Remove NEON vmovn intrinsic, replacing it with vector truncate operations.
[oota-llvm.git] / lib / VMCore / AutoUpgrade.cpp
index af4417e0f0bd851b2511a88188e4756fb55eadff..c99433965338dde780c045c53cdb2a64dd8802ee 100644 (file)
@@ -2,8 +2,8 @@
 //
 //                     The LLVM Compiler Infrastructure
 //
-// This file was developed by Reid Spencer 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.
 //
 //===----------------------------------------------------------------------===//
 //
 //
 //===----------------------------------------------------------------------===//
 
-#include "llvm/Assembly/AutoUpgrade.h"
+#include "llvm/AutoUpgrade.h"
 #include "llvm/Constants.h"
-#include "llvm/DerivedTypes.h"
 #include "llvm/Function.h"
+#include "llvm/LLVMContext.h"
 #include "llvm/Module.h"
-#include "llvm/Instructions.h"
-#include "llvm/Intrinsics.h"
-#include "llvm/SymbolTable.h"
-#include <iostream>
+#include "llvm/IntrinsicInst.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/Support/CallSite.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/IRBuilder.h"
+#include <cstring>
 using namespace llvm;
 
-static Function *getUpgradedUnaryFn(Function *F) {
-  const std::string &Name = F->getName();
-  Module *M = F->getParent();
-  switch (F->getReturnType()->getTypeID()) {
-  default: return 0;
-  case Type::UByteTyID:
-  case Type::SByteTyID:
-    return M->getOrInsertFunction(Name+".i8", 
-                                  Type::UByteTy, Type::UByteTy, NULL);
-  case Type::UShortTyID:
-  case Type::ShortTyID:
-    return M->getOrInsertFunction(Name+".i16", 
-                                  Type::UShortTy, Type::UShortTy, NULL);
-  case Type::UIntTyID:
-  case Type::IntTyID:
-    return M->getOrInsertFunction(Name+".i32", 
-                                  Type::UIntTy, Type::UIntTy, NULL);
-  case Type::ULongTyID:
-  case Type::LongTyID:
-    return M->getOrInsertFunction(Name+".i64",
-                                  Type::ULongTy, Type::ULongTy, NULL);
-  case Type::FloatTyID:
-    return M->getOrInsertFunction(Name+".f32",
-                                  Type::FloatTy, Type::FloatTy, NULL);
-  case Type::DoubleTyID:
-    return M->getOrInsertFunction(Name+".f64",
-                                  Type::DoubleTy, Type::DoubleTy, NULL);
-  }
-}
 
-static Function *getUpgradedIntrinsic(Function *F) {
-  // If there's no function, we can't get the argument type.
-  if (!F) return 0;
+static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
+  assert(F && "Illegal to upgrade a non-existent Function.");
 
   // Get the Function's name.
   const std::string& Name = F->getName();
 
+  // Convenience
+  const FunctionType *FTy = F->getFunctionType();
+
   // Quickly eliminate it, if it's not a candidate.
   if (Name.length() <= 8 || Name[0] != 'l' || Name[1] != 'l' || 
       Name[2] != 'v' || Name[3] != 'm' || Name[4] != '.')
-    return 0;
+    return false;
 
   Module *M = F->getParent();
   switch (Name[5]) {
   default: break;
-  case 'b':
-    if (Name == "llvm.bswap") return getUpgradedUnaryFn(F);
-    break;
-  case 'c':
-    if (Name == "llvm.ctpop" || Name == "llvm.ctlz" || Name == "llvm.cttz")
-      return getUpgradedUnaryFn(F);
-    break;
-  case 'd':
-    if (Name == "llvm.dbg.stoppoint") {
-      PointerType *ESP =
-                  PointerType::get(StructType::get(std::vector<const Type*>()));
-      if (F->getReturnType() != Type::VoidTy ||
-          F->getFunctionType()->getParamType(2) != ESP) {
-        return M->getOrInsertFunction(Name, Type::VoidTy,
-                                      Type::UIntTy, Type::UIntTy, ESP, NULL);
+  case 'a':
+    // This upgrades the llvm.atomic.lcs, llvm.atomic.las, llvm.atomic.lss,
+    // and atomics with default address spaces to their new names to their new
+    // function name (e.g. llvm.atomic.add.i32 => llvm.atomic.add.i32.p0i32)
+    if (Name.compare(5,7,"atomic.",7) == 0) {
+      if (Name.compare(12,3,"lcs",3) == 0) {
+        std::string::size_type delim = Name.find('.',12);
+        F->setName("llvm.atomic.cmp.swap" + Name.substr(delim) +
+                   ".p0" + Name.substr(delim+1));
+        NewFn = F;
+        return true;
+      }
+      else if (Name.compare(12,3,"las",3) == 0) {
+        std::string::size_type delim = Name.find('.',12);
+        F->setName("llvm.atomic.load.add"+Name.substr(delim)
+                   + ".p0" + Name.substr(delim+1));
+        NewFn = F;
+        return true;
       }
-    } else if (Name == "llvm.dbg.func.start") {
-      PointerType *ESP =
-                  PointerType::get(StructType::get(std::vector<const Type*>()));
-      if (F->getReturnType()  != Type::VoidTy ||
-          F->getFunctionType()->getParamType(0) != ESP) {
-        return M->getOrInsertFunction(Name, Type::VoidTy, ESP, NULL);
+      else if (Name.compare(12,3,"lss",3) == 0) {
+        std::string::size_type delim = Name.find('.',12);
+        F->setName("llvm.atomic.load.sub"+Name.substr(delim)
+                   + ".p0" + Name.substr(delim+1));
+        NewFn = F;
+        return true;
       }
-    } else if (Name == "llvm.dbg.region.start") {
-      PointerType *ESP =
-                  PointerType::get(StructType::get(std::vector<const Type*>()));
-      if (F->getReturnType() != Type::VoidTy ||
-          F->getFunctionType()->getParamType(0) != ESP) {
-        return M->getOrInsertFunction(Name, Type::VoidTy,  ESP, NULL);
+      else if (Name.rfind(".p") == std::string::npos) {
+        // We don't have an address space qualifier so this has be upgraded
+        // to the new name.  Copy the type name at the end of the intrinsic
+        // and add to it
+        std::string::size_type delim = Name.find_last_of('.');
+        assert(delim != std::string::npos && "can not find type");
+        F->setName(Name + ".p0" + Name.substr(delim+1));
+        NewFn = F;
+        return true;
       }
-    } else if (Name == "llvm.dbg.region.end") {
-      PointerType *ESP =
-                  PointerType::get(StructType::get(std::vector<const Type*>()));
-      if (F->getReturnType() != Type::VoidTy ||
-          F->getFunctionType()->getParamType(0) != ESP) {
-         return M->getOrInsertFunction(Name, Type::VoidTy,  ESP, NULL);
+    } else if (Name.compare(5, 9, "arm.neon.", 9) == 0) {
+      if (((Name.compare(14, 5, "vmovl", 5) == 0 ||
+            Name.compare(14, 5, "vaddl", 5) == 0 ||
+            Name.compare(14, 5, "vsubl", 5) == 0) &&
+           (Name.compare(19, 2, "s.", 2) == 0 ||
+            Name.compare(19, 2, "u.", 2) == 0)) ||
+
+          ((Name.compare(14, 5, "vaddw", 5) == 0 ||
+            Name.compare(14, 5, "vsubw", 5) == 0) &&
+           (Name.compare(19, 2, "s.", 2) == 0 ||
+            Name.compare(19, 2, "u.", 2) == 0)) ||
+
+          (Name.compare(14, 6, "vmovn.", 6) == 0)) {
+
+        // Calls to these are transformed into IR without intrinsics.
+        NewFn = 0;
+        return true;
       }
-    } else if (Name == "llvm.dbg.declare") {
-      PointerType *ESP =
-                  PointerType::get(StructType::get(std::vector<const Type*>()));
-      if (F->getReturnType() != Type::VoidTy ||
-          F->getFunctionType()->getParamType(0) != ESP ||
-          F->getFunctionType()->getParamType(1) != ESP) {
-        return M->getOrInsertFunction(Name, Type::VoidTy, ESP, ESP, NULL);
+      // Old versions of NEON ld/st intrinsics are missing alignment arguments.
+      bool isVLd = (Name.compare(14, 3, "vld", 3) == 0);
+      bool isVSt = (Name.compare(14, 3, "vst", 3) == 0);
+      if (isVLd || isVSt) {
+        unsigned NumVecs = Name.at(17) - '0';
+        if (NumVecs == 0 || NumVecs > 4)
+          return false;
+        bool isLaneOp = (Name.compare(18, 5, "lane.", 5) == 0);
+        if (!isLaneOp && Name.at(18) != '.')
+          return false;
+        unsigned ExpectedArgs = 2; // for the address and alignment
+        if (isVSt || isLaneOp)
+          ExpectedArgs += NumVecs;
+        if (isLaneOp)
+          ExpectedArgs += 1; // for the lane number
+        unsigned NumP = FTy->getNumParams();
+        if (NumP != ExpectedArgs - 1)
+          return false;
+
+        // Change the name of the old (bad) intrinsic, because 
+        // its type is incorrect, but we cannot overload that name.
+        F->setName("");
+
+        // One argument is missing: add the alignment argument.
+        std::vector<const Type*> NewParams;
+        for (unsigned p = 0; p < NumP; ++p)
+          NewParams.push_back(FTy->getParamType(p));
+        NewParams.push_back(Type::getInt32Ty(F->getContext()));
+        FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(),
+                                                 NewParams, false);
+        NewFn = cast<Function>(M->getOrInsertFunction(Name, NewFTy));
+        return true;
       }
     }
     break;
-  case 'i':
-    if (Name == "llvm.isunordered" && F->arg_begin() != F->arg_end()) {
-      if (F->arg_begin()->getType() == Type::FloatTy)
-        return M->getOrInsertFunction(Name+".f32", F->getFunctionType());
-      if (F->arg_begin()->getType() == Type::DoubleTy)
-        return M->getOrInsertFunction(Name+".f64", F->getFunctionType());
+  case 'b':
+    //  This upgrades the name of the llvm.bswap intrinsic function to only use 
+    //  a single type name for overloading. We only care about the old format
+    //  'llvm.bswap.i*.i*', so check for 'bswap.' and then for there being 
+    //  a '.' after 'bswap.'
+    if (Name.compare(5,6,"bswap.",6) == 0) {
+      std::string::size_type delim = Name.find('.',11);
+      
+      if (delim != std::string::npos) {
+        //  Construct the new name as 'llvm.bswap' + '.i*'
+        F->setName(Name.substr(0,10)+Name.substr(delim));
+        NewFn = F;
+        return true;
+      }
     }
     break;
-  case 'm':
-    if (Name == "llvm.memcpy" || Name == "llvm.memset" || 
-        Name == "llvm.memmove") {
-      if (F->getFunctionType()->getParamType(2) == Type::UIntTy ||
-          F->getFunctionType()->getParamType(2) == Type::IntTy)
-        return M->getOrInsertFunction(Name+".i32", Type::VoidTy,
-                                      PointerType::get(Type::SByteTy),
-                                      F->getFunctionType()->getParamType(1),
-                                      Type::UIntTy, Type::UIntTy, NULL);
-      if (F->getFunctionType()->getParamType(2) == Type::ULongTy ||
-          F->getFunctionType()->getParamType(2) == Type::LongTy)
-        return M->getOrInsertFunction(Name+".i64", Type::VoidTy,
-                                      PointerType::get(Type::SByteTy),
-                                      F->getFunctionType()->getParamType(1),
-                                      Type::ULongTy, Type::UIntTy, NULL);
+
+  case 'c':
+    //  We only want to fix the 'llvm.ct*' intrinsics which do not have the 
+    //  correct return type, so we check for the name, and then check if the 
+    //  return type does not match the parameter type.
+    if ( (Name.compare(5,5,"ctpop",5) == 0 ||
+          Name.compare(5,4,"ctlz",4) == 0 ||
+          Name.compare(5,4,"cttz",4) == 0) &&
+        FTy->getReturnType() != FTy->getParamType(0)) {
+      //  We first need to change the name of the old (bad) intrinsic, because 
+      //  its type is incorrect, but we cannot overload that name. We 
+      //  arbitrarily unique it here allowing us to construct a correctly named 
+      //  and typed function below.
+      F->setName("");
+
+      //  Now construct the new intrinsic with the correct name and type. We 
+      //  leave the old function around in order to query its type, whatever it 
+      //  may be, and correctly convert up to the new type.
+      NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                    FTy->getParamType(0),
+                                                    FTy->getParamType(0),
+                                                    (Type *)0));
+      return true;
     }
     break;
-  case 's':
-    if (Name == "llvm.sqrt")
-      return getUpgradedUnaryFn(F);
+
+  case 'e':
+    //  The old llvm.eh.selector.i32 is equivalent to the new llvm.eh.selector.
+    if (Name.compare("llvm.eh.selector.i32") == 0) {
+      F->setName("llvm.eh.selector");
+      NewFn = F;
+      return true;
+    }
+    //  The old llvm.eh.typeid.for.i32 is equivalent to llvm.eh.typeid.for.
+    if (Name.compare("llvm.eh.typeid.for.i32") == 0) {
+      F->setName("llvm.eh.typeid.for");
+      NewFn = F;
+      return true;
+    }
+    //  Convert the old llvm.eh.selector.i64 to a call to llvm.eh.selector.
+    if (Name.compare("llvm.eh.selector.i64") == 0) {
+      NewFn = Intrinsic::getDeclaration(M, Intrinsic::eh_selector);
+      return true;
+    }
+    //  Convert the old llvm.eh.typeid.for.i64 to a call to llvm.eh.typeid.for.
+    if (Name.compare("llvm.eh.typeid.for.i64") == 0) {
+      NewFn = Intrinsic::getDeclaration(M, Intrinsic::eh_typeid_for);
+      return true;
+    }
     break;
-  }
-  return 0;
-}
 
-// Occasionally upgraded function call site arguments need to be permutated to
-// some new order.  The result of getArgumentPermutation is an array of size 
-// F->getFunctionType()getNumParams() indicating the new operand order.  A value
-// of zero in the array indicates replacing with UndefValue for the arg type.
-// NULL is returned if there is no permutation.  It's assumed that the function
-// name is in the form "llvm.?????"
-static unsigned *getArgumentPermutation(Function* Fn, Function* NewFn) {
-  const std::string& Name = Fn->getName();
-  unsigned N = Fn->getFunctionType()->getNumParams();
-  unsigned M = NewFn->getFunctionType()->getNumParams();
-  
-  switch (Name[5]) {
-  case 'd':
-    if (Name == "llvm.dbg.stoppoint") {
-      static unsigned Permutation[] = { 2, 3, 4 };
-      assert(M == (sizeof(Permutation) / sizeof(unsigned)) &&
-             "Permutation is wrong length");
-      if (N == 4) return Permutation;
-    } else if (Name == "llvm.dbg.region.start") {
-      static unsigned Permutation[] = { 0 };
-      assert(M == (sizeof(Permutation) / sizeof(unsigned)) &&
-             "Permutation is wrong length");
-      if (N == 0) return Permutation;
-    } else if (Name == "llvm.dbg.region.end") {
-      static unsigned Permutation[] = { 0 };
-      assert(M == (sizeof(Permutation) / sizeof(unsigned)) &&
-             "Permutation is wrong length");
-      if (N == 0) return Permutation;
-    } else if (Name == "llvm.dbg.declare") {
-      static unsigned Permutation[] = { 0, 0 };
-      assert(M == (sizeof(Permutation) / sizeof(unsigned)) &&
-             "Permutation is wrong length");
-      if (N == 0) return Permutation;
+  case 'm': {
+    // This upgrades the llvm.memcpy, llvm.memmove, and llvm.memset to the
+    // new format that allows overloading the pointer for different address
+    // space (e.g., llvm.memcpy.i16 => llvm.memcpy.p0i8.p0i8.i16)
+    const char* NewFnName = NULL;
+    if (Name.compare(5,8,"memcpy.i",8) == 0) {
+      if (Name[13] == '8')
+        NewFnName = "llvm.memcpy.p0i8.p0i8.i8";
+      else if (Name.compare(13,2,"16") == 0)
+        NewFnName = "llvm.memcpy.p0i8.p0i8.i16";
+      else if (Name.compare(13,2,"32") == 0)
+        NewFnName = "llvm.memcpy.p0i8.p0i8.i32";
+      else if (Name.compare(13,2,"64") == 0)
+        NewFnName = "llvm.memcpy.p0i8.p0i8.i64";
+    } else if (Name.compare(5,9,"memmove.i",9) == 0) {
+      if (Name[14] == '8')
+        NewFnName = "llvm.memmove.p0i8.p0i8.i8";
+      else if (Name.compare(14,2,"16") == 0)
+        NewFnName = "llvm.memmove.p0i8.p0i8.i16";
+      else if (Name.compare(14,2,"32") == 0)
+        NewFnName = "llvm.memmove.p0i8.p0i8.i32";
+      else if (Name.compare(14,2,"64") == 0)
+        NewFnName = "llvm.memmove.p0i8.p0i8.i64";
+    }
+    else if (Name.compare(5,8,"memset.i",8) == 0) {
+      if (Name[13] == '8')
+        NewFnName = "llvm.memset.p0i8.i8";
+      else if (Name.compare(13,2,"16") == 0)
+        NewFnName = "llvm.memset.p0i8.i16";
+      else if (Name.compare(13,2,"32") == 0)
+        NewFnName = "llvm.memset.p0i8.i32";
+      else if (Name.compare(13,2,"64") == 0)
+        NewFnName = "llvm.memset.p0i8.i64";
+    }
+    if (NewFnName) {
+      NewFn = cast<Function>(M->getOrInsertFunction(NewFnName, 
+                                            FTy->getReturnType(),
+                                            FTy->getParamType(0),
+                                            FTy->getParamType(1),
+                                            FTy->getParamType(2),
+                                            FTy->getParamType(3),
+                                            Type::getInt1Ty(F->getContext()),
+                                            (Type *)0));
+      return true;
     }
     break;
   }
-  return NULL;
-}
+  case 'p':
+    //  This upgrades the llvm.part.select overloaded intrinsic names to only 
+    //  use one type specifier in the name. We only care about the old format
+    //  'llvm.part.select.i*.i*', and solve as above with bswap.
+    if (Name.compare(5,12,"part.select.",12) == 0) {
+      std::string::size_type delim = Name.find('.',17);
+      
+      if (delim != std::string::npos) {
+        //  Construct a new name as 'llvm.part.select' + '.i*'
+        F->setName(Name.substr(0,16)+Name.substr(delim));
+        NewFn = F;
+        return true;
+      }
+      break;
+    }
 
-// UpgradeIntrinsicFunction - Convert overloaded intrinsic function names to
-// their non-overloaded variants by appending the appropriate suffix based on
-// the argument types.
-Function *llvm::UpgradeIntrinsicFunction(Function* F) {
-  // See if its one of the name's we're interested in.
-  if (Function *R = getUpgradedIntrinsic(F)) {
-    std::cerr << "WARNING: change " << F->getName() << " to "
-              << R->getName() << "\n";
-    return R;
-  }
-  return 0;
-}
+    //  This upgrades the llvm.part.set intrinsics similarly as above, however 
+    //  we care about 'llvm.part.set.i*.i*.i*', but only the first two types 
+    //  must match. There is an additional type specifier after these two 
+    //  matching types that we must retain when upgrading.  Thus, we require 
+    //  finding 2 periods, not just one, after the intrinsic name.
+    if (Name.compare(5,9,"part.set.",9) == 0) {
+      std::string::size_type delim = Name.find('.',14);
 
-// CastArg - Perform the appropriate cast of an upgraded argument.
-//
-static Value *CastArg(Value *Arg, const Type *Ty) {
-  if (Constant *C = dyn_cast<Constant>(Arg)) {
-    return ConstantExpr::getCast(C, Ty);
-  } else {
-    return new CastInst(Arg, Ty, "autoupgrade_cast");
-  }
-}
+      if (delim != std::string::npos &&
+          Name.find('.',delim+1) != std::string::npos) {
+        //  Construct a new name as 'llvm.part.select' + '.i*.i*'
+        F->setName(Name.substr(0,13)+Name.substr(delim));
+        NewFn = F;
+        return true;
+      }
+      break;
+    }
 
-Instruction* llvm::MakeUpgradedCall(Function *F, 
-                                    const std::vector<Value*> &Params,
-                                    BasicBlock *BB, bool isTailCall,
-                                    unsigned CallingConv) {
-  assert(F && "Need a Function to make a CallInst");
-  assert(BB && "Need a BasicBlock to make a CallInst");
-
-  // Convert the params
-  bool signedArg = false;
-  std::vector<Value*> Oprnds;
-  for (std::vector<Value*>::const_iterator PI = Params.begin(), 
-       PE = Params.end(); PI != PE; ++PI) {
-    const Type* opTy = (*PI)->getType();
-    if (opTy->isSigned()) {
-      signedArg = true;
-      Value *cast = CastArg(*PI, opTy->getUnsignedVersion());
-      if (Instruction *I = dyn_cast<Instruction>(cast))
-        BB->getInstList().push_back(I);
-      Oprnds.push_back(cast);
+    break;
+  case 'x': 
+    // This fixes all MMX shift intrinsic instructions to take a
+    // v1i64 instead of a v2i32 as the second parameter.
+    if (Name.compare(5,10,"x86.mmx.ps",10) == 0 &&
+        (Name.compare(13,4,"psll", 4) == 0 ||
+         Name.compare(13,4,"psra", 4) == 0 ||
+         Name.compare(13,4,"psrl", 4) == 0) && Name[17] != 'i') {
+      
+      const llvm::Type *VT =
+                    VectorType::get(IntegerType::get(FTy->getContext(), 64), 1);
+      
+      // We don't have to do anything if the parameter already has
+      // the correct type.
+      if (FTy->getParamType(1) == VT)
+        break;
+      
+      //  We first need to change the name of the old (bad) intrinsic, because 
+      //  its type is incorrect, but we cannot overload that name. We 
+      //  arbitrarily unique it here allowing us to construct a correctly named 
+      //  and typed function below.
+      F->setName("");
+
+      assert(FTy->getNumParams() == 2 && "MMX shift intrinsics take 2 args!");
+      
+      //  Now construct the new intrinsic with the correct name and type. We 
+      //  leave the old function around in order to query its type, whatever it 
+      //  may be, and correctly convert up to the new type.
+      NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                    FTy->getReturnType(),
+                                                    FTy->getParamType(0),
+                                                    VT,
+                                                    (Type *)0));
+      return true;
+    } else if (Name.compare(5,17,"x86.sse2.loadh.pd",17) == 0 ||
+               Name.compare(5,17,"x86.sse2.loadl.pd",17) == 0 ||
+               Name.compare(5,16,"x86.sse2.movl.dq",16) == 0 ||
+               Name.compare(5,15,"x86.sse2.movs.d",15) == 0 ||
+               Name.compare(5,16,"x86.sse2.shuf.pd",16) == 0 ||
+               Name.compare(5,18,"x86.sse2.unpckh.pd",18) == 0 ||
+               Name.compare(5,18,"x86.sse2.unpckl.pd",18) == 0 ||
+               Name.compare(5,20,"x86.sse2.punpckh.qdq",20) == 0 ||
+               Name.compare(5,20,"x86.sse2.punpckl.qdq",20) == 0) {
+      // Calls to these intrinsics are transformed into ShuffleVector's.
+      NewFn = 0;
+      return true;
+    } else if (Name.compare(5, 16, "x86.sse41.pmulld", 16) == 0) {
+      // Calls to these intrinsics are transformed into vector multiplies.
+      NewFn = 0;
+      return true;
+    } else if (Name.compare(5, 18, "x86.ssse3.palign.r", 18) == 0 ||
+               Name.compare(5, 22, "x86.ssse3.palign.r.128", 22) == 0) {
+      // Calls to these intrinsics are transformed into vector shuffles, shifts,
+      // or 0.
+      NewFn = 0;
+      return true;           
     }
-    else
-      Oprnds.push_back(*PI);
-  }
 
-  Instruction *result = new CallInst(F, Oprnds);
-  if (result->getType() != Type::VoidTy) result->setName("autoupgrade_call");
-  if (isTailCall) cast<CallInst>(result)->setTailCall();
-  if (CallingConv) cast<CallInst>(result)->setCallingConv(CallingConv);
-  if (signedArg) {
-    const Type* newTy = F->getReturnType()->getUnsignedVersion();
-    CastInst* final = new CastInst(result, newTy, "autoupgrade_uncast");
-    BB->getInstList().push_back(result);
-    result = final;
+    break;
   }
-  return result;
+
+  //  This may not belong here. This function is effectively being overloaded 
+  //  to both detect an intrinsic which needs upgrading, and to provide the 
+  //  upgraded form of the intrinsic. We should perhaps have two separate 
+  //  functions for this.
+  return false;
 }
 
-// UpgradeIntrinsicCall - In the BC reader, change a call to an intrinsic to be
-// a call to an upgraded intrinsic.  We may have to permute the order or promote
-// some arguments with a cast.
+bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
+  NewFn = 0;
+  bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
+
+  // Upgrade intrinsic attributes.  This does not change the function.
+  if (NewFn)
+    F = NewFn;
+  if (unsigned id = F->getIntrinsicID())
+    F->setAttributes(Intrinsic::getAttributes((Intrinsic::ID)id));
+  return Upgraded;
+}
+
+// UpgradeIntrinsicCall - Upgrade a call to an old intrinsic to be a call the 
+// upgraded intrinsic. All argument and return casting must be provided in 
+// order to seamlessly integrate with existing context.
 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
   Function *F = CI->getCalledFunction();
+  LLVMContext &C = CI->getContext();
+  ImmutableCallSite CS(CI);
 
-  const FunctionType *NewFnTy = NewFn->getFunctionType();
-  std::vector<Value*> Oprnds;
-  
-  unsigned *Permutation = getArgumentPermutation(F, NewFn);
-  unsigned N = NewFnTy->getNumParams();
+  assert(F && "CallInst has no function associated with it.");
 
-  if (Permutation) {
-    for (unsigned i = 0; i != N; ++i) {
-      unsigned p = Permutation[i];
-      
-      if (p) {
-        Value *V = CI->getOperand(p);
-        if (V->getType() != NewFnTy->getParamType(i))
-          V = CastArg(V, NewFnTy->getParamType(i));
-        Oprnds.push_back(V);
-      } else
-        Oprnds.push_back(UndefValue::get(NewFnTy->getParamType(i)));
+  if (!NewFn) {
+    // Get the Function's name.
+    const std::string& Name = F->getName();
+
+    // Upgrade ARM NEON intrinsics.
+    if (Name.compare(5, 9, "arm.neon.", 9) == 0) {
+      Instruction *NewI;
+      if (Name.compare(14, 7, "vmovls.", 7) == 0) {
+        NewI = new SExtInst(CI->getArgOperand(0), CI->getType(),
+                            "upgraded." + CI->getName(), CI);
+      } else if (Name.compare(14, 7, "vmovlu.", 7) == 0) {
+        NewI = new ZExtInst(CI->getArgOperand(0), CI->getType(),
+                            "upgraded." + CI->getName(), CI);
+
+      } else if (Name.compare(14, 4, "vadd", 4) == 0 ||
+                 Name.compare(14, 4, "vsub", 4) == 0) {
+        // Extend one (vaddw/vsubw) or both (vaddl/vsubl) operands.
+        Value *V0 = CI->getArgOperand(0);
+        Value *V1 = CI->getArgOperand(1);
+        if (Name.at(19) == 's') {
+          if (Name.at(18) == 'l')
+            V0 = new SExtInst(CI->getArgOperand(0), CI->getType(), "", CI);
+          V1 = new SExtInst(CI->getArgOperand(1), CI->getType(), "", CI);
+        } else {
+          assert(Name.at(19) == 'u' && "unexpected vadd/vsub intrinsic");
+          if (Name.at(18) == 'l')
+            V0 = new ZExtInst(CI->getArgOperand(0), CI->getType(), "", CI);
+          V1 = new ZExtInst(CI->getArgOperand(1), CI->getType(), "", CI);
+        }
+        if (Name.compare(14, 4, "vadd", 4) == 0)
+          NewI = BinaryOperator::CreateAdd(V0, V1,"upgraded."+CI->getName(),CI);
+        else
+          NewI = BinaryOperator::CreateSub(V0, V1,"upgraded."+CI->getName(),CI);
+
+      } else if (Name.compare(14, 6, "vmovn.", 6) == 0) {
+        NewI = new TruncInst(CI->getArgOperand(0), CI->getType(),
+                             "upgraded." + CI->getName(), CI);
+      } else {
+        llvm_unreachable("Unknown arm.neon function for CallInst upgrade.");
+      }
+      // Replace any uses of the old CallInst.
+      if (!CI->use_empty())
+        CI->replaceAllUsesWith(NewI);
+      CI->eraseFromParent();
+      return;
     }
-  } else if (N) {
-    assert(N == (CI->getNumOperands() - 1) &&
-           "Upgraded function needs permutation");
-    for (unsigned i = 0; i != N; ++i) {
-      Value *V = CI->getOperand(i + 1);
-      if (V->getType() != NewFnTy->getParamType(i))
-        V = CastArg(V, NewFnTy->getParamType(i));
-      Oprnds.push_back(V);
+
+    bool isLoadH = false, isLoadL = false, isMovL = false;
+    bool isMovSD = false, isShufPD = false;
+    bool isUnpckhPD = false, isUnpcklPD = false;
+    bool isPunpckhQPD = false, isPunpcklQPD = false;
+    if (F->getName() == "llvm.x86.sse2.loadh.pd")
+      isLoadH = true;
+    else if (F->getName() == "llvm.x86.sse2.loadl.pd")
+      isLoadL = true;
+    else if (F->getName() == "llvm.x86.sse2.movl.dq")
+      isMovL = true;
+    else if (F->getName() == "llvm.x86.sse2.movs.d")
+      isMovSD = true;
+    else if (F->getName() == "llvm.x86.sse2.shuf.pd")
+      isShufPD = true;
+    else if (F->getName() == "llvm.x86.sse2.unpckh.pd")
+      isUnpckhPD = true;
+    else if (F->getName() == "llvm.x86.sse2.unpckl.pd")
+      isUnpcklPD = true;
+    else if (F->getName() ==  "llvm.x86.sse2.punpckh.qdq")
+      isPunpckhQPD = true;
+    else if (F->getName() ==  "llvm.x86.sse2.punpckl.qdq")
+      isPunpcklQPD = true;
+
+    if (isLoadH || isLoadL || isMovL || isMovSD || isShufPD ||
+        isUnpckhPD || isUnpcklPD || isPunpckhQPD || isPunpcklQPD) {
+      std::vector<Constant*> Idxs;
+      Value *Op0 = CI->getArgOperand(0);
+      ShuffleVectorInst *SI = NULL;
+      if (isLoadH || isLoadL) {
+        Value *Op1 = UndefValue::get(Op0->getType());
+        Value *Addr = new BitCastInst(CI->getArgOperand(1), 
+                                  Type::getDoublePtrTy(C),
+                                      "upgraded.", CI);
+        Value *Load = new LoadInst(Addr, "upgraded.", false, 8, CI);
+        Value *Idx = ConstantInt::get(Type::getInt32Ty(C), 0);
+        Op1 = InsertElementInst::Create(Op1, Load, Idx, "upgraded.", CI);
+
+        if (isLoadH) {
+          Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 0));
+          Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
+        } else {
+          Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
+          Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
+        }
+        Value *Mask = ConstantVector::get(Idxs);
+        SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
+      } else if (isMovL) {
+        Constant *Zero = ConstantInt::get(Type::getInt32Ty(C), 0);
+        Idxs.push_back(Zero);
+        Idxs.push_back(Zero);
+        Idxs.push_back(Zero);
+        Idxs.push_back(Zero);
+        Value *ZeroV = ConstantVector::get(Idxs);
+
+        Idxs.clear(); 
+        Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 4));
+        Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 5));
+        Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
+        Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 3));
+        Value *Mask = ConstantVector::get(Idxs);
+        SI = new ShuffleVectorInst(ZeroV, Op0, Mask, "upgraded.", CI);
+      } else if (isMovSD ||
+                 isUnpckhPD || isUnpcklPD || isPunpckhQPD || isPunpcklQPD) {
+        Value *Op1 = CI->getArgOperand(1);
+        if (isMovSD) {
+          Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
+          Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
+        } else if (isUnpckhPD || isPunpckhQPD) {
+          Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
+          Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 3));
+        } else {
+          Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 0));
+          Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
+        }
+        Value *Mask = ConstantVector::get(Idxs);
+        SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
+      } else if (isShufPD) {
+        Value *Op1 = CI->getArgOperand(1);
+        unsigned MaskVal =
+                        cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
+        Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), MaskVal & 1));
+        Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C),
+                                               ((MaskVal >> 1) & 1)+2));
+        Value *Mask = ConstantVector::get(Idxs);
+        SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
+      }
+
+      assert(SI && "Unexpected!");
+
+      // Handle any uses of the old CallInst.
+      if (!CI->use_empty())
+        //  Replace all uses of the old call with the new cast which has the 
+        //  correct type.
+        CI->replaceAllUsesWith(SI);
+      
+      //  Clean up the old call now that it has been completely upgraded.
+      CI->eraseFromParent();
+    } else if (F->getName() == "llvm.x86.sse41.pmulld") {
+      // Upgrade this set of intrinsics into vector multiplies.
+      Instruction *Mul = BinaryOperator::CreateMul(CI->getArgOperand(0),
+                                                   CI->getArgOperand(1),
+                                                   CI->getName(),
+                                                   CI);
+      // Fix up all the uses with our new multiply.
+      if (!CI->use_empty())
+        CI->replaceAllUsesWith(Mul);
+        
+      // Remove upgraded multiply.
+      CI->eraseFromParent();
+    } else if (F->getName() == "llvm.x86.ssse3.palign.r") {
+      Value *Op1 = CI->getArgOperand(0);
+      Value *Op2 = CI->getArgOperand(1);
+      Value *Op3 = CI->getArgOperand(2);
+      unsigned shiftVal = cast<ConstantInt>(Op3)->getZExtValue();
+      Value *Rep;
+      IRBuilder<> Builder(C);
+      Builder.SetInsertPoint(CI->getParent(), CI);
+
+      // If palignr is shifting the pair of input vectors less than 9 bytes,
+      // emit a shuffle instruction.
+      if (shiftVal <= 8) {
+        const Type *IntTy = Type::getInt32Ty(C);
+        const Type *EltTy = Type::getInt8Ty(C);
+        const Type *VecTy = VectorType::get(EltTy, 8);
+        
+        Op2 = Builder.CreateBitCast(Op2, VecTy);
+        Op1 = Builder.CreateBitCast(Op1, VecTy);
+
+        llvm::SmallVector<llvm::Constant*, 8> Indices;
+        for (unsigned i = 0; i != 8; ++i)
+          Indices.push_back(ConstantInt::get(IntTy, shiftVal + i));
+
+        Value *SV = ConstantVector::get(Indices.begin(), Indices.size());
+        Rep = Builder.CreateShuffleVector(Op2, Op1, SV, "palignr");
+        Rep = Builder.CreateBitCast(Rep, F->getReturnType());
+      }
+
+      // If palignr is shifting the pair of input vectors more than 8 but less
+      // than 16 bytes, emit a logical right shift of the destination.
+      else if (shiftVal < 16) {
+        // MMX has these as 1 x i64 vectors for some odd optimization reasons.
+        const Type *EltTy = Type::getInt64Ty(C);
+        const Type *VecTy = VectorType::get(EltTy, 1);
+
+        Op1 = Builder.CreateBitCast(Op1, VecTy, "cast");
+        Op2 = ConstantInt::get(VecTy, (shiftVal-8) * 8);
+
+        // create i32 constant
+        Function *I =
+          Intrinsic::getDeclaration(F->getParent(), Intrinsic::x86_mmx_psrl_q);
+        Rep = Builder.CreateCall2(I, Op1, Op2, "palignr");
+      }
+
+      // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
+      else {
+        Rep = Constant::getNullValue(F->getReturnType());
+      }
+      
+      // Replace any uses with our new instruction.
+      if (!CI->use_empty())
+        CI->replaceAllUsesWith(Rep);
+        
+      // Remove upgraded instruction.
+      CI->eraseFromParent();
+      
+    } else if (F->getName() == "llvm.x86.ssse3.palign.r.128") {
+      Value *Op1 = CI->getArgOperand(0);
+      Value *Op2 = CI->getArgOperand(1);
+      Value *Op3 = CI->getArgOperand(2);
+      unsigned shiftVal = cast<ConstantInt>(Op3)->getZExtValue();
+      Value *Rep;
+      IRBuilder<> Builder(C);
+      Builder.SetInsertPoint(CI->getParent(), CI);
+
+      // If palignr is shifting the pair of input vectors less than 17 bytes,
+      // emit a shuffle instruction.
+      if (shiftVal <= 16) {
+        const Type *IntTy = Type::getInt32Ty(C);
+        const Type *EltTy = Type::getInt8Ty(C);
+        const Type *VecTy = VectorType::get(EltTy, 16);
+        
+        Op2 = Builder.CreateBitCast(Op2, VecTy);
+        Op1 = Builder.CreateBitCast(Op1, VecTy);
+
+        llvm::SmallVector<llvm::Constant*, 16> Indices;
+        for (unsigned i = 0; i != 16; ++i)
+          Indices.push_back(ConstantInt::get(IntTy, shiftVal + i));
+
+        Value *SV = ConstantVector::get(Indices.begin(), Indices.size());
+        Rep = Builder.CreateShuffleVector(Op2, Op1, SV, "palignr");
+        Rep = Builder.CreateBitCast(Rep, F->getReturnType());
+      }
+
+      // If palignr is shifting the pair of input vectors more than 16 but less
+      // than 32 bytes, emit a logical right shift of the destination.
+      else if (shiftVal < 32) {
+        const Type *EltTy = Type::getInt64Ty(C);
+        const Type *VecTy = VectorType::get(EltTy, 2);
+        const Type *IntTy = Type::getInt32Ty(C);
+
+        Op1 = Builder.CreateBitCast(Op1, VecTy, "cast");
+        Op2 = ConstantInt::get(IntTy, (shiftVal-16) * 8);
+
+        // create i32 constant
+        Function *I =
+          Intrinsic::getDeclaration(F->getParent(), Intrinsic::x86_sse2_psrl_dq);
+        Rep = Builder.CreateCall2(I, Op1, Op2, "palignr");
+      }
+
+      // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
+      else {
+        Rep = Constant::getNullValue(F->getReturnType());
+      }
+      
+      // Replace any uses with our new instruction.
+      if (!CI->use_empty())
+        CI->replaceAllUsesWith(Rep);
+        
+      // Remove upgraded instruction.
+      CI->eraseFromParent();
+      
+    } else {
+      llvm_unreachable("Unknown function for CallInst upgrade.");
     }
+    return;
   }
-  
-  bool NewIsVoid = NewFn->getReturnType() == Type::VoidTy;
-  
-  CallInst *NewCI = new CallInst(NewFn, Oprnds,
-                                 NewIsVoid ? "" : CI->getName(),
-                                 CI);
-  NewCI->setTailCall(CI->isTailCall());
-  NewCI->setCallingConv(CI->getCallingConv());
-  
-  if (!CI->use_empty()) {
-    if (NewIsVoid) {
-      CI->replaceAllUsesWith(UndefValue::get(CI->getType()));
-    } else {
-      Instruction *RetVal = NewCI;
+
+  switch (NewFn->getIntrinsicID()) {
+  default:  llvm_unreachable("Unknown function for CallInst upgrade.");
+  case Intrinsic::arm_neon_vld1:
+  case Intrinsic::arm_neon_vld2:
+  case Intrinsic::arm_neon_vld3:
+  case Intrinsic::arm_neon_vld4:
+  case Intrinsic::arm_neon_vst1:
+  case Intrinsic::arm_neon_vst2:
+  case Intrinsic::arm_neon_vst3:
+  case Intrinsic::arm_neon_vst4:
+  case Intrinsic::arm_neon_vld2lane:
+  case Intrinsic::arm_neon_vld3lane:
+  case Intrinsic::arm_neon_vld4lane:
+  case Intrinsic::arm_neon_vst2lane:
+  case Intrinsic::arm_neon_vst3lane:
+  case Intrinsic::arm_neon_vst4lane: {
+    // Add a default alignment argument of 1.
+    SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
+    Operands.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
+    CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
+                                       CI->getName(), CI);
+    NewCI->setTailCall(CI->isTailCall());
+    NewCI->setCallingConv(CI->getCallingConv());
+
+    //  Handle any uses of the old CallInst.
+    if (!CI->use_empty())
+      //  Replace all uses of the old call with the new cast which has the 
+      //  correct type.
+      CI->replaceAllUsesWith(NewCI);
+    
+    //  Clean up the old call now that it has been completely upgraded.
+    CI->eraseFromParent();
+    break;
+  }        
+
+  case Intrinsic::x86_mmx_psll_d:
+  case Intrinsic::x86_mmx_psll_q:
+  case Intrinsic::x86_mmx_psll_w:
+  case Intrinsic::x86_mmx_psra_d:
+  case Intrinsic::x86_mmx_psra_w:
+  case Intrinsic::x86_mmx_psrl_d:
+  case Intrinsic::x86_mmx_psrl_q:
+  case Intrinsic::x86_mmx_psrl_w: {
+    Value *Operands[2];
+    
+    Operands[0] = CI->getArgOperand(0);
+    
+    // Cast the second parameter to the correct type.
+    BitCastInst *BC = new BitCastInst(CI->getArgOperand(1), 
+                                      NewFn->getFunctionType()->getParamType(1),
+                                      "upgraded.", CI);
+    Operands[1] = BC;
+    
+    //  Construct a new CallInst
+    CallInst *NewCI = CallInst::Create(NewFn, Operands, Operands+2, 
+                                       "upgraded."+CI->getName(), CI);
+    NewCI->setTailCall(CI->isTailCall());
+    NewCI->setCallingConv(CI->getCallingConv());
+    
+    //  Handle any uses of the old CallInst.
+    if (!CI->use_empty())
+      //  Replace all uses of the old call with the new cast which has the 
+      //  correct type.
+      CI->replaceAllUsesWith(NewCI);
+    
+    //  Clean up the old call now that it has been completely upgraded.
+    CI->eraseFromParent();
+    break;
+  }        
+  case Intrinsic::ctlz:
+  case Intrinsic::ctpop:
+  case Intrinsic::cttz: {
+    //  Build a small vector of the original arguments.
+    SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
+
+    //  Construct a new CallInst
+    CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
+                                       "upgraded."+CI->getName(), CI);
+    NewCI->setTailCall(CI->isTailCall());
+    NewCI->setCallingConv(CI->getCallingConv());
+
+    //  Handle any uses of the old CallInst.
+    if (!CI->use_empty()) {
+      //  Check for sign extend parameter attributes on the return values.
+      bool SrcSExt = NewFn->getAttributes().paramHasAttr(0, Attribute::SExt);
+      bool DestSExt = F->getAttributes().paramHasAttr(0, Attribute::SExt);
       
-      if (F->getReturnType() != NewFn->getReturnType()) {
-        RetVal = new CastInst(NewCI, F->getReturnType(), 
-                              NewCI->getName(), CI);
-        NewCI->moveBefore(RetVal);
+      //  Construct an appropriate cast from the new return type to the old.
+      CastInst *RetCast = CastInst::Create(
+                            CastInst::getCastOpcode(NewCI, SrcSExt,
+                                                    F->getReturnType(),
+                                                    DestSExt),
+                            NewCI, F->getReturnType(),
+                            NewCI->getName(), CI);
+      NewCI->moveBefore(RetCast);
+
+      //  Replace all uses of the old call with the new cast which has the 
+      //  correct type.
+      CI->replaceAllUsesWith(RetCast);
+    }
+
+    //  Clean up the old call now that it has been completely upgraded.
+    CI->eraseFromParent();
+  }
+  break;
+  case Intrinsic::eh_selector:
+  case Intrinsic::eh_typeid_for: {
+    // Only the return type changed.
+    SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
+    CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
+                                       "upgraded." + CI->getName(), CI);
+    NewCI->setTailCall(CI->isTailCall());
+    NewCI->setCallingConv(CI->getCallingConv());
+
+    //  Handle any uses of the old CallInst.
+    if (!CI->use_empty()) {
+      //  Construct an appropriate cast from the new return type to the old.
+      CastInst *RetCast =
+        CastInst::Create(CastInst::getCastOpcode(NewCI, true,
+                                                 F->getReturnType(), true),
+                         NewCI, F->getReturnType(), NewCI->getName(), CI);
+      CI->replaceAllUsesWith(RetCast);
+    }
+    CI->eraseFromParent();
+  }
+  break;
+  case Intrinsic::memcpy:
+  case Intrinsic::memmove:
+  case Intrinsic::memset: {
+    // Add isVolatile
+    const llvm::Type *I1Ty = llvm::Type::getInt1Ty(CI->getContext());
+    Value *Operands[5] = { CI->getArgOperand(0), CI->getArgOperand(1),
+                           CI->getArgOperand(2), CI->getArgOperand(3),
+                           llvm::ConstantInt::get(I1Ty, 0) };
+    CallInst *NewCI = CallInst::Create(NewFn, Operands, Operands+5,
+                                       CI->getName(), CI);
+    NewCI->setTailCall(CI->isTailCall());
+    NewCI->setCallingConv(CI->getCallingConv());
+    //  Handle any uses of the old CallInst.
+    if (!CI->use_empty())
+      //  Replace all uses of the old call with the new cast which has the 
+      //  correct type.
+      CI->replaceAllUsesWith(NewCI);
+    
+    //  Clean up the old call now that it has been completely upgraded.
+    CI->eraseFromParent();
+    break;
+  }
+  }
+}
+
+// This tests each Function to determine if it needs upgrading. When we find 
+// one we are interested in, we then upgrade all calls to reflect the new 
+// function.
+void llvm::UpgradeCallsToIntrinsic(Function* F) {
+  assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
+
+  // Upgrade the function and check if it is a totaly new function.
+  Function* NewFn;
+  if (UpgradeIntrinsicFunction(F, NewFn)) {
+    if (NewFn != F) {
+      // Replace all uses to the old function with the new one if necessary.
+      for (Value::use_iterator UI = F->use_begin(), UE = F->use_end();
+           UI != UE; ) {
+        if (CallInst* CI = dyn_cast<CallInst>(*UI++))
+          UpgradeIntrinsicCall(CI, NewFn);
       }
-      
-      CI->replaceAllUsesWith(RetVal);
+      // Remove old function, no longer used, from the module.
+      F->eraseFromParent();
     }
   }
-  CI->eraseFromParent();
 }
 
-bool llvm::UpgradeCallsToIntrinsic(Function* F) {
-  if (Function* NewFn = UpgradeIntrinsicFunction(F)) {
-    for (Value::use_iterator UI = F->use_begin(), UE = F->use_end();
-         UI != UE; ) {
-      if (CallInst* CI = dyn_cast<CallInst>(*UI++)) 
-        UpgradeIntrinsicCall(CI, NewFn);
+/// This function strips all debug info intrinsics, except for llvm.dbg.declare.
+/// If an llvm.dbg.declare intrinsic is invalid, then this function simply
+/// strips that use.
+void llvm::CheckDebugInfoIntrinsics(Module *M) {
+
+
+  if (Function *FuncStart = M->getFunction("llvm.dbg.func.start")) {
+    while (!FuncStart->use_empty()) {
+      CallInst *CI = cast<CallInst>(FuncStart->use_back());
+      CI->eraseFromParent();
+    }
+    FuncStart->eraseFromParent();
+  }
+  
+  if (Function *StopPoint = M->getFunction("llvm.dbg.stoppoint")) {
+    while (!StopPoint->use_empty()) {
+      CallInst *CI = cast<CallInst>(StopPoint->use_back());
+      CI->eraseFromParent();
+    }
+    StopPoint->eraseFromParent();
+  }
+
+  if (Function *RegionStart = M->getFunction("llvm.dbg.region.start")) {
+    while (!RegionStart->use_empty()) {
+      CallInst *CI = cast<CallInst>(RegionStart->use_back());
+      CI->eraseFromParent();
+    }
+    RegionStart->eraseFromParent();
+  }
+
+  if (Function *RegionEnd = M->getFunction("llvm.dbg.region.end")) {
+    while (!RegionEnd->use_empty()) {
+      CallInst *CI = cast<CallInst>(RegionEnd->use_back());
+      CI->eraseFromParent();
+    }
+    RegionEnd->eraseFromParent();
+  }
+  
+  if (Function *Declare = M->getFunction("llvm.dbg.declare")) {
+    if (!Declare->use_empty()) {
+      DbgDeclareInst *DDI = cast<DbgDeclareInst>(Declare->use_back());
+      if (!isa<MDNode>(DDI->getArgOperand(0)) ||
+          !isa<MDNode>(DDI->getArgOperand(1))) {
+        while (!Declare->use_empty()) {
+          CallInst *CI = cast<CallInst>(Declare->use_back());
+          CI->eraseFromParent();
+        }
+        Declare->eraseFromParent();
+      }
     }
-    if (NewFn != F)
-      F->eraseFromParent();
-    return true;
   }
-  return false;
 }