to run ACCEPT optimizations in isolation.
However, we can't do this now since it seems to be impossible to cleanly require an analysis pass from a FunctionPass. That is, using addRequired<>() from the getAnalysisUsage method in a FunctionPass leads to an assertion failure along the lines of "Unable to schedule…". As far as I can tell this is by design. It may be solved in later LLVM versions, which have recently undergone a complete refactor of the pass manager system. Or we could move to ModulePasses.
Running the ACCEPT passes currently requires invoking
optor whatever with the-O1optimization level, which wastes time since we don't need all those other optimizations at this point. It should instead be possible to run ACCEPT optimizations in isolation.However, we can't do this now since it seems to be impossible to cleanly require an analysis pass from a FunctionPass. That is, using
addRequired<>()from thegetAnalysisUsagemethod in aFunctionPassleads to an assertion failure along the lines of "Unable to schedule…". As far as I can tell this is by design. It may be solved in later LLVM versions, which have recently undergone a complete refactor of the pass manager system. Or we could move toModulePasses.