Don't use potentially invalidated iterator

[sberg]

Summary

If the lhs is evaluated before the rhs, FuncletI's operator-> can trigger the

  
assert(isHandleInSync() && "invalid iterator access!");

at include/llvm/ADT/DenseMap.h:1061. (Happens e.g. when compiled with GCC 6.)

[sberg]

• That code had originally been introduced with David Majnemer's r249257 "[WinEH] Permit branch folding in the face of funclets", hence setting him as reviewer.
• No idea what those int stored in FuncletMembership represent, so didn't come up with a better name than generic "n" for the var.

[sberg]

friendly ping

(I have commit access now, so all it would need is a "lgtm" :)

[dblaikie]

LGTM, thanks!

(drop the const from const auto - we don't usually constify value types
defensively like that)

Also - what type is 'n'? Something simple/cheap? Or should it be referenced
by (const) ref instead of by value?

[sberg]

Dropped the const. The type of n is int, so taking by value is fine. (But guess it's better to stick to auto, in case the value type of FuncletMembership should ever change.)

llvm/trunk/lib/CodeGen/BranchFolding.cpp

MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
   // NewMBB inherits CurMBB's block frequency.
   MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB));
 
   // For targets that use the register scavenger, we must maintain LiveIns.
   MaintainLiveIns(&CurMBB, NewMBB);
 
   // Add the new block to the funclet.
   const auto &FuncletI = FuncletMembership.find(&CurMBB);
-  if (FuncletI != FuncletMembership.end())
-    FuncletMembership[NewMBB] = FuncletI->second;
+  if (FuncletI != FuncletMembership.end()) {
+    auto n = FuncletI->second;
+    FuncletMembership[NewMBB] = n;
+  }
 
   return NewMBB;
 }
 
 /// EstimateRuntime - Make a rough estimate for how long it will take to run
 /// the specified code.
 static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
                                 MachineBasicBlock::iterator E) {