Try to keep the cached inliner costs around for a bit longer for big functions.

The Caller cost info would be reset everytime a callee was inlined. If the
caller has lots of calls and there is some mutual recursion going on, the
caller cost info could be calculated many times.

This patch reduces inliner runtime from 240s to 0.5s for a function with 20000
small function calls.

This is a more conservative version of r98089 that doesn't break the clang
test CodeGenCXX/temp-order.cpp. That test relies on rather extreme inlining
for constant folding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@98099 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 12 insertions, 6 deletions

diff --git a/lib/Transforms/IPO/InlineAlways.cpp b/lib/Transforms/IPO/InlineAlways.cpp
index f11ecae8df..bc8028c020 100644
--- a/lib/Transforms/IPO/InlineAlways.cpp
+++ b/lib/Transforms/IPO/InlineAlways.cpp
@@ -45,7 +45,10 @@ namespace {
       return CA.getInlineFudgeFactor(CS);
     }
     void resetCachedCostInfo(Function *Caller) {
-      return CA.resetCachedCostInfo(Caller);
+      CA.resetCachedCostInfo(Caller);
+    }
+    void growCachedCostInfo(Function* Caller, Function* Callee) {
+      CA.growCachedCostInfo(Caller, Callee);
     }
     virtual bool doFinalization(CallGraph &CG) { 
       return removeDeadFunctions(CG, &NeverInline); 
diff --git a/lib/Transforms/IPO/InlineSimple.cpp b/lib/Transforms/IPO/InlineSimple.cpp
index 598043de69..46cf4b25e4 100644
--- a/lib/Transforms/IPO/InlineSimple.cpp
+++ b/lib/Transforms/IPO/InlineSimple.cpp
@@ -45,6 +45,9 @@ namespace {
     void resetCachedCostInfo(Function *Caller) {
       CA.resetCachedCostInfo(Caller);
     }
+    void growCachedCostInfo(Function* Caller, Function* Callee) {
+      CA.growCachedCostInfo(Caller, Callee);
+    }
     virtual bool doInitialization(CallGraph &CG);
   };
 }
diff --git a/lib/Transforms/IPO/Inliner.cpp b/lib/Transforms/IPO/Inliner.cpp
index 3d0309f87c..03ec72c030 100644
--- a/lib/Transforms/IPO/Inliner.cpp
+++ b/lib/Transforms/IPO/Inliner.cpp
@@ -369,6 +369,8 @@ bool Inliner::runOnSCC(std::vector<CallGraphNode*> &SCC) {
         CG[Caller]->removeCallEdgeFor(CS);
         CS.getInstruction()->eraseFromParent();
         ++NumCallsDeleted;
+        // Update the cached cost info with the missing call
+        growCachedCostInfo(Caller, NULL);
       } else {
         // We can only inline direct calls to non-declarations.
         if (Callee == 0 || Callee->isDeclaration()) continue;
@@ -382,6 +384,9 @@ bool Inliner::runOnSCC(std::vector<CallGraphNode*> &SCC) {
         if (!InlineCallIfPossible(CS, CG, TD, InlinedArrayAllocas))
           continue;
         ++NumInlined;
+
+        // Update the cached cost info with the inlined call.
+        growCachedCostInfo(Caller, Callee);
       }
       
       // If we inlined or deleted the last possible call site to the function,
@@ -407,11 +412,6 @@ bool Inliner::runOnSCC(std::vector<CallGraphNode*> &SCC) {
         delete CG.removeFunctionFromModule(CalleeNode);
         ++NumDeleted;
       }
-      
-      // Remove any cached cost info for this caller, as inlining the
-      // callee has increased the size of the caller (which may be the
-      // same as the callee).
-      resetCachedCostInfo(Caller);
 
       // Remove this call site from the list.  If possible, use 
       // swap/pop_back for efficiency, but do not use it if doing so would