It turns out that allocating one word on stack every loop iteration doesn't scale very well.

It also turns out that if you test the cache every loop iteration but never write to it that this doesn't really make things any faster...

opts/IMPCacher.cpp

@@ -139,7 +139,7 @@ void GNUstep::IMPCacher::CacheLookup(Instruction *lookup, Value *slot, Value
   // version is not 0 at the start and an occasional redundant store is
   // probably better than a branch every time.
   B.CreateStore(lookup, slot);
-  //B.CreateStore(B.CreateLoad(B.CreateStructGEP(lookup, 3)), version);
+  B.CreateStore(B.CreateLoad(B.CreateStructGEP(lookup, 3)), version);
   cls = B.CreateLoad(B.CreateBitCast(receiver, IdTy));
   B.CreateStore(cls, B.CreateStructGEP(lookup, 1));
   B.CreateBr(afterLookupBB);
@@ -232,7 +232,8 @@ void GNUstep::IMPCacher::SpeculativelyInline(Instruction *call, Function
 CallSite GNUstep::IMPCacher::SplitSend(CallSite msgSend)
 {
   BasicBlock *lookupBB = msgSend->getParent();
-  Module *M = lookupBB->getParent()->getParent();
+  Function *F = lookupBB->getParent();
+  Module *M = F->getParent();
   Function *send = M->getFunction("objc_msgSend");
   Function *send_stret = M->getFunction("objc_msgSend_stret");
   Function *send_fpret = M->getFunction("objc_msgSend_fpret");
@@ -251,8 +252,9 @@ CallSite GNUstep::IMPCacher::SplitSend(CallSite msgSend)
     abort();
     return CallSite();
   }
-  CGBuilder B(msgSend.getInstruction());
+  CGBuilder B(&F->getEntryBlock(), F->getEntryBlock().begin());
   Value *selfPtr = B.CreateAlloca(self->getType());
+  B.SetInsertPoint(msgSend.getInstruction());
   B.CreateStore(self, selfPtr, true);
   LLVMType *impTy = msgSend.getCalledValue()->getType();
   LLVMType *slotTy = PointerType::getUnqual(StructType::get(PtrTy, PtrTy, PtrTy,