X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FUtils%2FInlineCost.cpp;h=6ecd060991238c9f46ff137f400a09ef0db616aa;hb=dd4924c564c7a661b78b604ebf16dfef7aa62b35;hp=ab459c2e3f923e42cd4e6c6935809fc57f1da4bb;hpb=161660e6d282515252866a0a12642211cf5639be;p=oota-llvm.git

diff --git a/lib/Transforms/Utils/InlineCost.cpp b/lib/Transforms/Utils/InlineCost.cpp
index ab459c2e3f9..6ecd0609912 100644
--- a/lib/Transforms/Utils/InlineCost.cpp
+++ b/lib/Transforms/Utils/InlineCost.cpp
@@ -2,8 +2,8 @@
 //
 //                     The LLVM Compiler Infrastructure
 //
-// This file was developed by the LLVM research group and is distributed under
-// the University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
@@ -93,14 +93,41 @@ unsigned InlineCostAnalyzer::FunctionInfo::
 /// analyzeFunction - Fill in the current structure with information gleaned
 /// from the specified function.
 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
-  unsigned NumInsts = 0, NumBlocks = 0;
+  unsigned NumInsts = 0, NumBlocks = 0, NumVectorInsts = 0;
 
   // Look at the size of the callee.  Each basic block counts as 20 units, and
   // each instruction counts as 5.
   for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
     for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
          II != E; ++II) {
-      if (isa<DbgInfoIntrinsic>(II)) continue;  // Debug intrinsics don't count.
+      if (isa<PHINode>(II)) continue;           // PHI nodes don't count.
+
+      // Special handling for calls.
+      if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
+        if (isa<DbgInfoIntrinsic>(II))
+          continue;  // Debug intrinsics don't count as size.
+        
+        CallSite CS = CallSite::get(const_cast<Instruction*>(&*II));
+        
+        // If this function contains a call to setjmp or _setjmp, never inline
+        // it.  This is a hack because we depend on the user marking their local
+        // variables as volatile if they are live across a setjmp call, and they
+        // probably won't do this in callers.
+        if (Function *F = CS.getCalledFunction())
+          if (F->isDeclaration() && 
+              (F->isName("setjmp") || F->isName("_setjmp"))) {
+            NeverInline = true;
+            return;
+          }
+        
+        // Calls often compile into many machine instructions.  Bump up their
+        // cost to reflect this.
+        if (!isa<IntrinsicInst>(II))
+          NumInsts += 5;
+      }
+      
+      if (isa<ExtractElementInst>(II) || isa<VectorType>(II->getType()))
+        ++NumVectorInsts; 
       
       // Noop casts, including ptr <-> int,  don't count.
       if (const CastInst *CI = dyn_cast<CastInst>(II)) {
@@ -108,7 +135,7 @@ void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
             isa<PtrToIntInst>(CI))
           continue;
       } else if (const GetElementPtrInst *GEPI =
-                         dyn_cast<GetElementPtrInst>(II)) {
+                 dyn_cast<GetElementPtrInst>(II)) {
         // If a GEP has all constant indices, it will probably be folded with
         // a load/store.
         bool AllConstant = true;
@@ -126,8 +153,9 @@ void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
     ++NumBlocks;
   }
 
-  this->NumBlocks = NumBlocks;
-  this->NumInsts  = NumInsts;
+  this->NumBlocks      = NumBlocks;
+  this->NumInsts       = NumInsts;
+  this->NumVectorInsts = NumVectorInsts;
 
   // Check out all of the arguments to the function, figuring out how much
   // code can be eliminated if one of the arguments is a constant.
@@ -141,7 +169,8 @@ void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
 // getInlineCost - The heuristic used to determine if we should inline the
 // function call or not.
 //
-int InlineCostAnalyzer::getInlineCost(CallSite CS, SmallPtrSet<const Function *, 16> &NeverInline) {
+int InlineCostAnalyzer::getInlineCost(CallSite CS,
+                               SmallPtrSet<const Function *, 16> &NeverInline) {
   Instruction *TheCall = CS.getInstruction();
   Function *Callee = CS.getCalledFunction();
   const Function *Caller = TheCall->getParent()->getParent();
@@ -166,7 +195,7 @@ int InlineCostAnalyzer::getInlineCost(CallSite CS, SmallPtrSet<const Function *,
   // make it almost guaranteed to be inlined.
   //
   if (Callee->hasInternalLinkage() && Callee->hasOneUse())
-    InlineCost -= 30000;
+    InlineCost -= 15000;
   
   // If this function uses the coldcc calling convention, prefer not to inline
   // it.
@@ -188,6 +217,13 @@ int InlineCostAnalyzer::getInlineCost(CallSite CS, SmallPtrSet<const Function *,
   // If we haven't calculated this information yet, do so now.
   if (CalleeFI.NumBlocks == 0)
     CalleeFI.analyzeFunction(Callee);
+  
+  // If we should never inline this, return a huge cost.
+  if (CalleeFI.NeverInline)
+    return 2000000000;
+
+  if (!Callee->isDeclaration() && Callee->hasNote(FN_NOTE_AlwaysInline))
+    return -2000000000;
     
   // Add to the inline quality for properties that make the call valuable to
   // inline.  This includes factors that indicate that the result of inlining
@@ -227,15 +263,38 @@ int InlineCostAnalyzer::getInlineCost(CallSite CS, SmallPtrSet<const Function *,
   // Now that we have considered all of the factors that make the call site more
   // likely to be inlined, look at factors that make us not want to inline it.
   
-  // Don't inline into something too big, which would make it bigger.  Here, we
-  // count each basic block as a single unit.
+  // Don't inline into something too big, which would make it bigger.
   //
-  InlineCost += Caller->size()/20;
-  
+  InlineCost += Caller->size()/15;
   
-  // Look at the size of the callee.  Each basic block counts as 20 units, and
-  // each instruction counts as 5.
-  InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20;
+  // Look at the size of the callee. Each instruction counts as 5.
+  InlineCost += CalleeFI.NumInsts*5;
+
   return InlineCost;
 }
 
+// getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
+// higher threshold to determine if the function call should be inlined.
+float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
+  Function *Callee = CS.getCalledFunction();
+  
+  // Get information about the callee...
+  FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
+  
+  // If we haven't calculated this information yet, do so now.
+  if (CalleeFI.NumBlocks == 0)
+    CalleeFI.analyzeFunction(Callee);
+
+  float Factor = 1.0f;
+  // Single BB functions are often written to be inlined.
+  if (CalleeFI.NumBlocks == 1)
+    Factor += 0.5f;
+
+  // Be more aggressive if the function contains a good chunk (if it mades up
+  // at least 10% of the instructions) of vector instructions.
+  if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/2)
+    Factor += 2.0f;
+  else if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/10)
+    Factor += 1.5f;
+  return Factor;
+}