[PATCH] D18576: Initial implementation of optimization bisect

[Matthias Braun via llvm-commits]

MatzeB added a comment.

In http://reviews.llvm.org/D18576#400442, @andrew.w.kaylor wrote:

> In http://reviews.llvm.org/D18576#399046, @MatzeB wrote:
>
> > We have to count how many passes we skipped/did not skip and compare that against the bisect number. This is global information.
>
>
> Is there a good way to get the context from a CallGraphSCC?  This is what I came up with:
>
>   static LLVMContext *getCGNodeContext(CallGraphNode *Node) {
>     Function *F = Node->getFunction();
>     if (F)
>       return &F->getContext();
>     for (CallGraphNode::iterator I = Node->begin(), E = Node->end(); I != E;
>          ++I) {
>       if (I->first)
>         return &cast<Function>(I->first)->getContext();
>       Function *CalledFn = I->second->getFunction();
>       // A null CalledFn indicates a call to an external node.
>       if (!CalledFn)
>         continue;
>       return &(CalledFn->getContext());
>     }
>   }
>  
>   bool CallGraphSCCPass::skipSCC(CallGraphSCC &SCC) const {
>     CallGraphNode *Node = *(SCC.begin());
>     assert(Node);
>     if (!Node)
>       return false;
>     LLVMContext *Context = nullptr;
>     for (CallGraphNode *Node : SCC) {
>       Context = getCGNodeContext(Node);
>       if (Context)
>         break;
>     }
>     assert(Context);
>     // If we can't get the context, assume the pass can't be skipped.
>     if (!Context)
>       return false;
>     return !Context->getOptBisect().shouldRunPass(this, &SCC);
>   }
>
>
> It feels awful.  Is there a better way to do that?


This does indeed look ugly. The best solution I can see would be to give the CallGraphSCC class a reference to the CallGraph (from there you should be able to go CallGraph.getModule().getContext()). CallGraphSCC is only instantiated once per CallGraphSCCPassManager as far as I can see so an extra reference shouldn't hurt.


Repository:
  rL LLVM

http://reviews.llvm.org/D18576