Always skip ptr-to-ptr bitcasts when counting,

per Chris' suggestion.  Slightly faster.

llvm-svn: 65999

2 files changed, 10 insertions, 7 deletions

diff --git a/llvm/lib/Transforms/Scalar/InstructionCombining.cpp b/llvm/lib/Transforms/Scalar/InstructionCombining.cpp
index 5b50ab56a3e..66ac48181e6 100644
--- a/llvm/lib/Transforms/Scalar/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/Scalar/InstructionCombining.cpp
@@ -11484,12 +11484,12 @@ Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
   for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
        --ScanInsts) {
     --BBI;
-    // Don't count debug info directives, lest they affect codegen.
-    // Likewise, we skip bitcasts that feed into a llvm.dbg.declare; these are
-    // not present when debugging is off.
+    // Don't count debug info directives, lest they affect codegen,
+    // and we skip pointer-to-pointer bitcasts, which are NOPs.
+    // It is necessary for correctness to skip those that feed into a
+    // llvm.dbg.declare, as these are not present when debugging is off.
     if (isa<DbgInfoIntrinsic>(BBI) ||
-        (isa<BitCastInst>(BBI) && BBI->hasOneUse() &&
-         isa<DbgDeclareInst>(BBI->use_begin()))) {
+        (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
       ScanInsts++;
       continue;
     }    
diff --git a/llvm/lib/Transforms/Utils/InlineFunction.cpp b/llvm/lib/Transforms/Utils/InlineFunction.cpp
index a96c7ceaa8e..9cd38cc55e0 100644
--- a/llvm/lib/Transforms/Utils/InlineFunction.cpp
+++ b/llvm/lib/Transforms/Utils/InlineFunction.cpp
@@ -235,7 +235,7 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) {
   // function.
   std::vector<ReturnInst*> Returns;
   ClonedCodeInfo InlinedFunctionInfo;
-  Function::iterator FirstNewBlock;
+  Function::iterator FirstNewBlock, LastNewBlock;
 
   { // Scope to destroy ValueMap after cloning.
     DenseMap<const Value*, Value*> ValueMap;
@@ -312,6 +312,7 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) {
 
     // Remember the first block that is newly cloned over.
     FirstNewBlock = LastBlock; ++FirstNewBlock;
+    LastNewBlock = &Caller->back();
 
     // Update the callgraph if requested.
     if (CG)
@@ -537,7 +538,9 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) {
     // Add a branch to the merge points and remove return instructions.
     for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
       ReturnInst *RI = Returns[i];
-      BranchInst::Create(AfterCallBB, RI);
+      // A return in the last block in the function falls through.
+//      if (isa<InvokeInst>(TheCall) || RI->getParent() != LastNewBlock)
+        BranchInst::Create(AfterCallBB, RI);
       RI->eraseFromParent();
     }
   } else if (!Returns.empty()) {