[X86] Lowering SAD (sum of absolute differences) intrinsics to native IR (LLVM side)

[mike.dvoretsky]

Summary

This patch lowers the SAD intrinsics to native LLVM IR. Comes with a clang patch (D45722).

[craig.topper]

lib/IR/AutoUpgrade.cpp

2389 ↗

(On Diff #142771)

Can't N be calculated from CI.getType()?

[mike.dvoretsky]

Updated per comments.

[RKSimon]

Don't remove the -fast-isel tests - they should be updated to test the generic codegen instead

[craig.topper]

How much value are we getting out of this change? Does this expose a lot of optimization potential to the middle end? This is a pretty complex sequence. How easy/likely is for the middle end to mess this up and make it hard for the backend to recognize?

[mike.dvoretsky]

Fast-isel tests restored and updated.

[mike.dvoretsky]

Updated per comments in the clang part.

[mike.dvoretsky]

Style fix.

[RKSimon]

In D45723#1071518, @craig.topper wrote:

How much value are we getting out of this change? Does this expose a lot of optimization potential to the middle end? This is a pretty complex sequence. How easy/likely is for the middle end to mess this up and make it hard for the backend to recognize?

This will be the 3rd code path we'll have for PSADBW recognition.....

llvm/lib/Target/X86/X86ISelLowering.cpp

38075

Can we reuse/tweak matchBinOpReduction to do this for us?

[mike.dvoretsky]

llvm/lib/Target/X86/X86ISelLowering.cpp

38075

This is not a scalar reduction. The patterns calls for a sum of specifically formed vectors (hence all the checks below) to form the PSADBW instruction where it is exactly semantically fitting rather than where it can be used as a reduction tool. This is also why the third path to recognize it is being added - other paths use it for reductions and so don't actually need the input pattern to match it in terms of which qword the specific byte corresponds to.

llvm/include/llvm/IR/IntrinsicsX86.td

def int_x86_sse2_pmulhu_w : GCCBuiltin<"__builtin_ia32_pmulhuw128">,
               Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_pmulh_w : GCCBuiltin<"__builtin_ia32_pmulhw128">,
               Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem, Commutative]>;
   def int_x86_sse2_pmadd_wd : GCCBuiltin<"__builtin_ia32_pmaddwd128">,
               Intrinsic<[llvm_v4i32_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem, Commutative]>;
-  def int_x86_sse2_psad_bw : GCCBuiltin<"__builtin_ia32_psadbw128">,
-              Intrinsic<[llvm_v2i64_ty], [llvm_v16i8_ty,
-                         llvm_v16i8_ty], [IntrNoMem, Commutative]>;
 }
 
 // Integer shift ops.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse2_psll_w : GCCBuiltin<"__builtin_ia32_psllw128">,
               Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
   def int_x86_sse2_psll_d : GCCBuiltin<"__builtin_ia32_pslld128">,
def int_x86_avx2_pmulhu_w : GCCBuiltin<"__builtin_ia32_pmulhuw256">,
               Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
                          llvm_v16i16_ty], [IntrNoMem, Commutative]>;
   def int_x86_avx2_pmulh_w : GCCBuiltin<"__builtin_ia32_pmulhw256">,
               Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
                          llvm_v16i16_ty], [IntrNoMem, Commutative]>;
   def int_x86_avx2_pmadd_wd : GCCBuiltin<"__builtin_ia32_pmaddwd256">,
               Intrinsic<[llvm_v8i32_ty], [llvm_v16i16_ty,
                          llvm_v16i16_ty], [IntrNoMem, Commutative]>;
-  def int_x86_avx2_psad_bw : GCCBuiltin<"__builtin_ia32_psadbw256">,
-              Intrinsic<[llvm_v4i64_ty], [llvm_v32i8_ty,
-                         llvm_v32i8_ty], [IntrNoMem, Commutative]>;
 }
 
 // Integer shift ops.
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_avx2_psll_w : GCCBuiltin<"__builtin_ia32_psllw256">,
               Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
   def int_x86_avx2_psll_d : GCCBuiltin<"__builtin_ia32_pslld256">,
let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_avx512_rsqrt28_ss : GCCBuiltin<"__builtin_ia32_rsqrt28ss_round_mask">,
             Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
                                         llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty],
                       [IntrNoMem]>;
   def int_x86_avx512_rsqrt28_sd : GCCBuiltin<"__builtin_ia32_rsqrt28sd_round_mask">,
             Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                                         llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty],
                       [IntrNoMem]>;
-  def int_x86_avx512_psad_bw_512 : GCCBuiltin<"__builtin_ia32_psadbw512">,
-              Intrinsic<[llvm_v8i64_ty], [llvm_v64i8_ty, llvm_v64i8_ty],
-                        [IntrNoMem, Commutative]>;
 }
 // Integer arithmetic ops
 let TargetPrefix = "x86" in {
   def int_x86_avx512_mask_padds_b_128 : // FIXME: remove this intrinsic
           Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty,
                      llvm_v16i8_ty, llvm_i16_ty], [IntrNoMem]>;
   def int_x86_avx512_mask_padds_b_256 : // FIXME: remove this intrinsic
           Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty, llvm_v32i8_ty,

llvm/lib/IR/AutoUpgrade.cpp

if (Name=="ssse3.pabs.b.128" || // Added in 6.0
       Name.startswith("avx512.mask.move.s") || // Added in 4.0
       Name.startswith("avx512.cvtmask2") || // Added in 5.0
       (Name.startswith("xop.vpcom") && // Added in 3.2
        F->arg_size() == 2) ||
       Name.startswith("avx512.ptestm") || //Added in 6.0
       Name.startswith("avx512.ptestnm") || //Added in 6.0
       Name.startswith("sse2.pavg") || // Added in 6.0
       Name.startswith("avx2.pavg") || // Added in 6.0
-      Name.startswith("avx512.mask.pavg")) // Added in 6.0
+      Name.startswith("avx512.mask.pavg") || // Added in 6.0
+      Name == "sse2.psad.bw" || // Added in 7.0
+      Name == "avx2.psad.bw" || // Added in 7.0
+      Name == "avx512.psad.bw.512") // Added in 7.0
     return true;
 
   return false;
 }
 
 static bool UpgradeX86IntrinsicFunction(Function *F, StringRef Name,
                                         Function *&NewFn) {
   // Only handle intrinsics that start with "x86.".
void llvm::UpgradeInlineAsmString(std::string *AsmStr) {
   if (AsmStr->find("mov\tfp") == 0 &&
       AsmStr->find("objc_retainAutoreleaseReturnValue") != std::string::npos &&
       (Pos = AsmStr->find("# marker")) != std::string::npos) {
     AsmStr->replace(Pos, 1, ";");
   }
   return;
 }
 
+// Upgrades calls to PSAD intrinsics.
+static Value *UpgradeX86SAD(IRBuilder<> &Builder, CallInst &CI) {
+  // The operands arrive already bitcast to byte vectors.
+  Value *A = CI.getArgOperand(0);
+  Value *B = CI.getArgOperand(1);
+  // N shows the corresponding number of qwords.
+  unsigned N = (cast<VectorType>(CI.getType()))->getBitWidth() / 64;
+
+  Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_UGT, A, B);
+  Value *AD = Builder.CreateSelect(Cmp, Builder.CreateSub(A, B),
+                                   Builder.CreateSub(B, A));
+
+  Type *QTy = Builder.getInt64Ty();
+  Type *VTy = VectorType::get(QTy, N);
+  SmallVector<uint32_t, 8> ShuffleMask(N);
+  for (unsigned i = 0; i < N; ++i)
+    ShuffleMask[i] = i * 8;
+  Value *Res = Builder.CreateZExt(
+      Builder.CreateShuffleVector(AD, AD, ShuffleMask), VTy);
+  for (unsigned i = 1; i < 8; ++i) {
+    for (unsigned j = 0; j < N; ++j)
+      ShuffleMask[j] = i + j * 8;
+    Value *Sum = Builder.CreateShuffleVector(AD, AD, ShuffleMask);
+    Res = Builder.CreateAdd(Res, Builder.CreateZExt(Sum, VTy));
+  }
+  return Res;
+
+}
+
 /// Upgrade a call to an old intrinsic. All argument and return casting must be
 /// provided to seamlessly integrate with existing context.
 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
   Function *F = CI->getCalledFunction();
   LLVMContext &C = CI->getContext();
   IRBuilder<> Builder(C);
   Builder.SetInsertPoint(CI->getParent(), CI->getIterator());
 

       Rep = Builder.CreateTrunc(ShiftR, A->getType());
       if (CI->getNumArgOperands() > 2) {
         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
                             CI->getArgOperand(2));
       }
     } else if (IsX86 && Name.startswith("avx512.mask.") &&
                upgradeAVX512MaskToSelect(Name, Builder, *CI, Rep)) {
       // Rep will be updated by the call in the condition.
+    } else if (IsX86 &&
+               (Name.startswith("sse2.psad") || Name.startswith("avx2.psad") ||
+                Name.startswith("avx512.psad"))) {
+      // llvm.x86.sse2.psad.bw, llvm.x86.avx2.psad.bw,
+      // llvm.x86.avx512.psad.bw.512
+      Rep = UpgradeX86SAD(Builder, *CI);
     } else if (IsNVVM && (Name == "abs.i" || Name == "abs.ll")) {
       Value *Arg = CI->getArgOperand(0);
       Value *Neg = Builder.CreateNeg(Arg, "neg");
       Value *Cmp = Builder.CreateICmpSGE(
           Arg, llvm::Constant::getNullValue(Arg->getType()), "abs.cond");
       Rep = Builder.CreateSelect(Cmp, Arg, Neg, "abs");
     } else if (IsNVVM && (Name == "max.i" || Name == "max.ll" ||
                           Name == "max.ui" || Name == "max.ull")) {

llvm/lib/Target/X86/X86ISelLowering.cpp

return DAG.getNode(X86ISD::VPMADDWD, DL, ResVT,
                        DAG.getNode(ISD::TRUNCATE, DL, TruncVT, Ops[0]),
                        DAG.getNode(ISD::TRUNCATE, DL, TruncVT, Ops[1]));
   };
   return SplitOpsAndApply(DAG, Subtarget, DL, VT,
                           { Mul.getOperand(0), Mul.getOperand(1) },
                           PMADDBuilder);
 }
 
+// Try to find a sum-of-shuffles pattern for PSADBW.
+static SDValue combineScalarSADPattern(SDNode *Node, SelectionDAG &DAG,
+                                       const X86Subtarget &Subtarget) {
+  if (Node->getOpcode() != ISD::ADD)
+    return SDValue();
+  EVT Ty = Node->getValueType(0);
+  if (!Ty.isVector() || Ty.getScalarSizeInBits() != 64)
+    return SDValue();
+  unsigned N = Ty.getVectorNumElements();
+  if ((N != 2 && N != 4 && N != 8) || (N == 2 && !Subtarget.hasSSE2()) ||
+      (N == 4 && !Subtarget.hasAVX2()) ||
+      (N == 8 && (!Subtarget.hasAVX512() || !Subtarget.hasBWI())))
+    return SDValue();
+  SDValue ByteGroup[8], Cap, Top;
+  bool TopSpotted = false;
+  ByteGroup[0] = SDValue(Node, 0);
+  bool ByteGroupDone[8] = {false, false, false, false,
+                           false, false, false, false};
+  unsigned NumByteGroups = 1;
+  // For the vector SAD, we expect the following pattern: nodes are extracted
+  // by vectors of 0-7 offsets, creating 8 vectors of byte differences that are
+  // then summed.
+  while (NumByteGroups) {
+    // If it's a sum, go up the tree.
+    if (ByteGroup[NumByteGroups - 1].getOpcode() == ISD::ADD) {
+      if (NumByteGroups == 8)
+        return SDValue();
+      ByteGroup[NumByteGroups] = ByteGroup[NumByteGroups - 1].getOperand(1);
+      ByteGroup[NumByteGroups - 1] = ByteGroup[NumByteGroups - 1].getOperand(0);
+      ++NumByteGroups;
+    // Check that the pattern above zext corresponds to a shuffle of the AD
+    // pattern's output with the <pos, pos + 8, pos + 16, ...> shuffle mask.
+    } else if (ByteGroup[NumByteGroups - 1].getOpcode() == ISD::ZERO_EXTEND &&
+               ByteGroup[NumByteGroups - 1]
+                       .getValueType()
+                       .getVectorNumElements() == N) {
+      --NumByteGroups;
+      Cap = ByteGroup[NumByteGroups].getOperand(0);
+      if (Cap.getOpcode() != ISD::BUILD_VECTOR)
+        return SDValue();
+      if (Cap.getNumOperands() < N)
+        return SDValue();
+      SDValue Extract = Cap.getOperand(0);
+      if (Extract.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
+        return SDValue();
+      if (!TopSpotted) {
+        Top = Extract.getOperand(0);
+        TopSpotted = true;
+      } else if (Extract.getOperand(0) != Top)
+        return SDValue();
+      auto C = dyn_cast<ConstantSDNode>(Extract.getOperand(1).getNode());
+      if (!C)
+        return SDValue();
+      unsigned Pos = C->getZExtValue();
+      if (Pos >= 8 || ByteGroupDone[Pos])
+        return SDValue();
+      for (unsigned i = 1; i < N; ++i) {
+        Extract = Cap.getOperand(i);
+        if (Extract.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
+            Extract.getOperand(0) != Top)
+          return SDValue();
+        auto C = dyn_cast<ConstantSDNode>(Extract.getOperand(1).getNode());
+        if (!C || C->getZExtValue() != Pos + 8 * i)
+          return SDValue();
+      }
+      ByteGroupDone[Pos] = true;
+    } else
+      return SDValue();
+  }
+  // Check that we've found 8 shuffles.
+  for (unsigned i = 0; i < 8; ++i)
+    if (!ByteGroupDone[i])
+      return SDValue();
+  // Check that Top points to an absolute difference pattern, that is
+  // AD[i] = (A[i] > B[i]) ? A[i] - B[i] : B[i] - A[i]
+  if (Top.getOpcode() != ISD::VSELECT)
+    return SDValue();
+  SDValue SubP = Top.getOperand(1);
+  SDValue SubN = Top.getOperand(2);
+  if (SubP.getOpcode() != ISD::SUB || SubN.getOpcode() != ISD::SUB)
+    return SDValue();
+  Top = Top.getOperand(0);
+  if (Top.getOpcode() != ISD::SETCC)
+    return SDValue();
+  ISD::CondCode CC = cast<CondCodeSDNode>(Top.getOperand(2))->get();
+  if (CC != ISD::SETUGT && CC != ISD::SETUGE)
+    return SDValue();
+  SDValue A = Top.getOperand(0);
+  SDValue B = Top.getOperand(1);
+  if (SubP.getOperand(0) != A || SubP.getOperand(1) != B ||
+      SubN.getOperand(0) != B || SubN.getOperand(1) != A)
+    return SDValue();
+
+  SDLoc DL(Node);
+  MVT VT = MVT::getVectorVT(MVT::i64, N);
+  return DAG.getNode(X86ISD::PSADBW, DL, VT, A, B);
+}
+
 static SDValue combineAdd(SDNode *N, SelectionDAG &DAG,
                           const X86Subtarget &Subtarget) {
   const SDNodeFlags Flags = N->getFlags();
+  if (SDValue VSad = combineScalarSADPattern(N, DAG, Subtarget))
+    return VSad;
   if (Flags.hasVectorReduction()) {
     if (SDValue Sad = combineLoopSADPattern(N, DAG, Subtarget))
       return Sad;
     if (SDValue MAdd = combineLoopMAddPattern(N, DAG, Subtarget))
       return MAdd;
   }
   EVT VT = N->getValueType(0);
   SDValue Op0 = N->getOperand(0);

llvm/lib/Target/X86/X86IntrinsicsInfo.h

static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx2_phsub_d, INTR_TYPE_2OP, X86ISD::HSUB, 0),
   X86_INTRINSIC_DATA(avx2_phsub_w, INTR_TYPE_2OP, X86ISD::HSUB, 0),
   X86_INTRINSIC_DATA(avx2_pmadd_ub_sw, INTR_TYPE_2OP, X86ISD::VPMADDUBSW, 0),
   X86_INTRINSIC_DATA(avx2_pmadd_wd, INTR_TYPE_2OP, X86ISD::VPMADDWD, 0),
   X86_INTRINSIC_DATA(avx2_pmovmskb, INTR_TYPE_1OP, X86ISD::MOVMSK, 0),
   X86_INTRINSIC_DATA(avx2_pmul_hr_sw, INTR_TYPE_2OP, X86ISD::MULHRS, 0),
   X86_INTRINSIC_DATA(avx2_pmulh_w, INTR_TYPE_2OP, ISD::MULHS, 0),
   X86_INTRINSIC_DATA(avx2_pmulhu_w, INTR_TYPE_2OP, ISD::MULHU, 0),
-  X86_INTRINSIC_DATA(avx2_psad_bw, INTR_TYPE_2OP, X86ISD::PSADBW, 0),
   X86_INTRINSIC_DATA(avx2_pshuf_b, INTR_TYPE_2OP, X86ISD::PSHUFB, 0),
   X86_INTRINSIC_DATA(avx2_psll_d, INTR_TYPE_2OP, X86ISD::VSHL, 0),
   X86_INTRINSIC_DATA(avx2_psll_q, INTR_TYPE_2OP, X86ISD::VSHL, 0),
   X86_INTRINSIC_DATA(avx2_psll_w, INTR_TYPE_2OP, X86ISD::VSHL, 0),
   X86_INTRINSIC_DATA(avx2_pslli_d, VSHIFT, X86ISD::VSHLI, 0),
   X86_INTRINSIC_DATA(avx2_pslli_q, VSHIFT, X86ISD::VSHLI, 0),
   X86_INTRINSIC_DATA(avx2_pslli_w, VSHIFT, X86ISD::VSHLI, 0),
   X86_INTRINSIC_DATA(avx2_psllv_d, INTR_TYPE_2OP, ISD::SHL, 0),
X86_INTRINSIC_DATA(avx512_mask_vpermi2var_d_128, VPERM_3OP_MASK,
   X86_INTRINSIC_DATA(avx512_packuswb_512, INTR_TYPE_2OP, X86ISD::PACKUS, 0),
   X86_INTRINSIC_DATA(avx512_pmaddubs_w_512, INTR_TYPE_2OP,
                      X86ISD::VPMADDUBSW, 0),
   X86_INTRINSIC_DATA(avx512_pmaddw_d_512, INTR_TYPE_2OP,
                      X86ISD::VPMADDWD, 0),
   X86_INTRINSIC_DATA(avx512_pmul_hr_sw_512, INTR_TYPE_2OP, X86ISD::MULHRS, 0),
   X86_INTRINSIC_DATA(avx512_pmulh_w_512, INTR_TYPE_2OP, ISD::MULHS, 0),
   X86_INTRINSIC_DATA(avx512_pmulhu_w_512, INTR_TYPE_2OP, ISD::MULHU, 0),
-  X86_INTRINSIC_DATA(avx512_psad_bw_512, INTR_TYPE_2OP, X86ISD::PSADBW, 0),
   X86_INTRINSIC_DATA(avx512_pshuf_b_512, INTR_TYPE_2OP, X86ISD::PSHUFB, 0),
   X86_INTRINSIC_DATA(avx512_psll_d_512, INTR_TYPE_2OP, X86ISD::VSHL, 0),
   X86_INTRINSIC_DATA(avx512_psll_q_512, INTR_TYPE_2OP, X86ISD::VSHL, 0),
   X86_INTRINSIC_DATA(avx512_psll_w_512, INTR_TYPE_2OP, X86ISD::VSHL, 0),
   X86_INTRINSIC_DATA(avx512_pslli_d_512, VSHIFT, X86ISD::VSHLI, 0),
   X86_INTRINSIC_DATA(avx512_pslli_q_512, VSHIFT, X86ISD::VSHLI, 0),
   X86_INTRINSIC_DATA(avx512_pslli_w_512, VSHIFT, X86ISD::VSHLI, 0),
   X86_INTRINSIC_DATA(avx512_psllv_d_512, INTR_TYPE_2OP, ISD::SHL, 0),
X86_INTRINSIC_DATA(avx512_mask_vpermi2var_d_128, VPERM_3OP_MASK,
   X86_INTRINSIC_DATA(sse2_padds_b,      INTR_TYPE_2OP, X86ISD::ADDS, 0),
   X86_INTRINSIC_DATA(sse2_padds_w,      INTR_TYPE_2OP, X86ISD::ADDS, 0),
   X86_INTRINSIC_DATA(sse2_paddus_b,     INTR_TYPE_2OP, X86ISD::ADDUS, 0),
   X86_INTRINSIC_DATA(sse2_paddus_w,     INTR_TYPE_2OP, X86ISD::ADDUS, 0),
   X86_INTRINSIC_DATA(sse2_pmadd_wd,     INTR_TYPE_2OP, X86ISD::VPMADDWD, 0),
   X86_INTRINSIC_DATA(sse2_pmovmskb_128, INTR_TYPE_1OP, X86ISD::MOVMSK, 0),
   X86_INTRINSIC_DATA(sse2_pmulh_w,      INTR_TYPE_2OP, ISD::MULHS, 0),
   X86_INTRINSIC_DATA(sse2_pmulhu_w,     INTR_TYPE_2OP, ISD::MULHU, 0),
-  X86_INTRINSIC_DATA(sse2_psad_bw,      INTR_TYPE_2OP, X86ISD::PSADBW, 0),
   X86_INTRINSIC_DATA(sse2_psll_d,       INTR_TYPE_2OP, X86ISD::VSHL, 0),
   X86_INTRINSIC_DATA(sse2_psll_q,       INTR_TYPE_2OP, X86ISD::VSHL, 0),
   X86_INTRINSIC_DATA(sse2_psll_w,       INTR_TYPE_2OP, X86ISD::VSHL, 0),
   X86_INTRINSIC_DATA(sse2_pslli_d,      VSHIFT, X86ISD::VSHLI, 0),
   X86_INTRINSIC_DATA(sse2_pslli_q,      VSHIFT, X86ISD::VSHLI, 0),
   X86_INTRINSIC_DATA(sse2_pslli_w,      VSHIFT, X86ISD::VSHLI, 0),
   X86_INTRINSIC_DATA(sse2_psra_d,       INTR_TYPE_2OP, X86ISD::VSRA, 0),
   X86_INTRINSIC_DATA(sse2_psra_w,       INTR_TYPE_2OP, X86ISD::VSRA, 0),

llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp

case Intrinsic::x86_mmx_packuswb:
       handleVectorPackIntrinsic(I, 16);
       break;
 
     case Intrinsic::x86_mmx_packssdw:
       handleVectorPackIntrinsic(I, 32);
       break;
 
     case Intrinsic::x86_mmx_psad_bw:
-    case Intrinsic::x86_sse2_psad_bw:
-    case Intrinsic::x86_avx2_psad_bw:
       handleVectorSadIntrinsic(I);
       break;
 
     case Intrinsic::x86_sse2_pmadd_wd:
     case Intrinsic::x86_avx2_pmadd_wd:
     case Intrinsic::x86_ssse3_pmadd_ub_sw_128:
     case Intrinsic::x86_avx2_pmadd_ub_sw:
       handleVectorPmaddIntrinsic(I);

llvm/test/CodeGen/X86/avx2-intrinsics-fast-isel.ll

 ; CHECK-NEXT:    ret{{[l|q]}}
   %arg1 = bitcast <4 x i64> %a1 to <8 x i32>
   %res = call <8 x float> @llvm.x86.avx2.permps(<8 x float> %a0, <8 x i32> %arg1)
   ret <8 x float> %res
 }
 declare <8 x float> @llvm.x86.avx2.permps(<8 x float>, <8 x i32>) nounwind readonly
 
 define <4 x i64> @test_mm256_sad_epu8(<4 x i64> %a0, <4 x i64> %a1) {
-; CHECK-LABEL: test_mm256_sad_epu8:
-; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; CHECK-NEXT:    ret{{[l|q]}}
-  %arg0 = bitcast <4 x i64> %a0 to <32 x i8>
-  %arg1 = bitcast <4 x i64> %a1 to <32 x i8>
-  %res = call <4 x i64> @llvm.x86.avx2.psad.bw(<32 x i8> %arg0, <32 x i8> %arg1)
+; X86-LABEL: test_mm256_sad_epu8:
+; X86:       # %bb.0:
+; X86-NEXT:    vpminub %ymm1, %ymm0, %ymm2
+; X86-NEXT:    vpcmpeqb %ymm2, %ymm0, %ymm2
+; X86-NEXT:    vpcmpeqd %ymm3, %ymm3, %ymm3
+; X86-NEXT:    vpxor %ymm3, %ymm2, %ymm2
+; X86-NEXT:    vpsubb %ymm1, %ymm0, %ymm3
+; X86-NEXT:    vpsubb %ymm0, %ymm1, %ymm0
+; X86-NEXT:    vpblendvb %ymm2, %ymm3, %ymm0, %ymm0
+; X86-NEXT:    vpand {{\.LCPI.*}}, %ymm0, %ymm1
+; X86-NEXT:    vpshufb {{.*#+}} ymm2 = ymm0[1],zero,zero,zero,zero,zero,zero,zero,ymm0[9],zero,zero,zero,zero,zero,zero,zero,ymm0[17],zero,zero,zero,zero,zero,zero,zero,ymm0[25],zero,zero,zero,zero,zero,zero,zero
+; X86-NEXT:    vpshufb {{.*#+}} ymm3 = ymm0[2],zero,zero,zero,zero,zero,zero,zero,ymm0[10],zero,zero,zero,zero,zero,zero,zero,ymm0[18],zero,zero,zero,zero,zero,zero,zero,ymm0[26],zero,zero,zero,zero,zero,zero,zero
+; X86-NEXT:    vpaddq %ymm3, %ymm2, %ymm2
+; X86-NEXT:    vpshufb {{.*#+}} ymm3 = ymm0[3],zero,zero,zero,zero,zero,zero,zero,ymm0[11],zero,zero,zero,zero,zero,zero,zero,ymm0[19],zero,zero,zero,zero,zero,zero,zero,ymm0[27],zero,zero,zero,zero,zero,zero,zero
+; X86-NEXT:    vpaddq %ymm3, %ymm2, %ymm2
+; X86-NEXT:    vpshufb {{.*#+}} ymm3 = ymm0[4],zero,zero,zero,zero,zero,zero,zero,ymm0[12],zero,zero,zero,zero,zero,zero,zero,ymm0[20],zero,zero,zero,zero,zero,zero,zero,ymm0[28],zero,zero,zero,zero,zero,zero,zero
+; X86-NEXT:    vpaddq %ymm3, %ymm2, %ymm2
+; X86-NEXT:    vpshufb {{.*#+}} ymm3 = ymm0[5],zero,zero,zero,zero,zero,zero,zero,ymm0[13],zero,zero,zero,zero,zero,zero,zero,ymm0[21],zero,zero,zero,zero,zero,zero,zero,ymm0[29],zero,zero,zero,zero,zero,zero,zero
+; X86-NEXT:    vpaddq %ymm3, %ymm2, %ymm2
+; X86-NEXT:    vpshufb {{.*#+}} ymm3 = ymm0[6],zero,zero,zero,zero,zero,zero,zero,ymm0[14],zero,zero,zero,zero,zero,zero,zero,ymm0[22],zero,zero,zero,zero,zero,zero,zero,ymm0[30],zero,zero,zero,zero,zero,zero,zero
+; X86-NEXT:    vpaddq %ymm3, %ymm2, %ymm2
+; X86-NEXT:    vpsrlq $56, %ymm0, %ymm0
+; X86-NEXT:    vpaddq %ymm0, %ymm2, %ymm0
+; X86-NEXT:    vpaddq %ymm0, %ymm1, %ymm0
+; X86-NEXT:    ret{{[l|q]}}
+;
+; X64-LABEL: test_mm256_sad_epu8:
+; X64:       # %bb.0:
+; X64-NEXT:    vpminub %ymm1, %ymm0, %ymm2
+; X64-NEXT:    vpcmpeqb %ymm2, %ymm0, %ymm2
+; X64-NEXT:    vpcmpeqd %ymm3, %ymm3, %ymm3
+; X64-NEXT:    vpxor %ymm3, %ymm2, %ymm2
+; X64-NEXT:    vpsubb %ymm1, %ymm0, %ymm3
+; X64-NEXT:    vpsubb %ymm0, %ymm1, %ymm0
+; X64-NEXT:    vpblendvb %ymm2, %ymm3, %ymm0, %ymm0
+; X64-NEXT:    vpand {{.*}}(%rip), %ymm0, %ymm1
+; X64-NEXT:    vpshufb {{.*#+}} ymm2 = ymm0[1],zero,zero,zero,zero,zero,zero,zero,ymm0[9],zero,zero,zero,zero,zero,zero,zero,ymm0[17],zero,zero,zero,zero,zero,zero,zero,ymm0[25],zero,zero,zero,zero,zero,zero,zero
+; X64-NEXT:    vpshufb {{.*#+}} ymm3 = ymm0[2],zero,zero,zero,zero,zero,zero,zero,ymm0[10],zero,zero,zero,zero,zero,zero,zero,ymm0[18],zero,zero,zero,zero,zero,zero,zero,ymm0[26],zero,zero,zero,zero,zero,zero,zero
+; X64-NEXT:    vpaddq %ymm3, %ymm2, %ymm2
+; X64-NEXT:    vpshufb {{.*#+}} ymm3 = ymm0[3],zero,zero,zero,zero,zero,zero,zero,ymm0[11],zero,zero,zero,zero,zero,zero,zero,ymm0[19],zero,zero,zero,zero,zero,zero,zero,ymm0[27],zero,zero,zero,zero,zero,zero,zero
+; X64-NEXT:    vpaddq %ymm3, %ymm2, %ymm2
+; X64-NEXT:    vpshufb {{.*#+}} ymm3 = ymm0[4],zero,zero,zero,zero,zero,zero,zero,ymm0[12],zero,zero,zero,zero,zero,zero,zero,ymm0[20],zero,zero,zero,zero,zero,zero,zero,ymm0[28],zero,zero,zero,zero,zero,zero,zero
+; X64-NEXT:    vpaddq %ymm3, %ymm2, %ymm2
+; X64-NEXT:    vpshufb {{.*#+}} ymm3 = ymm0[5],zero,zero,zero,zero,zero,zero,zero,ymm0[13],zero,zero,zero,zero,zero,zero,zero,ymm0[21],zero,zero,zero,zero,zero,zero,zero,ymm0[29],zero,zero,zero,zero,zero,zero,zero
+; X64-NEXT:    vpaddq %ymm3, %ymm2, %ymm2
+; X64-NEXT:    vpshufb {{.*#+}} ymm3 = ymm0[6],zero,zero,zero,zero,zero,zero,zero,ymm0[14],zero,zero,zero,zero,zero,zero,zero,ymm0[22],zero,zero,zero,zero,zero,zero,zero,ymm0[30],zero,zero,zero,zero,zero,zero,zero
+; X64-NEXT:    vpaddq %ymm3, %ymm2, %ymm2
+; X64-NEXT:    vpsrlq $56, %ymm0, %ymm0
+; X64-NEXT:    vpaddq %ymm0, %ymm2, %ymm0
+; X64-NEXT:    vpaddq %ymm0, %ymm1, %ymm0
+; X64-NEXT:    ret{{[l|q]}}
+  %arg0 = bitcast <4 x i64> %a0 to <32 x i8>
+  %arg1 = bitcast <4 x i64> %a1 to <32 x i8>
+  %1 = icmp ugt <32 x i8> %arg0, %arg1
+  %2 = sub <32 x i8> %arg0, %arg1
+  %3 = sub <32 x i8> %arg1, %arg0
+  %4 = select <32 x i1> %1, <32 x i8> %2, <32 x i8> %3
+  %5 = shufflevector <32 x i8> %4, <32 x i8> %4, <4 x i32> <i32 0, i32 8, i32 16, i32 24>
+  %6 = zext <4 x i8> %5 to <4 x i64>
+  %7 = shufflevector <32 x i8> %4, <32 x i8> %4, <4 x i32> <i32 1, i32 9, i32 17, i32 25>
+  %8 = zext <4 x i8> %7 to <4 x i64>
+  %9 = add <4 x i64> %6, %8
+  %10 = shufflevector <32 x i8> %4, <32 x i8> %4, <4 x i32> <i32 2, i32 10, i32 18, i32 26>
+  %11 = zext <4 x i8> %10 to <4 x i64>
+  %12 = add <4 x i64> %9, %11
+  %13 = shufflevector <32 x i8> %4, <32 x i8> %4, <4 x i32> <i32 3, i32 11, i32 19, i32 27>
+  %14 = zext <4 x i8> %13 to <4 x i64>
+  %15 = add <4 x i64> %12, %14
+  %16 = shufflevector <32 x i8> %4, <32 x i8> %4, <4 x i32> <i32 4, i32 12, i32 20, i32 28>
+  %17 = zext <4 x i8> %16 to <4 x i64>
+  %18 = add <4 x i64> %15, %17
+  %19 = shufflevector <32 x i8> %4, <32 x i8> %4, <4 x i32> <i32 5, i32 13, i32 21, i32 29>
+  %20 = zext <4 x i8> %19 to <4 x i64>
+  %21 = add <4 x i64> %18, %20
+  %22 = shufflevector <32 x i8> %4, <32 x i8> %4, <4 x i32> <i32 6, i32 14, i32 22, i32 30>
+  %23 = zext <4 x i8> %22 to <4 x i64>
+  %24 = add <4 x i64> %21, %23
+  %25 = shufflevector <32 x i8> %4, <32 x i8> %4, <4 x i32> <i32 7, i32 15, i32 23, i32 31>
+  %26 = zext <4 x i8> %25 to <4 x i64>
+  %res = add <4 x i64> %24, %26
   ret <4 x i64> %res
 }
-declare <4 x i64> @llvm.x86.avx2.psad.bw(<32 x i8>, <32 x i8>) nounwind readnone
 
 define <4 x i64> @test_mm256_shuffle_epi32(<4 x i64> %a0) {
 ; CHECK-LABEL: test_mm256_shuffle_epi32:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vpermilps {{.*#+}} ymm0 = ymm0[3,3,0,0,7,7,4,4]
 ; CHECK-NEXT:    ret{{[l|q]}}
   %arg0 = bitcast <4 x i64> %a0 to <8 x i32>
   %shuf = shufflevector <8 x i32> %arg0, <8 x i32> undef, <8 x i32> <i32 3, i32 3, i32 0, i32 0, i32 7, i32 7, i32 4, i32 4>

llvm/test/CodeGen/X86/avx2-intrinsics-x86-upgrade.ll

 ; X64-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4,5,6,7]
 ; X64-NEXT:    retq
   %res = call <8 x i32> @llvm.x86.avx2.pblendd.256(<8 x i32> %a0, <8 x i32> %a1, i32 7) ; <<8 x i32>> [#uses=1]
   ret <8 x i32> %res
 }
 declare <8 x i32> @llvm.x86.avx2.pblendd.256(<8 x i32>, <8 x i32>, i32) nounwind readnone
 
 
+define <4 x i64> @test_x86_avx2_psad_bw(<32 x i8> %a0, <32 x i8> %a1) {
+; X86-LABEL: test_x86_avx2_psad_bw:
+; X86:       ## %bb.0:
+; X86-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
+; X86-NEXT:    retl
+;
+; X64-LABEL: test_x86_avx2_psad_bw:
+; X64:       ## %bb.0:
+; X64-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
+; X64-NEXT:    retq
+  %res = call <4 x i64> @llvm.x86.avx2.psad.bw(<32 x i8> %a0, <32 x i8> %a1) ; <<4 x i64>> [#uses=1]
+  ret <4 x i64> %res
+}
+declare <4 x i64> @llvm.x86.avx2.psad.bw(<32 x i8>, <32 x i8>) nounwind readnone
+
+
 define <4 x i64> @test_x86_avx2_movntdqa(i8* %a0) {
 ; X86-LABEL: test_x86_avx2_movntdqa:
 ; X86:       ## %bb.0:
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    vmovntdqa (%eax), %ymm0
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_x86_avx2_movntdqa:

llvm/test/CodeGen/X86/avx2-intrinsics-x86.ll

 ; X64-AVX512VL-NEXT:    vpmulhuw %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xe4,0xc1]
 ; X64-AVX512VL-NEXT:    retq ## encoding: [0xc3]
   %res = call <16 x i16> @llvm.x86.avx2.pmulhu.w(<16 x i16> %a0, <16 x i16> %a1) ; <<16 x i16>> [#uses=1]
   ret <16 x i16> %res
 }
 declare <16 x i16> @llvm.x86.avx2.pmulhu.w(<16 x i16>, <16 x i16>) nounwind readnone
 
 
-define <4 x i64> @test_x86_avx2_psad_bw(<32 x i8> %a0, <32 x i8> %a1) {
-; X86-AVX-LABEL: test_x86_avx2_psad_bw:
-; X86-AVX:       ## %bb.0:
-; X86-AVX-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0 ## encoding: [0xc5,0xfd,0xf6,0xc1]
-; X86-AVX-NEXT:    retl ## encoding: [0xc3]
-;
-; X86-AVX512VL-LABEL: test_x86_avx2_psad_bw:
-; X86-AVX512VL:       ## %bb.0:
-; X86-AVX512VL-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xf6,0xc1]
-; X86-AVX512VL-NEXT:    retl ## encoding: [0xc3]
-;
-; X64-AVX-LABEL: test_x86_avx2_psad_bw:
-; X64-AVX:       ## %bb.0:
-; X64-AVX-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0 ## encoding: [0xc5,0xfd,0xf6,0xc1]
-; X64-AVX-NEXT:    retq ## encoding: [0xc3]
-;
-; X64-AVX512VL-LABEL: test_x86_avx2_psad_bw:
-; X64-AVX512VL:       ## %bb.0:
-; X64-AVX512VL-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xf6,0xc1]
-; X64-AVX512VL-NEXT:    retq ## encoding: [0xc3]
-  %res = call <4 x i64> @llvm.x86.avx2.psad.bw(<32 x i8> %a0, <32 x i8> %a1) ; <<4 x i64>> [#uses=1]
-  ret <4 x i64> %res
-}
-declare <4 x i64> @llvm.x86.avx2.psad.bw(<32 x i8>, <32 x i8>) nounwind readnone
-
-
 define <8 x i32> @test_x86_avx2_psll_d(<8 x i32> %a0, <4 x i32> %a1) {
 ; X86-AVX-LABEL: test_x86_avx2_psll_d:
 ; X86-AVX:       ## %bb.0:
 ; X86-AVX-NEXT:    vpslld %xmm1, %ymm0, %ymm0 ## encoding: [0xc5,0xfd,0xf2,0xc1]
 ; X86-AVX-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X86-AVX512VL-LABEL: test_x86_avx2_psll_d:
 ; X86-AVX512VL:       ## %bb.0:

 declare <16 x i16> @llvm.x86.avx2.packusdw(<8 x i32>, <8 x i32>) nounwind readnone
 
 
 define <16 x i16> @test_x86_avx2_packusdw_fold() {
 ; X86-AVX-LABEL: test_x86_avx2_packusdw_fold:
 ; X86-AVX:       ## %bb.0:
 ; X86-AVX-NEXT:    vmovaps {{.*#+}} ymm0 = [0,0,0,0,255,32767,65535,0,0,0,0,0,0,0,0,0]
 ; X86-AVX-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x05,A,A,A,A]
-; X86-AVX-NEXT:    ## fixup A - offset: 4, value: LCPI54_0, kind: FK_Data_4
+; X86-AVX-NEXT:    ## fixup A - offset: 4, value: LCPI53_0, kind: FK_Data_4
 ; X86-AVX-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X86-AVX512VL-LABEL: test_x86_avx2_packusdw_fold:
 ; X86-AVX512VL:       ## %bb.0:
-; X86-AVX512VL-NEXT:    vmovaps LCPI54_0, %ymm0 ## EVEX TO VEX Compression ymm0 = [0,0,0,0,255,32767,65535,0,0,0,0,0,0,0,0,0]
+; X86-AVX512VL-NEXT:    vmovaps LCPI53_0, %ymm0 ## EVEX TO VEX Compression ymm0 = [0,0,0,0,255,32767,65535,0,0,0,0,0,0,0,0,0]
 ; X86-AVX512VL-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x05,A,A,A,A]
-; X86-AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI54_0, kind: FK_Data_4
+; X86-AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI53_0, kind: FK_Data_4
 ; X86-AVX512VL-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X64-AVX-LABEL: test_x86_avx2_packusdw_fold:
 ; X64-AVX:       ## %bb.0:
 ; X64-AVX-NEXT:    vmovaps {{.*#+}} ymm0 = [0,0,0,0,255,32767,65535,0,0,0,0,0,0,0,0,0]
 ; X64-AVX-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x05,A,A,A,A]
-; X64-AVX-NEXT:    ## fixup A - offset: 4, value: LCPI54_0-4, kind: reloc_riprel_4byte
+; X64-AVX-NEXT:    ## fixup A - offset: 4, value: LCPI53_0-4, kind: reloc_riprel_4byte
 ; X64-AVX-NEXT:    retq ## encoding: [0xc3]
 ;
 ; X64-AVX512VL-LABEL: test_x86_avx2_packusdw_fold:
 ; X64-AVX512VL:       ## %bb.0:
 ; X64-AVX512VL-NEXT:    vmovaps {{.*}}(%rip), %ymm0 ## EVEX TO VEX Compression ymm0 = [0,0,0,0,255,32767,65535,0,0,0,0,0,0,0,0,0]
 ; X64-AVX512VL-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x05,A,A,A,A]
-; X64-AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI54_0-4, kind: reloc_riprel_4byte
+; X64-AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI53_0-4, kind: reloc_riprel_4byte
 ; X64-AVX512VL-NEXT:    retq ## encoding: [0xc3]
   %res = call <16 x i16> @llvm.x86.avx2.packusdw(<8 x i32> zeroinitializer, <8 x i32> <i32 255, i32 32767, i32 65535, i32 -1, i32 -32767, i32 -65535, i32 0, i32 -256>)
   ret <16 x i16> %res
 }
 
 
 define <32 x i8> @test_x86_avx2_pblendvb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> %a2) {
 ; X86-LABEL: test_x86_avx2_pblendvb:
; X64-AVX512VL-NEXT:    retq ## encoding: [0xc3]
   ret <4 x i32> %res
 }
 
 define <4 x i32> @test_x86_avx2_psrav_d_const(<4 x i32> %a0, <4 x i32> %a1) {
 ; X86-AVX-LABEL: test_x86_avx2_psrav_d_const:
 ; X86-AVX:       ## %bb.0:
 ; X86-AVX-NEXT:    vmovdqa {{.*#+}} xmm0 = [2,9,4294967284,23]
 ; X86-AVX-NEXT:    ## encoding: [0xc5,0xf9,0x6f,0x05,A,A,A,A]
-; X86-AVX-NEXT:    ## fixup A - offset: 4, value: LCPI86_0, kind: FK_Data_4
-; X86-AVX-NEXT:    vpsravd LCPI86_1, %xmm0, %xmm0 ## encoding: [0xc4,0xe2,0x79,0x46,0x05,A,A,A,A]
-; X86-AVX-NEXT:    ## fixup A - offset: 5, value: LCPI86_1, kind: FK_Data_4
+; X86-AVX-NEXT:    ## fixup A - offset: 4, value: LCPI85_0, kind: FK_Data_4
+; X86-AVX-NEXT:    vpsravd LCPI85_1, %xmm0, %xmm0 ## encoding: [0xc4,0xe2,0x79,0x46,0x05,A,A,A,A]
+; X86-AVX-NEXT:    ## fixup A - offset: 5, value: LCPI85_1, kind: FK_Data_4
 ; X86-AVX-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X86-AVX512VL-LABEL: test_x86_avx2_psrav_d_const:
 ; X86-AVX512VL:       ## %bb.0:
-; X86-AVX512VL-NEXT:    vmovdqa LCPI86_0, %xmm0 ## EVEX TO VEX Compression xmm0 = [2,9,4294967284,23]
+; X86-AVX512VL-NEXT:    vmovdqa LCPI85_0, %xmm0 ## EVEX TO VEX Compression xmm0 = [2,9,4294967284,23]
 ; X86-AVX512VL-NEXT:    ## encoding: [0xc5,0xf9,0x6f,0x05,A,A,A,A]
-; X86-AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI86_0, kind: FK_Data_4
-; X86-AVX512VL-NEXT:    vpsravd LCPI86_1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x46,0x05,A,A,A,A]
-; X86-AVX512VL-NEXT:    ## fixup A - offset: 5, value: LCPI86_1, kind: FK_Data_4
+; X86-AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI85_0, kind: FK_Data_4
+; X86-AVX512VL-NEXT:    vpsravd LCPI85_1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x46,0x05,A,A,A,A]
+; X86-AVX512VL-NEXT:    ## fixup A - offset: 5, value: LCPI85_1, kind: FK_Data_4
 ; X86-AVX512VL-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X64-AVX-LABEL: test_x86_avx2_psrav_d_const:
 ; X64-AVX:       ## %bb.0:
 ; X64-AVX-NEXT:    vmovdqa {{.*#+}} xmm0 = [2,9,4294967284,23]
 ; X64-AVX-NEXT:    ## encoding: [0xc5,0xf9,0x6f,0x05,A,A,A,A]
-; X64-AVX-NEXT:    ## fixup A - offset: 4, value: LCPI86_0-4, kind: reloc_riprel_4byte
+; X64-AVX-NEXT:    ## fixup A - offset: 4, value: LCPI85_0-4, kind: reloc_riprel_4byte
 ; X64-AVX-NEXT:    vpsravd {{.*}}(%rip), %xmm0, %xmm0 ## encoding: [0xc4,0xe2,0x79,0x46,0x05,A,A,A,A]
-; X64-AVX-NEXT:    ## fixup A - offset: 5, value: LCPI86_1-4, kind: reloc_riprel_4byte
+; X64-AVX-NEXT:    ## fixup A - offset: 5, value: LCPI85_1-4, kind: reloc_riprel_4byte
 ; X64-AVX-NEXT:    retq ## encoding: [0xc3]
 ;
 ; X64-AVX512VL-LABEL: test_x86_avx2_psrav_d_const:
 ; X64-AVX512VL:       ## %bb.0:
 ; X64-AVX512VL-NEXT:    vmovdqa {{.*}}(%rip), %xmm0 ## EVEX TO VEX Compression xmm0 = [2,9,4294967284,23]
 ; X64-AVX512VL-NEXT:    ## encoding: [0xc5,0xf9,0x6f,0x05,A,A,A,A]
-; X64-AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI86_0-4, kind: reloc_riprel_4byte
+; X64-AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI85_0-4, kind: reloc_riprel_4byte
 ; X64-AVX512VL-NEXT:    vpsravd {{.*}}(%rip), %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x46,0x05,A,A,A,A]
-; X64-AVX512VL-NEXT:    ## fixup A - offset: 5, value: LCPI86_1-4, kind: reloc_riprel_4byte
+; X64-AVX512VL-NEXT:    ## fixup A - offset: 5, value: LCPI85_1-4, kind: reloc_riprel_4byte
 ; X64-AVX512VL-NEXT:    retq ## encoding: [0xc3]
   %res = call <4 x i32> @llvm.x86.avx2.psrav.d(<4 x i32> <i32 2, i32 9, i32 -12, i32 23>, <4 x i32> <i32 1, i32 18, i32 35, i32 52>)
   ret <4 x i32> %res
 }
 declare <4 x i32> @llvm.x86.avx2.psrav.d(<4 x i32>, <4 x i32>) nounwind readnone
 
 define <8 x i32> @test_x86_avx2_psrav_d_256(<8 x i32> %a0, <8 x i32> %a1) {
 ; X86-AVX-LABEL: test_x86_avx2_psrav_d_256:
; X64-AVX512VL-NEXT:    retq ## encoding: [0xc3]
   ret <8 x i32> %res
 }
 
 define <8 x i32> @test_x86_avx2_psrav_d_256_const(<8 x i32> %a0, <8 x i32> %a1) {
 ; X86-AVX-LABEL: test_x86_avx2_psrav_d_256_const:
 ; X86-AVX:       ## %bb.0:
 ; X86-AVX-NEXT:    vmovdqa {{.*#+}} ymm0 = [2,9,4294967284,23,4294967270,37,4294967256,51]
 ; X86-AVX-NEXT:    ## encoding: [0xc5,0xfd,0x6f,0x05,A,A,A,A]
-; X86-AVX-NEXT:    ## fixup A - offset: 4, value: LCPI88_0, kind: FK_Data_4
-; X86-AVX-NEXT:    vpsravd LCPI88_1, %ymm0, %ymm0 ## encoding: [0xc4,0xe2,0x7d,0x46,0x05,A,A,A,A]
-; X86-AVX-NEXT:    ## fixup A - offset: 5, value: LCPI88_1, kind: FK_Data_4
+; X86-AVX-NEXT:    ## fixup A - offset: 4, value: LCPI87_0, kind: FK_Data_4
+; X86-AVX-NEXT:    vpsravd LCPI87_1, %ymm0, %ymm0 ## encoding: [0xc4,0xe2,0x7d,0x46,0x05,A,A,A,A]
+; X86-AVX-NEXT:    ## fixup A - offset: 5, value: LCPI87_1, kind: FK_Data_4
 ; X86-AVX-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X86-AVX512VL-LABEL: test_x86_avx2_psrav_d_256_const:
 ; X86-AVX512VL:       ## %bb.0:
-; X86-AVX512VL-NEXT:    vmovdqa LCPI88_0, %ymm0 ## EVEX TO VEX Compression ymm0 = [2,9,4294967284,23,4294967270,37,4294967256,51]
+; X86-AVX512VL-NEXT:    vmovdqa LCPI87_0, %ymm0 ## EVEX TO VEX Compression ymm0 = [2,9,4294967284,23,4294967270,37,4294967256,51]
 ; X86-AVX512VL-NEXT:    ## encoding: [0xc5,0xfd,0x6f,0x05,A,A,A,A]
-; X86-AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI88_0, kind: FK_Data_4
-; X86-AVX512VL-NEXT:    vpsravd LCPI88_1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x46,0x05,A,A,A,A]
-; X86-AVX512VL-NEXT:    ## fixup A - offset: 5, value: LCPI88_1, kind: FK_Data_4
+; X86-AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI87_0, kind: FK_Data_4
+; X86-AVX512VL-NEXT:    vpsravd LCPI87_1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x46,0x05,A,A,A,A]
+; X86-AVX512VL-NEXT:    ## fixup A - offset: 5, value: LCPI87_1, kind: FK_Data_4
 ; X86-AVX512VL-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X64-AVX-LABEL: test_x86_avx2_psrav_d_256_const:
 ; X64-AVX:       ## %bb.0:
 ; X64-AVX-NEXT:    vmovdqa {{.*#+}} ymm0 = [2,9,4294967284,23,4294967270,37,4294967256,51]
 ; X64-AVX-NEXT:    ## encoding: [0xc5,0xfd,0x6f,0x05,A,A,A,A]
-; X64-AVX-NEXT:    ## fixup A - offset: 4, value: LCPI88_0-4, kind: reloc_riprel_4byte
+; X64-AVX-NEXT:    ## fixup A - offset: 4, value: LCPI87_0-4, kind: reloc_riprel_4byte
 ; X64-AVX-NEXT:    vpsravd {{.*}}(%rip), %ymm0, %ymm0 ## encoding: [0xc4,0xe2,0x7d,0x46,0x05,A,A,A,A]
-; X64-AVX-NEXT:    ## fixup A - offset: 5, value: LCPI88_1-4, kind: reloc_riprel_4byte
+; X64-AVX-NEXT:    ## fixup A - offset: 5, value: LCPI87_1-4, kind: reloc_riprel_4byte
 ; X64-AVX-NEXT:    retq ## encoding: [0xc3]
 ;
 ; X64-AVX512VL-LABEL: test_x86_avx2_psrav_d_256_const:
 ; X64-AVX512VL:       ## %bb.0:
 ; X64-AVX512VL-NEXT:    vmovdqa {{.*}}(%rip), %ymm0 ## EVEX TO VEX Compression ymm0 = [2,9,4294967284,23,4294967270,37,4294967256,51]
 ; X64-AVX512VL-NEXT:    ## encoding: [0xc5,0xfd,0x6f,0x05,A,A,A,A]
-; X64-AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI88_0-4, kind: reloc_riprel_4byte
+; X64-AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI87_0-4, kind: reloc_riprel_4byte
 ; X64-AVX512VL-NEXT:    vpsravd {{.*}}(%rip), %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x46,0x05,A,A,A,A]
-; X64-AVX512VL-NEXT:    ## fixup A - offset: 5, value: LCPI88_1-4, kind: reloc_riprel_4byte
+; X64-AVX512VL-NEXT:    ## fixup A - offset: 5, value: LCPI87_1-4, kind: reloc_riprel_4byte
 ; X64-AVX512VL-NEXT:    retq ## encoding: [0xc3]
   %res = call <8 x i32> @llvm.x86.avx2.psrav.d.256(<8 x i32> <i32 2, i32 9, i32 -12, i32 23, i32 -26, i32 37, i32 -40, i32 51>, <8 x i32> <i32 1, i32 18, i32 35, i32 52, i32 69, i32 15, i32 32, i32 49>)
   ret <8 x i32> %res
 }
 declare <8 x i32> @llvm.x86.avx2.psrav.d.256(<8 x i32>, <8 x i32>) nounwind readnone
 
 define <2 x double> @test_x86_avx2_gather_d_pd(<2 x double> %a0, i8* %a1, <4 x i32> %idx, <2 x double> %mask) {
 ; X86-LABEL: test_x86_avx2_gather_d_pd:

llvm/test/CodeGen/X86/avx512bw-intrinsics-upgrade.ll

; AVX512F-32-NEXT:    retl
     %res = call <32 x i16> @llvm.x86.avx512.mask.pbroadcast.w.gpr.512(i16 %x0, <32 x i16> %x1, i32 -1)
     %res1 = call <32 x i16> @llvm.x86.avx512.mask.pbroadcast.w.gpr.512(i16 %x0, <32 x i16> %x1, i32 %mask)
    %res2 = call <32 x i16> @llvm.x86.avx512.mask.pbroadcast.w.gpr.512(i16 %x0, <32 x i16> zeroinitializer, i32 %mask)
     %res3 = add <32 x i16> %res, %res1
    %res4 = add <32 x i16> %res2, %res3
     ret <32 x i16> %res4
  }
 
+declare  <8 x i64> @llvm.x86.avx512.psad.bw.512(<64 x i8>, <64 x i8>)
+
+define  <8 x i64>@test_int_x86_avx512_mask_psadb_w_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2){
+; AVX512BW-LABEL: test_int_x86_avx512_mask_psadb_w_512:
+; AVX512BW:       ## %bb.0:
+; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vpsadbw %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm0, %zmm1, %zmm0
+; AVX512BW-NEXT:    retq
+;
+; AVX512F-32-LABEL: test_int_x86_avx512_mask_psadb_w_512:
+; AVX512F-32:       # %bb.0:
+; AVX512F-32-NEXT:    vpsadbw %zmm1, %zmm0, %zmm1
+; AVX512F-32-NEXT:    vpsadbw %zmm2, %zmm0, %zmm0
+; AVX512F-32-NEXT:    vpaddq %zmm0, %zmm1, %zmm0
+; AVX512F-32-NEXT:    retl
+  %res = call  <8 x i64> @llvm.x86.avx512.psad.bw.512(<64 x i8> %x0, <64 x i8> %x1)
+  %res1 = call  <8 x i64> @llvm.x86.avx512.psad.bw.512(<64 x i8> %x0, <64 x i8> %x2)
+  %res2 = add  <8 x i64> %res, %res1
+  ret  <8 x i64> %res2
+}
+
 declare void @llvm.x86.avx512.mask.storeu.b.512(i8*, <64 x i8>, i64)
 
 define void@test_int_x86_avx512_mask_storeu_b_512(i8* %ptr1, i8* %ptr2, <64 x i8> %x1, i64 %x2) {
 ; AVX512BW-LABEL: test_int_x86_avx512_mask_storeu_b_512:
 ; AVX512BW:       ## %bb.0:
 ; AVX512BW-NEXT:    kmovq %rdx, %k1
 ; AVX512BW-NEXT:    vmovdqu8 %zmm0, (%rdi) {%k1}
 ; AVX512BW-NEXT:    vmovdqu64 %zmm0, (%rsi)

llvm/test/CodeGen/X86/avx512bw-intrinsics.ll

; AVX512F-32-NEXT:    retl
   %res = call <32 x i16> @llvm.x86.avx512.mask.dbpsadbw.512(<64 x i8> %x0, <64 x i8> %x1, i32 2, <32 x i16> %x3, i32 %x4)
   %res1 = call <32 x i16> @llvm.x86.avx512.mask.dbpsadbw.512(<64 x i8> %x0, <64 x i8> %x1, i32 2, <32 x i16> zeroinitializer, i32 %x4)
   %res2 = call <32 x i16> @llvm.x86.avx512.mask.dbpsadbw.512(<64 x i8> %x0, <64 x i8> %x1, i32 2, <32 x i16> %x3, i32 -1)
   %res3 = add <32 x i16> %res, %res1
   %res4 = add <32 x i16> %res3, %res2
   ret <32 x i16> %res4
 }
 
-declare  <8 x i64> @llvm.x86.avx512.psad.bw.512(<64 x i8>, <64 x i8>)
-
-define  <8 x i64>@test_int_x86_avx512_mask_psadb_w_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2){
-; AVX512BW-LABEL: test_int_x86_avx512_mask_psadb_w_512:
-; AVX512BW:       ## %bb.0:
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm1
-; AVX512BW-NEXT:    vpsadbw %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpaddq %zmm0, %zmm1, %zmm0
-; AVX512BW-NEXT:    retq
-;
-; AVX512F-32-LABEL: test_int_x86_avx512_mask_psadb_w_512:
-; AVX512F-32:       # %bb.0:
-; AVX512F-32-NEXT:    vpsadbw %zmm1, %zmm0, %zmm1
-; AVX512F-32-NEXT:    vpsadbw %zmm2, %zmm0, %zmm0
-; AVX512F-32-NEXT:    vpaddq %zmm0, %zmm1, %zmm0
-; AVX512F-32-NEXT:    retl
-  %res = call  <8 x i64> @llvm.x86.avx512.psad.bw.512(<64 x i8> %x0, <64 x i8> %x1)
-  %res1 = call  <8 x i64> @llvm.x86.avx512.psad.bw.512(<64 x i8> %x0, <64 x i8> %x2)
-  %res2 = add  <8 x i64> %res, %res1
-  ret  <8 x i64> %res2
-}
-
 declare <32 x i16> @llvm.x86.avx512.mask.psrlv32hi(<32 x i16>, <32 x i16>, <32 x i16>, i32)
 
 define <32 x i16>@test_int_x86_avx512_mask_psrlv32hi(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
 ; AVX512BW-LABEL: test_int_x86_avx512_mask_psrlv32hi:
 ; AVX512BW:       ## %bb.0:
 ; AVX512BW-NEXT:    vpsrlvw %zmm1, %zmm0, %zmm3
 ; AVX512BW-NEXT:    kmovd %edi, %k1
 ; AVX512BW-NEXT:    vpsrlvw %zmm1, %zmm0, %zmm2 {%k1}

llvm/test/CodeGen/X86/sse2-intrinsics-fast-isel.ll

; X64-NEXT:    retq
   call void @llvm.x86.sse2.pause()
   ret void
 }
 declare void @llvm.x86.sse2.pause() nounwind readnone
 
 define <2 x i64> @test_mm_sad_epu8(<2 x i64> %a0, <2 x i64> %a1) nounwind {
 ; X32-LABEL: test_mm_sad_epu8:
 ; X32:       # %bb.0:
-; X32-NEXT:    psadbw %xmm1, %xmm0
+; X32-NEXT:    movdqa %xmm0, %xmm2
+; X32-NEXT:    pminub %xmm1, %xmm0
+; X32-NEXT:    pcmpeqb %xmm2, %xmm0
+; X32-NEXT:    pcmpeqd %xmm3, %xmm3
+; X32-NEXT:    pxor %xmm0, %xmm3
+; X32-NEXT:    movdqa %xmm2, %xmm4
+; X32-NEXT:    psubb %xmm1, %xmm4
+; X32-NEXT:    psubb %xmm2, %xmm1
+; X32-NEXT:    pandn %xmm1, %xmm3
+; X32-NEXT:    pandn %xmm4, %xmm0
+; X32-NEXT:    por %xmm3, %xmm0
+; X32-NEXT:    movdqa {{.*#+}} xmm2 = [255,0,0,0,0,0,0,0,255,0,0,0,0,0,0,0]
+; X32-NEXT:    movdqa %xmm0, %xmm1
+; X32-NEXT:    pand %xmm2, %xmm1
+; X32-NEXT:    movdqa %xmm0, %xmm3
+; X32-NEXT:    psrlw $8, %xmm3
+; X32-NEXT:    pand %xmm2, %xmm3
+; X32-NEXT:    movdqa %xmm0, %xmm4
+; X32-NEXT:    psrld $16, %xmm4
+; X32-NEXT:    pand %xmm2, %xmm4
+; X32-NEXT:    paddq %xmm3, %xmm4
+; X32-NEXT:    movdqa %xmm0, %xmm3
+; X32-NEXT:    psrld $24, %xmm3
+; X32-NEXT:    pand %xmm2, %xmm3
+; X32-NEXT:    paddq %xmm4, %xmm3
+; X32-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
+; X32-NEXT:    pand %xmm2, %xmm4
+; X32-NEXT:    paddq %xmm3, %xmm4
+; X32-NEXT:    movdqa %xmm0, %xmm3
+; X32-NEXT:    psrlq $40, %xmm3
+; X32-NEXT:    pand %xmm2, %xmm3
+; X32-NEXT:    paddq %xmm4, %xmm3
+; X32-NEXT:    movdqa %xmm0, %xmm4
+; X32-NEXT:    psrlq $48, %xmm4
+; X32-NEXT:    pand %xmm2, %xmm4
+; X32-NEXT:    paddq %xmm3, %xmm4
+; X32-NEXT:    psrlq $56, %xmm0
+; X32-NEXT:    paddq %xmm4, %xmm0
+; X32-NEXT:    paddq %xmm1, %xmm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: test_mm_sad_epu8:
 ; X64:       # %bb.0:
-; X64-NEXT:    psadbw %xmm1, %xmm0
+; X64-NEXT:    movdqa %xmm0, %xmm2
+; X64-NEXT:    pminub %xmm1, %xmm0
+; X64-NEXT:    pcmpeqb %xmm2, %xmm0
+; X64-NEXT:    pcmpeqd %xmm3, %xmm3
+; X64-NEXT:    pxor %xmm0, %xmm3
+; X64-NEXT:    movdqa %xmm2, %xmm4
+; X64-NEXT:    psubb %xmm1, %xmm4
+; X64-NEXT:    psubb %xmm2, %xmm1
+; X64-NEXT:    pandn %xmm1, %xmm3
+; X64-NEXT:    pandn %xmm4, %xmm0
+; X64-NEXT:    por %xmm3, %xmm0
+; X64-NEXT:    movdqa {{.*#+}} xmm2 = [255,0,0,0,0,0,0,0,255,0,0,0,0,0,0,0]
+; X64-NEXT:    movdqa %xmm0, %xmm1
+; X64-NEXT:    pand %xmm2, %xmm1
+; X64-NEXT:    movdqa %xmm0, %xmm3
+; X64-NEXT:    psrlw $8, %xmm3
+; X64-NEXT:    pand %xmm2, %xmm3
+; X64-NEXT:    movdqa %xmm0, %xmm4
+; X64-NEXT:    psrld $16, %xmm4
+; X64-NEXT:    pand %xmm2, %xmm4
+; X64-NEXT:    paddq %xmm3, %xmm4
+; X64-NEXT:    movdqa %xmm0, %xmm3
+; X64-NEXT:    psrld $24, %xmm3
+; X64-NEXT:    pand %xmm2, %xmm3
+; X64-NEXT:    paddq %xmm4, %xmm3
+; X64-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
+; X64-NEXT:    pand %xmm2, %xmm4
+; X64-NEXT:    paddq %xmm3, %xmm4
+; X64-NEXT:    movdqa %xmm0, %xmm3
+; X64-NEXT:    psrlq $40, %xmm3
+; X64-NEXT:    pand %xmm2, %xmm3
+; X64-NEXT:    paddq %xmm4, %xmm3
+; X64-NEXT:    movdqa %xmm0, %xmm4
+; X64-NEXT:    psrlq $48, %xmm4
+; X64-NEXT:    pand %xmm2, %xmm4
+; X64-NEXT:    paddq %xmm3, %xmm4
+; X64-NEXT:    psrlq $56, %xmm0
+; X64-NEXT:    paddq %xmm4, %xmm0
+; X64-NEXT:    paddq %xmm1, %xmm0
 ; X64-NEXT:    retq
   %arg0 = bitcast <2 x i64> %a0 to <16 x i8>
   %arg1 = bitcast <2 x i64> %a1 to <16 x i8>
-  %res = call <2 x i64> @llvm.x86.sse2.psad.bw(<16 x i8> %arg0, <16 x i8> %arg1)
+  %1 = icmp ugt <16 x i8> %arg0, %arg1
+  %2 = sub <16 x i8> %arg0, %arg1
+  %3 = sub <16 x i8> %arg1, %arg0
+  %4 = select <16 x i1> %1, <16 x i8> %2, <16 x i8> %3
+  %5 = shufflevector <16 x i8> %4, <16 x i8> %4, <2 x i32> <i32 0, i32 8>
+  %6 = zext <2 x i8> %5 to <2 x i64>
+  %7 = shufflevector <16 x i8> %4, <16 x i8> %4, <2 x i32> <i32 1, i32 9>
+  %8 = zext <2 x i8> %7 to <2 x i64>
+  %9 = add <2 x i64> %6, %8
+  %10 = shufflevector <16 x i8> %4, <16 x i8> %4, <2 x i32> <i32 2, i32 10>
+  %11 = zext <2 x i8> %10 to <2 x i64>
+  %12 = add <2 x i64> %9, %11
+  %13 = shufflevector <16 x i8> %4, <16 x i8> %4, <2 x i32> <i32 3, i32 11>
+  %14 = zext <2 x i8> %13 to <2 x i64>
+  %15 = add <2 x i64> %12, %14
+  %16 = shufflevector <16 x i8> %4, <16 x i8> %4, <2 x i32> <i32 4, i32 12>
+  %17 = zext <2 x i8> %16 to <2 x i64>
+  %18 = add <2 x i64> %15, %17
+  %19 = shufflevector <16 x i8> %4, <16 x i8> %4, <2 x i32> <i32 5, i32 13>
+  %20 = zext <2 x i8> %19 to <2 x i64>
+  %21 = add <2 x i64> %18, %20
+  %22 = shufflevector <16 x i8> %4, <16 x i8> %4, <2 x i32> <i32 6, i32 14>
+  %23 = zext <2 x i8> %22 to <2 x i64>
+  %24 = add <2 x i64> %21, %23
+  %25 = shufflevector <16 x i8> %4, <16 x i8> %4, <2 x i32> <i32 7, i32 15>
+  %26 = zext <2 x i8> %25 to <2 x i64>
+  %res = add <2 x i64> %24, %26
   ret <2 x i64> %res
 }
-declare <2 x i64> @llvm.x86.sse2.psad.bw(<16 x i8>, <16 x i8>) nounwind readnone
 
 define <2 x i64> @test_mm_set_epi8(i8 %a0, i8 %a1, i8 %a2, i8 %a3, i8 %a4, i8 %a5, i8 %a6, i8 %a7, i8 %a8, i8 %a9, i8 %a10, i8 %a11, i8 %a12, i8 %a13, i8 %a14, i8 %a15) nounwind {
 ; X32-LABEL: test_mm_set_epi8:
 ; X32:       # %bb.0:
 ; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
 ; X32-NEXT:    movd %eax, %xmm0
 ; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
 ; X32-NEXT:    movd %eax, %xmm1

llvm/test/CodeGen/X86/sse2-intrinsics-x86-upgrade.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+sse2 | FileCheck %s
 
+define <2 x i64> @test_x86_sse2_psad_bw(<16 x i8> %a0, <16 x i8> %a1) {
+; CHECK-LABEL: test_x86_sse2_psad_bw:
+; CHECK:       ## %bb.0:
+; CHECK-NEXT:    psadbw %xmm1, %xmm0
+; CHECK-NEXT:    retl
+  %res = call <2 x i64> @llvm.x86.sse2.psad.bw(<16 x i8> %a0, <16 x i8> %a1) ; <<2 x i64>> [#uses=1]
+  ret <2 x i64> %res
+}
+declare <2 x i64> @llvm.x86.sse2.psad.bw(<16 x i8>, <16 x i8>) nounwind readnone
+
+
 define <2 x i64> @test_x86_sse2_psll_dq_bs(<2 x i64> %a0) {
 ; CHECK-LABEL: test_x86_sse2_psll_dq_bs:
 ; CHECK:       ## %bb.0:
 ; CHECK-NEXT:    pslldq {{.*#+}} xmm0 = zero,zero,zero,zero,zero,zero,zero,xmm0[0,1,2,3,4,5,6,7,8]
 ; CHECK-NEXT:    retl
   %res = call <2 x i64> @llvm.x86.sse2.psll.dq.bs(<2 x i64> %a0, i32 7) ; <<2 x i64>> [#uses=1]
   ret <2 x i64> %res
 }

llvm/test/CodeGen/X86/sse2-intrinsics-x86.ll

 ; SKX-NEXT:    vpmulhuw %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xe4,0xc1]
 ; SKX-NEXT:    retl ## encoding: [0xc3]
   %res = call <8 x i16> @llvm.x86.sse2.pmulhu.w(<8 x i16> %a0, <8 x i16> %a1) ; <<8 x i16>> [#uses=1]
   ret <8 x i16> %res
 }
 declare <8 x i16> @llvm.x86.sse2.pmulhu.w(<8 x i16>, <8 x i16>) nounwind readnone
 
 
-define <2 x i64> @test_x86_sse2_psad_bw(<16 x i8> %a0, <16 x i8> %a1) {
-; SSE-LABEL: test_x86_sse2_psad_bw:
-; SSE:       ## %bb.0:
-; SSE-NEXT:    psadbw %xmm1, %xmm0 ## encoding: [0x66,0x0f,0xf6,0xc1]
-; SSE-NEXT:    retl ## encoding: [0xc3]
-;
-; AVX2-LABEL: test_x86_sse2_psad_bw:
-; AVX2:       ## %bb.0:
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0xf6,0xc1]
-; AVX2-NEXT:    retl ## encoding: [0xc3]
-;
-; SKX-LABEL: test_x86_sse2_psad_bw:
-; SKX:       ## %bb.0:
-; SKX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xf6,0xc1]
-; SKX-NEXT:    retl ## encoding: [0xc3]
-  %res = call <2 x i64> @llvm.x86.sse2.psad.bw(<16 x i8> %a0, <16 x i8> %a1) ; <<2 x i64>> [#uses=1]
-  ret <2 x i64> %res
-}
-declare <2 x i64> @llvm.x86.sse2.psad.bw(<16 x i8>, <16 x i8>) nounwind readnone
-
-
 define <4 x i32> @test_x86_sse2_psll_d(<4 x i32> %a0, <4 x i32> %a1) {
 ; SSE-LABEL: test_x86_sse2_psll_d:
 ; SSE:       ## %bb.0:
 ; SSE-NEXT:    pslld %xmm1, %xmm0 ## encoding: [0x66,0x0f,0xf2,0xc1]
 ; SSE-NEXT:    retl ## encoding: [0xc3]
 ;
 ; AVX2-LABEL: test_x86_sse2_psll_d:
 ; AVX2:       ## %bb.0:

llvm/test/Instrumentation/MemorySanitizer/vector_arith.ll

 ; CHECK: or i64
 ; CHECK: bitcast i64 {{.*}} to <4 x i16>
 ; CHECK: icmp ne <4 x i16> {{.*}}, zeroinitializer
 ; CHECK: sext <4 x i1> {{.*}} to <4 x i16>
 ; CHECK: bitcast <4 x i16> {{.*}} to i64
 ; CHECK: ret x86_mmx
 
 
-define <2 x i64> @Test_x86_sse2_psad_bw(<16 x i8> %a, <16 x i8> %b) sanitize_memory {
-  %c = tail call <2 x i64> @llvm.x86.sse2.psad.bw(<16 x i8> %a, <16 x i8> %b)
-  ret <2 x i64> %c
-}
-
-; CHECK-LABEL: @Test_x86_sse2_psad_bw(
-; CHECK: or <16 x i8> {{.*}}, {{.*}}
-; CHECK: bitcast <16 x i8> {{.*}} to <2 x i64>
-; CHECK: icmp ne <2 x i64> {{.*}}, zeroinitializer
-; CHECK: sext <2 x i1> {{.*}} to <2 x i64>
-; CHECK: lshr <2 x i64> {{.*}}, <i64 48, i64 48>
-; CHECK: ret <2 x i64>
-
-
 define x86_mmx @Test_x86_mmx_psad_bw(x86_mmx %a, x86_mmx %b) sanitize_memory {
 entry:
   %c = tail call x86_mmx @llvm.x86.mmx.psad.bw(x86_mmx %a, x86_mmx %b) nounwind
   ret x86_mmx %c
 }
 
 ; CHECK-LABEL: @Test_x86_mmx_psad_bw(
 ; CHECK: or i64
 ; CHECK: icmp ne i64
 ; CHECK: sext i1 {{.*}} to i64
 ; CHECK: lshr i64 {{.*}}, 48
 ; CHECK: ret x86_mmx