Massive rewrite of MMX:

The x86_mmx type is used for MMX intrinsics, parameters and
return values where these use MMX registers, and is also
supported in load, store, and bitcast.

Only the above operations generate MMX instructions, and optimizations
do not operate on or produce MMX intrinsics. 

MMX-sized vectors <2 x i32> etc. are lowered to XMM or split into
smaller pieces.  Optimizations may occur on these forms and the
result casted back to x86_mmx, provided the result feeds into a
previous existing x86_mmx operation.

The point of all this is prevent optimizations from introducing
MMX operations, which is unsafe due to the EMMS problem.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115243 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 506 insertions, 57 deletions

diff --git a/lib/VMCore/AutoUpgrade.cpp b/lib/VMCore/AutoUpgrade.cpp
index 9330e14..5b5955d 100644
--- a/lib/VMCore/AutoUpgrade.cpp
+++ b/lib/VMCore/AutoUpgrade.cpp
@@ -288,37 +288,224 @@ static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
     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)
+    // x86_mmx instead of a v1i64, v2i32, v4i16, or v8i8.
+    if (Name.compare(5, 8, "x86.mmx.", 8) == 0) {
+      const Type *X86_MMXTy = VectorType::getX86_MMXTy(FTy->getContext());
+
+      if (Name.compare(13, 4, "padd", 4) == 0   ||
+          Name.compare(13, 4, "psub", 4) == 0   ||
+          Name.compare(13, 4, "pmul", 4) == 0   ||
+          Name.compare(13, 5, "pmadd", 5) == 0  ||
+          Name.compare(13, 4, "pand", 4) == 0   ||
+          Name.compare(13, 3, "por", 3) == 0    ||
+          Name.compare(13, 4, "pxor", 4) == 0   ||
+          Name.compare(13, 4, "pavg", 4) == 0   ||
+          Name.compare(13, 4, "pmax", 4) == 0   ||
+          Name.compare(13, 4, "pmin", 4) == 0   ||
+          Name.compare(13, 4, "psad", 4) == 0   ||
+          Name.compare(13, 4, "psll", 4) == 0   ||
+          Name.compare(13, 4, "psrl", 4) == 0   ||
+          Name.compare(13, 4, "psra", 4) == 0   ||
+          Name.compare(13, 4, "pack", 4) == 0   ||
+          Name.compare(13, 6, "punpck", 6) == 0 ||
+          Name.compare(13, 4, "pcmp", 4) == 0) {
+        assert(FTy->getNumParams() == 2 && "MMX intrinsic takes 2 args!");
+        const Type *SecondParamTy = X86_MMXTy;
+
+        if (Name.compare(13, 5, "pslli", 5) == 0 ||
+            Name.compare(13, 5, "psrli", 5) == 0 ||
+            Name.compare(13, 5, "psrai", 5) == 0)
+          SecondParamTy = FTy->getParamType(1);
+
+        // Don't do anything if it has the correct types.
+        if (FTy->getReturnType() == X86_MMXTy &&
+            FTy->getParamType(0) == X86_MMXTy &&
+            FTy->getParamType(1) == SecondParamTy)
+          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("");
+
+        // 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, 
+                                                      X86_MMXTy, X86_MMXTy,
+                                                      SecondParamTy, (Type*)0));
+        return true;
+      }
+
+      if (Name.compare(13, 8, "maskmovq", 8) == 0) {
+        // Don't do anything if it has the correct types.
+        if (FTy->getParamType(0) == X86_MMXTy &&
+            FTy->getParamType(1) == X86_MMXTy)
+          break;
+
+        F->setName("");
+        NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                      FTy->getReturnType(),
+                                                      X86_MMXTy,
+                                                      X86_MMXTy,
+                                                      FTy->getParamType(2),
+                                                      (Type*)0));
+        return true;
+      }
+
+      if (Name.compare(13, 8, "pmovmskb", 8) == 0) {
+        if (FTy->getParamType(0) == X86_MMXTy)
+          break;
+
+        F->setName("");
+        NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                      FTy->getReturnType(),
+                                                      X86_MMXTy,
+                                                      (Type*)0));
+        return true;
+      }
+
+      if (Name.compare(13, 5, "movnt", 5) == 0) {
+        if (FTy->getParamType(1) == X86_MMXTy)
+          break;
+
+        F->setName("");
+        NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                      FTy->getReturnType(),
+                                                      FTy->getParamType(0),
+                                                      X86_MMXTy,
+                                                      (Type*)0));
+        return true;
+      }
+
+      if (Name.compare(13, 7, "palignr", 7) == 0) {
+        if (FTy->getReturnType() == X86_MMXTy &&
+            FTy->getParamType(0) == X86_MMXTy &&
+            FTy->getParamType(1) == X86_MMXTy)
+          break;
+
+        F->setName("");
+        NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                      X86_MMXTy,
+                                                      X86_MMXTy,
+                                                      X86_MMXTy,
+                                                      FTy->getParamType(2),
+                                                      (Type*)0));
+        return true;
+      }
+
+      if (Name.compare(13, 5, "pextr", 5) == 0) {
+        if (FTy->getParamType(0) == X86_MMXTy)
+          break;
+
+        F->setName("");
+        NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                      FTy->getReturnType(),
+                                                      X86_MMXTy,
+                                                      FTy->getParamType(1),
+                                                      (Type*)0));
+        return true;
+      }
+
+      if (Name.compare(13, 5, "pinsr", 5) == 0) {
+        if (FTy->getReturnType() == X86_MMXTy &&
+            FTy->getParamType(0) == X86_MMXTy)
+          break;
+
+        F->setName("");
+        NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                      X86_MMXTy,
+                                                      X86_MMXTy,
+                                                      FTy->getParamType(1),
+                                                      FTy->getParamType(2),
+                                                      (Type*)0));
+        return true;
+      }
+
+      if (Name.compare(13, 12, "cvtsi32.si64", 12) == 0) {
+        if (FTy->getReturnType() == X86_MMXTy)
+          break;
+
+        F->setName("");
+        NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                      X86_MMXTy,
+                                                      FTy->getParamType(0),
+                                                      (Type*)0));
+        return true;
+      }
+
+      if (Name.compare(13, 12, "cvtsi64.si32", 12) == 0) {
+        if (FTy->getParamType(0) == X86_MMXTy)
+          break;
+
+        F->setName("");
+        NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                      FTy->getReturnType(),
+                                                      X86_MMXTy,
+                                                      (Type*)0));
+        return true;
+      }
+
+      if (Name.compare(13, 8, "vec.init", 8) == 0) {
+        if (FTy->getReturnType() == X86_MMXTy)
+          break;
+
+        F->setName("");
+
+        if (Name.compare(21, 2, ".b", 2) == 0)
+          NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                        X86_MMXTy,
+                                                        FTy->getParamType(0),
+                                                        FTy->getParamType(1),
+                                                        FTy->getParamType(2),
+                                                        FTy->getParamType(3),
+                                                        FTy->getParamType(4),
+                                                        FTy->getParamType(5),
+                                                        FTy->getParamType(6),
+                                                        FTy->getParamType(7),
+                                                        (Type*)0));
+        else if (Name.compare(21, 2, ".w", 2) == 0)
+          NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                        X86_MMXTy,
+                                                        FTy->getParamType(0),
+                                                        FTy->getParamType(1),
+                                                        FTy->getParamType(2),
+                                                        FTy->getParamType(3),
+                                                        (Type*)0));
+        else if (Name.compare(21, 2, ".d", 2) == 0)
+          NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                        X86_MMXTy,
+                                                        FTy->getParamType(0),
+                                                        FTy->getParamType(1),
+                                                        (Type*)0));
+        return true;
+      }
+
+
+      if (Name.compare(13, 9, "vec.ext.d", 9) == 0) {
+        if (FTy->getReturnType() == X86_MMXTy &&
+            FTy->getParamType(0) == X86_MMXTy)
+          break;
+
+        F->setName("");
+        NewFn = cast<Function>(M->getOrInsertFunction(Name, 
+                                                      X86_MMXTy,
+                                                      X86_MMXTy,
+                                                      FTy->getParamType(1),
+                                                      (Type*)0));
+        return true;
+      }
+
+      if (Name.compare(13, 9, "emms", 4) == 0 ||
+          Name.compare(13, 9, "femms", 5) == 0) {
+        NewFn = 0;
         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;
+      // We really shouldn't get here ever.
+      assert(0 && "Invalid MMX intrinsic!");
+      break;
     } 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 ||
@@ -432,6 +619,38 @@ static Instruction *CallVABD(CallInst *CI, Value *Arg0, Value *Arg1) {
                           "upgraded."+CI->getName(), CI);
 }
 
+/// ConstructNewCallInst - Construct a new CallInst with the signature of NewFn.
+static void ConstructNewCallInst(Function *NewFn, CallInst *OldCI,
+                                 Value **Operands, unsigned NumOps,
+                                 bool AssignName = true) {
+  // Construct a new CallInst.
+  CallInst *NewCI =
+    CallInst::Create(NewFn, Operands, Operands + NumOps,
+                     AssignName ? "upgraded." + OldCI->getName() : "", OldCI);
+
+  NewCI->setTailCall(OldCI->isTailCall());
+  NewCI->setCallingConv(OldCI->getCallingConv());
+
+  // Handle any uses of the old CallInst.
+  if (!OldCI->use_empty()) {
+    // If the type has changed, add a cast.
+    Instruction *I = OldCI;
+    if (OldCI->getType() != NewCI->getType()) {
+      Function *OldFn = OldCI->getCalledFunction();
+      CastInst *RetCast =
+        CastInst::Create(CastInst::getCastOpcode(NewCI, true,
+                                                 OldFn->getReturnType(), true),
+                         NewCI, OldFn->getReturnType(), NewCI->getName(),OldCI);
+      I = RetCast;
+    }
+    // Replace all uses of the old call with the new cast which has the 
+    // correct type.
+    OldCI->replaceAllUsesWith(I);
+  }
+  // Clean up the old call now that it has been completely upgraded.
+  OldCI->eraseFromParent();
+}
+
 // 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.
@@ -759,40 +978,246 @@ void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
     break;
   }        
 
+  case Intrinsic::x86_mmx_padd_b:
+  case Intrinsic::x86_mmx_padd_w:
+  case Intrinsic::x86_mmx_padd_d:
+  case Intrinsic::x86_mmx_padd_q:
+  case Intrinsic::x86_mmx_padds_b:
+  case Intrinsic::x86_mmx_padds_w:
+  case Intrinsic::x86_mmx_paddus_b:
+  case Intrinsic::x86_mmx_paddus_w:
+  case Intrinsic::x86_mmx_psub_b:
+  case Intrinsic::x86_mmx_psub_w:
+  case Intrinsic::x86_mmx_psub_d:
+  case Intrinsic::x86_mmx_psub_q:
+  case Intrinsic::x86_mmx_psubs_b:
+  case Intrinsic::x86_mmx_psubs_w:
+  case Intrinsic::x86_mmx_psubus_b:
+  case Intrinsic::x86_mmx_psubus_w:
+  case Intrinsic::x86_mmx_pmulh_w:
+  case Intrinsic::x86_mmx_pmull_w:
+  case Intrinsic::x86_mmx_pmulhu_w:
+  case Intrinsic::x86_mmx_pmulu_dq:
+  case Intrinsic::x86_mmx_pmadd_wd:
+  case Intrinsic::x86_mmx_pand:
+  case Intrinsic::x86_mmx_pandn:
+  case Intrinsic::x86_mmx_por:
+  case Intrinsic::x86_mmx_pxor:
+  case Intrinsic::x86_mmx_pavg_b:
+  case Intrinsic::x86_mmx_pavg_w:
+  case Intrinsic::x86_mmx_pmaxu_b:
+  case Intrinsic::x86_mmx_pmaxs_w:
+  case Intrinsic::x86_mmx_pminu_b:
+  case Intrinsic::x86_mmx_pmins_w:
+  case Intrinsic::x86_mmx_psad_bw:
+  case Intrinsic::x86_mmx_psll_w:
   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_pslli_w:
+  case Intrinsic::x86_mmx_pslli_d:
+  case Intrinsic::x86_mmx_pslli_q:
+  case Intrinsic::x86_mmx_psrl_w:
   case Intrinsic::x86_mmx_psrl_d:
   case Intrinsic::x86_mmx_psrl_q:
-  case Intrinsic::x86_mmx_psrl_w: {
+  case Intrinsic::x86_mmx_psrli_w:
+  case Intrinsic::x86_mmx_psrli_d:
+  case Intrinsic::x86_mmx_psrli_q:
+  case Intrinsic::x86_mmx_psra_w:
+  case Intrinsic::x86_mmx_psra_d:
+  case Intrinsic::x86_mmx_psrai_w:
+  case Intrinsic::x86_mmx_psrai_d:
+  case Intrinsic::x86_mmx_packsswb:
+  case Intrinsic::x86_mmx_packssdw:
+  case Intrinsic::x86_mmx_packuswb:
+  case Intrinsic::x86_mmx_punpckhbw:
+  case Intrinsic::x86_mmx_punpckhwd:
+  case Intrinsic::x86_mmx_punpckhdq:
+  case Intrinsic::x86_mmx_punpcklbw:
+  case Intrinsic::x86_mmx_punpcklwd:
+  case Intrinsic::x86_mmx_punpckldq:
+  case Intrinsic::x86_mmx_pcmpeq_b:
+  case Intrinsic::x86_mmx_pcmpeq_w:
+  case Intrinsic::x86_mmx_pcmpeq_d:
+  case Intrinsic::x86_mmx_pcmpgt_b:
+  case Intrinsic::x86_mmx_pcmpgt_w:
+  case Intrinsic::x86_mmx_pcmpgt_d: {
     Value *Operands[2];
     
+    // Cast the operand to the X86 MMX type.
+    Operands[0] = new BitCastInst(CI->getArgOperand(0), 
+                                  NewFn->getFunctionType()->getParamType(0),
+                                  "upgraded.", CI);
+
+    switch (NewFn->getIntrinsicID()) {
+    default:
+      // Cast to the X86 MMX type.
+      Operands[1] = new BitCastInst(CI->getArgOperand(1), 
+                                    NewFn->getFunctionType()->getParamType(1),
+                                    "upgraded.", CI);
+      break;
+    case Intrinsic::x86_mmx_pslli_w:
+    case Intrinsic::x86_mmx_pslli_d:
+    case Intrinsic::x86_mmx_pslli_q:
+    case Intrinsic::x86_mmx_psrli_w:
+    case Intrinsic::x86_mmx_psrli_d:
+    case Intrinsic::x86_mmx_psrli_q:
+    case Intrinsic::x86_mmx_psrai_w:
+    case Intrinsic::x86_mmx_psrai_d:
+      // These take an i32 as their second parameter.
+      Operands[1] = CI->getArgOperand(1);
+      break;
+    }
+
+    ConstructNewCallInst(NewFn, CI, Operands, 2);
+    break;
+  }
+  case Intrinsic::x86_mmx_maskmovq: {
+    Value *Operands[3];
+
+    // Cast the operands to the X86 MMX type.
+    Operands[0] = new BitCastInst(CI->getArgOperand(0), 
+                                  NewFn->getFunctionType()->getParamType(0),
+                                  "upgraded.", CI);
+    Operands[1] = new BitCastInst(CI->getArgOperand(1), 
+                                  NewFn->getFunctionType()->getParamType(1),
+                                  "upgraded.", CI);
+    Operands[2] = CI->getArgOperand(2);
+
+    ConstructNewCallInst(NewFn, CI, Operands, 3, false);
+    break;
+  }
+  case Intrinsic::x86_mmx_pmovmskb: {
+    Value *Operands[1];
+
+    // Cast the operand to the X86 MMX type.
+    Operands[0] = new BitCastInst(CI->getArgOperand(0), 
+                                  NewFn->getFunctionType()->getParamType(0),
+                                  "upgraded.", CI);
+
+    ConstructNewCallInst(NewFn, CI, Operands, 1);
+    break;
+  }
+  case Intrinsic::x86_mmx_movnt_dq: {
+    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();
+
+    // Cast the operand to the X86 MMX type.
+    Operands[1] = new BitCastInst(CI->getArgOperand(1),
+                                  NewFn->getFunctionType()->getParamType(1),
+                                  "upgraded.", CI);
+
+    ConstructNewCallInst(NewFn, CI, Operands, 2, false);
     break;
-  }        
+  }
+  case Intrinsic::x86_mmx_palignr_b: {
+    Value *Operands[3];
+
+    // Cast the operands to the X86 MMX type.
+    Operands[0] = new BitCastInst(CI->getArgOperand(0),
+                                  NewFn->getFunctionType()->getParamType(0),
+                                  "upgraded.", CI);
+    Operands[1] = new BitCastInst(CI->getArgOperand(1),
+                                  NewFn->getFunctionType()->getParamType(1),
+                                  "upgraded.", CI);
+    Operands[2] = CI->getArgOperand(2);
+
+    ConstructNewCallInst(NewFn, CI, Operands, 3);
+    break;
+  }
+  case Intrinsic::x86_mmx_pextr_w: {
+    Value *Operands[2];
+
+    // Cast the operands to the X86 MMX type.
+    Operands[0] = new BitCastInst(CI->getArgOperand(0),
+                                  NewFn->getFunctionType()->getParamType(0),
+                                  "upgraded.", CI);
+    Operands[1] = CI->getArgOperand(1);
+
+    ConstructNewCallInst(NewFn, CI, Operands, 2);
+    break;
+  }
+  case Intrinsic::x86_mmx_pinsr_w: {
+    Value *Operands[3];
+
+    // Cast the operands to the X86 MMX type.
+    Operands[0] = new BitCastInst(CI->getArgOperand(0),
+                                  NewFn->getFunctionType()->getParamType(0),
+                                  "upgraded.", CI);
+    Operands[1] = CI->getArgOperand(1);
+    Operands[2] = CI->getArgOperand(2);
+
+    ConstructNewCallInst(NewFn, CI, Operands, 3);
+    break;
+  }
+#if 0
+  case Intrinsic::x86_mmx_cvtsi32_si64: {
+    // The return type needs to be changed.
+    Value *Operands[1];
+    Operands[0] = CI->getArgOperand(0);
+    ConstructNewCallInst(NewFn, CI, Operands, 1);
+    break;
+  }
+  case Intrinsic::x86_mmx_cvtsi64_si32: {
+    Value *Operands[1];
+
+    // Cast the operand to the X86 MMX type.
+    Operands[0] = new BitCastInst(CI->getArgOperand(0),
+                                  NewFn->getFunctionType()->getParamType(0),
+                                  "upgraded.", CI);
+
+    ConstructNewCallInst(NewFn, CI, Operands, 1);
+    break;
+  }
+  case Intrinsic::x86_mmx_vec_init_b:
+  case Intrinsic::x86_mmx_vec_init_w:
+  case Intrinsic::x86_mmx_vec_init_d: {
+    // The return type needs to be changed.
+    Value *Operands[8];
+    unsigned NumOps = 0;
+
+    switch (NewFn->getIntrinsicID()) {
+    default: break;
+    case Intrinsic::x86_mmx_vec_init_b: NumOps = 8; break;
+    case Intrinsic::x86_mmx_vec_init_w: NumOps = 4; break;
+    case Intrinsic::x86_mmx_vec_init_d: NumOps = 2; break;
+    }
+
+    switch (NewFn->getIntrinsicID()) {
+    default: break;
+    case Intrinsic::x86_mmx_vec_init_b:
+      Operands[7] = CI->getArgOperand(7);
+      Operands[6] = CI->getArgOperand(6);
+      Operands[5] = CI->getArgOperand(5);
+      Operands[4] = CI->getArgOperand(4);
+      // FALLTHRU
+    case Intrinsic::x86_mmx_vec_init_w:
+      Operands[3] = CI->getArgOperand(3);
+      Operands[2] = CI->getArgOperand(2);
+      // FALLTHRU
+    case Intrinsic::x86_mmx_vec_init_d:
+      Operands[1] = CI->getArgOperand(1);
+      Operands[0] = CI->getArgOperand(0);
+      break;
+    }
+
+    ConstructNewCallInst(NewFn, CI, Operands, NumOps);
+    break;
+  }
+  case Intrinsic::x86_mmx_vec_ext_d: {
+    Value *Operands[2];
+
+    // Cast the operand to the X86 MMX type.
+    Operands[0] = new BitCastInst(CI->getArgOperand(0),
+                                  NewFn->getFunctionType()->getParamType(0),
+                                  "upgraded.", CI);
+    Operands[1] = CI->getArgOperand(1);
+
+    ConstructNewCallInst(NewFn, CI, Operands, 2);
+    break;
+  }
+#endif
+
   case Intrinsic::ctlz:
   case Intrinsic::ctpop:
   case Intrinsic::cttz: {
diff --git a/lib/VMCore/Instructions.cpp b/lib/VMCore/Instructions.cpp
index 7f8df58..754d992 100644
--- a/lib/VMCore/Instructions.cpp
+++ b/lib/VMCore/Instructions.cpp
@@ -2360,6 +2360,8 @@ bool CastInst::isCastable(const Type *SrcTy, const Type *DestTy) {
     } else {                                    // Casting from something else
       return false;
     }
+  } else if (DestTy->isX86_MMXTy()) {     
+    return SrcBits == 64;
   } else {                                      // Casting to something else
     return false;
   }
@@ -2441,6 +2443,10 @@ CastInst::getCastOpcode(
       return BitCast;                             // vector -> vector
     } else if (DestPTy->getBitWidth() == SrcBits) {
       return BitCast;                               // float/int -> vector
+    } else if (SrcTy->isX86_MMXTy()) {
+      assert(DestPTy->getBitWidth()==64 &&
+             "Casting X86_MMX to vector of wrong width");
+      return BitCast;                             // MMX to 64-bit vector
     } else {
       assert(!"Illegal cast to vector (wrong type or size)");
     }
@@ -2452,6 +2458,14 @@ CastInst::getCastOpcode(
     } else {
       assert(!"Casting pointer to other than pointer or int");
     }
+  } else if (DestTy->isX86_MMXTy()) {
+    if (const VectorType *SrcPTy = dyn_cast<VectorType>(SrcTy)) {
+      assert(SrcPTy->getBitWidth()==64 &&
+             "Casting vector of wrong width to X86_MMX");
+      return BitCast;                               // 64-bit vector to MMX
+    } else {
+      assert(!"Illegal cast to X86_MMX");
+    }
   } else {
     assert(!"Casting to type that is not first-class");
   }
diff --git a/lib/VMCore/Type.cpp b/lib/VMCore/Type.cpp
index 3704999..ba50e07 100644
--- a/lib/VMCore/Type.cpp
+++ b/lib/VMCore/Type.cpp
@@ -173,10 +173,20 @@ bool Type::canLosslesslyBitCastTo(const Type *Ty) const {
     return false;
 
   // Vector -> Vector conversions are always lossless if the two vector types
-  // have the same size, otherwise not.
-  if (const VectorType *thisPTy = dyn_cast<VectorType>(this))
+  // have the same size, otherwise not.  Also, 64-bit vector types can be
+  // converted to x86mmx.
+  if (const VectorType *thisPTy = dyn_cast<VectorType>(this)) {
     if (const VectorType *thatPTy = dyn_cast<VectorType>(Ty))
       return thisPTy->getBitWidth() == thatPTy->getBitWidth();
+    if (Ty->getTypeID() == Type::X86_MMXTyID &&
+        thisPTy->getBitWidth() == 64)
+      return true;
+  }
+
+  if (this->getTypeID() == Type::X86_MMXTyID)
+    if (const VectorType *thatPTy = dyn_cast<VectorType>(Ty))
+      if (thatPTy->getBitWidth() == 64)
+        return true;
 
   // At this point we have only various mismatches of the first class types
   // remaining and ptr->ptr. Just select the lossless conversions. Everything