[CodeExtractor]: Fix an extractor bug of missing replacing use of old phi with the new one

[davidxl]

Summary

This is triggered by partial inliner improvement patch (so a test case is not yet available) -- but the bug seems obvious. (found with clang built with partial inlining turned on -- it causes infinite loop during indirect call promotion).

[davide]

uh, how this ever worked is clearly beyond me.
If you can add these tests to the inliner improvement patch, that would be great.

llvm/trunk/lib/Transforms/Utils/CodeExtractor.cpp

if (NumPredsFromRegion) {
     // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
     BasicBlock::iterator AfterPHIs;
     for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
       PHINode *PN = cast<PHINode>(AfterPHIs);
       // Create a new PHI node in the new region, which has an incoming value
       // from OldPred of PN.
       PHINode *NewPN = PHINode::Create(PN->getType(), 1 + NumPredsFromRegion,
                                        PN->getName() + ".ce", &NewBB->front());
+      PN->replaceAllUsesWith(NewPN);
       NewPN->addIncoming(PN, OldPred);
 
       // Loop over all of the incoming value in PN, moving them to NewPN if they
       // are from the extracted region.
       for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
         if (Blocks.count(PN->getIncomingBlock(i))) {
           NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
           PN->removeIncomingValue(i);