Improve debugging output when there is a null function pointer in the IR dump.

[rtrieu]

Summary

Add new lines so debugging information isn't all on one line.

[reames]

LGTM, this seems like a strict improvement.

However, is this the best reporting we could give? From what I can gather, null function means either an indirect call or the external function node in the callgraph. Given the purpose of this pass, would simply skipping the null node entirely be clearer?

[rtrieu]

Closed by commit rL212250 (authored by @rtrieu).

[rtrieu]

However, is this the best reporting we could give? From what I can gather, null function means either an indirect call or the external function node in the callgraph. Given the purpose of this pass, would simply skipping the null node entirely be clearer?

I'm committed my patch. I don't have any answers since I don't this part of LLVM. Someone who regularly works with the optimizations would be able to answer which approach would be the most useful.

llvm/trunk/lib/Analysis/IPA/CallGraphSCCPass.cpp

public:
     }
 
     bool runOnSCC(CallGraphSCC &SCC) override {
       Out << Banner;
       for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
         if ((*I)->getFunction())
           (*I)->getFunction()->print(Out);
         else
-          Out << "Printing <null> Function";
+          Out << "\nPrinting <null> Function\n";
       }
       return false;
     }
   };
   
 } // end anonymous namespace.
 
 char PrintCallGraphPass::ID = 0;
 
 Pass *CallGraphSCCPass::createPrinterPass(raw_ostream &O,
                                           const std::string &Banner) const {
   return new PrintCallGraphPass(Banner, O);
 }
 

llvm/trunk/test/Transforms/Inline/null-function.ll

+; RUN: opt -print-before=always-inline -always-inline < %s -o /dev/null 2>&1 | FileCheck %s
+
+define i32 @main() #0 {
+entry:
+  ret i32 0
+}
+
+; CHECK: *** IR Dump Before Inliner for always_inline functions ***
+; CHECK: Printing <null> Function