[LLVMdev] How to write optimizer loop

[Andrew Clinton]

I've written an optimization loop, with the following form:

     PassManager lpm;

     lpm.add(createLoopDeletionPass());
     lpm.add(createSCCPPass());
     lpm.add(createAggressiveDCEPass());
     lpm.add(createGlobalOptimizerPass());
     lpm.add(createGlobalDCEPass());
     lpm.add(createDeadStoreEliminationPass());
     lpm.add(createLoopDeletionPass());
     lpm.add(createInstructionCombiningPass());
     lpm.add(createCFGSimplificationPass());

     const int           maxit = 100;
     int                 it = 0;
     bool                changed = true;
     while (changed && it < maxit)
     {
         changed = lpm.run(*myModule);
         it++;
     }

Aside from the possibility that the optimizations don't converge 
(handled by the "maxit" variable), this code is erroneous since the Loop 
Deletion pass incurs LCSSA and loop-simplify, which will likely always 
modify code that has been simplified via CFGSimplification.

Is there a recommended method to write an optimization loop that 
correctly detects when an iteration has made changes to the module?  I'm 
now thinking that it will be necessary to compare the new and previous 
module, possibly with a hash function.

Also, how could I embed this loop into another PassManager so that it 
doesn't need to recompute stuff such as DominatorTree?

Andrew