Remove NEON vaddl, vaddw, vsubl, and vsubw intrinsics. Instead, use llvm

IR add/sub operations with one or both operands sign- or zero-extended.
Auto-upgrade the old intrinsics.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@112416 91177308-0d34-0410-b5e6-96231b3b80d8

7 files changed, 225 insertions, 124 deletions

diff --git a/include/llvm/IntrinsicsARM.td b/include/llvm/IntrinsicsARM.td
index 7be283b29c..2528eaa973 100644
--- a/include/llvm/IntrinsicsARM.td
+++ b/include/llvm/IntrinsicsARM.td
@@ -73,10 +73,6 @@ let TargetPrefix = "arm" in {  // All intrinsics start with "llvm.arm.".
                 [LLVMTruncatedElementVectorType<0>,
                  LLVMTruncatedElementVectorType<0>],
                 [IntrNoMem]>;
-  class Neon_2Arg_Wide_Intrinsic
-    : Intrinsic<[llvm_anyvector_ty],
-                [LLVMMatchType<0>, LLVMTruncatedElementVectorType<0>],
-                [IntrNoMem]>;
   class Neon_3Arg_Intrinsic
     : Intrinsic<[llvm_anyvector_ty],
                 [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>],
@@ -128,10 +124,6 @@ let Properties = [IntrNoMem, Commutative] in {
   def int_arm_neon_vqaddu : Neon_2Arg_Intrinsic;
   def int_arm_neon_vaddhn : Neon_2Arg_Narrow_Intrinsic;
   def int_arm_neon_vraddhn : Neon_2Arg_Narrow_Intrinsic;
-  def int_arm_neon_vaddls : Neon_2Arg_Long_Intrinsic;
-  def int_arm_neon_vaddlu : Neon_2Arg_Long_Intrinsic;
-  def int_arm_neon_vaddws : Neon_2Arg_Wide_Intrinsic;
-  def int_arm_neon_vaddwu : Neon_2Arg_Wide_Intrinsic;
 
   // Vector Multiply.
   def int_arm_neon_vmulp : Neon_2Arg_Intrinsic;
@@ -172,10 +164,6 @@ def int_arm_neon_vqsubs : Neon_2Arg_Intrinsic;
 def int_arm_neon_vqsubu : Neon_2Arg_Intrinsic;
 def int_arm_neon_vsubhn : Neon_2Arg_Narrow_Intrinsic;
 def int_arm_neon_vrsubhn : Neon_2Arg_Narrow_Intrinsic;
-def int_arm_neon_vsubls : Neon_2Arg_Long_Intrinsic;
-def int_arm_neon_vsublu : Neon_2Arg_Long_Intrinsic;
-def int_arm_neon_vsubws : Neon_2Arg_Wide_Intrinsic;
-def int_arm_neon_vsubwu : Neon_2Arg_Wide_Intrinsic;
 
 // Vector Absolute Compare.
 let TargetPrefix = "arm" in {
diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 88d606c0f5..741df7e9d4 100644
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -1294,6 +1294,19 @@ class N3VNInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
   let isCommutable = Commutable;
 }
 
+// Long 3-register operations.
+class N3VL<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
+           InstrItinClass itin, string OpcodeStr, string Dt,
+           ValueType TyQ, ValueType TyD, SDNode OpNode, SDNode ExtOp,
+           bit Commutable>
+  : N3V<op24, op23, op21_20, op11_8, 0, op4,
+        (outs QPR:$dst), (ins DPR:$src1, DPR:$src2), N3RegFrm, itin,
+        OpcodeStr, Dt, "$dst, $src1, $src2", "",
+        [(set QPR:$dst, (OpNode (TyQ (ExtOp (TyD DPR:$src1))),
+                                (TyQ (ExtOp (TyD DPR:$src2)))))]> {
+  let isCommutable = Commutable;
+}
+
 // Long 3-register intrinsics.
 class N3VLInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
               InstrItinClass itin, string OpcodeStr, string Dt,
@@ -1325,14 +1338,15 @@ class N3VLIntSL16<bit op24, bits<2> op21_20, bits<4> op11_8,
                             (OpTy (NEONvduplane (OpTy DPR_8:$src2),
                                                 imm:$lane)))))]>;
 
-// Wide 3-register intrinsics.
-class N3VWInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
-              string OpcodeStr, string Dt, ValueType TyQ, ValueType TyD,
-              Intrinsic IntOp, bit Commutable>
+// Wide 3-register operations.
+class N3VW<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
+           string OpcodeStr, string Dt, ValueType TyQ, ValueType TyD,
+           SDNode OpNode, SDNode ExtOp, bit Commutable>
   : N3V<op24, op23, op21_20, op11_8, 0, op4,
         (outs QPR:$dst), (ins QPR:$src1, DPR:$src2), N3RegFrm, IIC_VSUBiD,
         OpcodeStr, Dt, "$dst, $src1, $src2", "",
-        [(set QPR:$dst, (TyQ (IntOp (TyQ QPR:$src1), (TyD DPR:$src2))))]> {
+        [(set QPR:$dst, (OpNode (TyQ QPR:$src1),
+                                (TyQ (ExtOp (TyD DPR:$src2)))))]> {
   let isCommutable = Commutable;
 }
 
@@ -1684,6 +1698,23 @@ multiclass N3VNInt_HSD<bit op24, bit op23, bits<4> op11_8, bit op4,
 }
 
 
+// Neon Long 3-register vector operations.
+
+multiclass N3VL_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
+                    InstrItinClass itin16, InstrItinClass itin32,
+                    string OpcodeStr, string Dt,
+                    SDNode OpNode, SDNode ExtOp, bit Commutable = 0> {
+  def v4i32 : N3VL<op24, op23, 0b01, op11_8, op4, itin16, 
+                   OpcodeStr, !strconcat(Dt, "16"),
+                   v4i32, v4i16, OpNode, ExtOp, Commutable>;
+  def v2i64 : N3VL<op24, op23, 0b10, op11_8, op4, itin32,
+                   OpcodeStr, !strconcat(Dt, "32"),
+                   v2i64, v2i32, OpNode, ExtOp, Commutable>;
+  def v8i16 : N3VL<op24, op23, 0b00, op11_8, op4, itin16,
+                   OpcodeStr, !strconcat(Dt, "8"),
+                   v8i16, v8i8, OpNode, ExtOp, Commutable>;
+}
+
 // Neon Long 3-register vector intrinsics.
 
 // First with only element sizes of 16 and 32 bits:
@@ -1723,18 +1754,18 @@ multiclass N3VLInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
 
 // Neon Wide 3-register vector intrinsics,
 //   source operand element sizes of 8, 16 and 32 bits:
-multiclass N3VWInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
-                       string OpcodeStr, string Dt,
-                       Intrinsic IntOp, bit Commutable = 0> {
-  def v8i16 : N3VWInt<op24, op23, 0b00, op11_8, op4,
-                      OpcodeStr, !strconcat(Dt, "8"),
-                      v8i16, v8i8, IntOp, Commutable>;
-  def v4i32 : N3VWInt<op24, op23, 0b01, op11_8, op4,
-                      OpcodeStr, !strconcat(Dt, "16"),
-                      v4i32, v4i16, IntOp, Commutable>;
-  def v2i64 : N3VWInt<op24, op23, 0b10, op11_8, op4,
-                      OpcodeStr, !strconcat(Dt, "32"),
-                      v2i64, v2i32, IntOp, Commutable>;
+multiclass N3VW_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
+                    string OpcodeStr, string Dt,
+                    SDNode OpNode, SDNode ExtOp, bit Commutable = 0> {
+  def v8i16 : N3VW<op24, op23, 0b00, op11_8, op4,
+                   OpcodeStr, !strconcat(Dt, "8"),
+                   v8i16, v8i8, OpNode, ExtOp, Commutable>;
+  def v4i32 : N3VW<op24, op23, 0b01, op11_8, op4,
+                   OpcodeStr, !strconcat(Dt, "16"),
+                   v4i32, v4i16, OpNode, ExtOp, Commutable>;
+  def v2i64 : N3VW<op24, op23, 0b10, op11_8, op4,
+                   OpcodeStr, !strconcat(Dt, "32"),
+                   v2i64, v2i32, OpNode, ExtOp, Commutable>;
 }
 
 
@@ -2073,13 +2104,13 @@ def  VADDfd   : N3VD<0, 0, 0b00, 0b1101, 0, IIC_VBIND, "vadd", "f32",
 def  VADDfq   : N3VQ<0, 0, 0b00, 0b1101, 0, IIC_VBINQ, "vadd", "f32",
                      v4f32, v4f32, fadd, 1>;
 //   VADDL    : Vector Add Long (Q = D + D)
-defm VADDLs   : N3VLInt_QHS<0,1,0b0000,0, IIC_VSHLiD, IIC_VSHLiD,
-                            "vaddl", "s", int_arm_neon_vaddls, 1>;
-defm VADDLu   : N3VLInt_QHS<1,1,0b0000,0, IIC_VSHLiD, IIC_VSHLiD,
-                            "vaddl", "u", int_arm_neon_vaddlu, 1>;
+defm VADDLs   : N3VL_QHS<0,1,0b0000,0, IIC_VSHLiD, IIC_VSHLiD,
+                         "vaddl", "s", add, sext, 1>;
+defm VADDLu   : N3VL_QHS<1,1,0b0000,0, IIC_VSHLiD, IIC_VSHLiD,
+                         "vaddl", "u", add, zext, 1>;
 //   VADDW    : Vector Add Wide (Q = Q + D)
-defm VADDWs   : N3VWInt_QHS<0,1,0b0001,0, "vaddw", "s", int_arm_neon_vaddws, 0>;
-defm VADDWu   : N3VWInt_QHS<1,1,0b0001,0, "vaddw", "u", int_arm_neon_vaddwu, 0>;
+defm VADDWs   : N3VW_QHS<0,1,0b0001,0, "vaddw", "s", add, sext, 0>;
+defm VADDWu   : N3VW_QHS<1,1,0b0001,0, "vaddw", "u", add, zext, 0>;
 //   VHADD    : Vector Halving Add
 defm VHADDs   : N3VInt_QHS<0, 0, 0b0000, 0, N3RegFrm,
                            IIC_VBINi4D, IIC_VBINi4D, IIC_VBINi4Q, IIC_VBINi4Q,
@@ -2324,13 +2355,13 @@ def  VSUBfd   : N3VD<0, 0, 0b10, 0b1101, 0, IIC_VBIND, "vsub", "f32",
 def  VSUBfq   : N3VQ<0, 0, 0b10, 0b1101, 0, IIC_VBINQ, "vsub", "f32",
                      v4f32, v4f32, fsub, 0>;
 //   VSUBL    : Vector Subtract Long (Q = D - D)
-defm VSUBLs   : N3VLInt_QHS<0,1,0b0010,0, IIC_VSHLiD, IIC_VSHLiD,
-                            "vsubl", "s", int_arm_neon_vsubls, 1>;
-defm VSUBLu   : N3VLInt_QHS<1,1,0b0010,0, IIC_VSHLiD, IIC_VSHLiD,
-                            "vsubl", "u", int_arm_neon_vsublu, 1>;
+defm VSUBLs   : N3VL_QHS<0,1,0b0010,0, IIC_VSHLiD, IIC_VSHLiD,
+                         "vsubl", "s", sub, sext, 0>;
+defm VSUBLu   : N3VL_QHS<1,1,0b0010,0, IIC_VSHLiD, IIC_VSHLiD,
+                         "vsubl", "u", sub, zext, 0>;
 //   VSUBW    : Vector Subtract Wide (Q = Q - D)
-defm VSUBWs   : N3VWInt_QHS<0,1,0b0011,0, "vsubw", "s", int_arm_neon_vsubws, 0>;
-defm VSUBWu   : N3VWInt_QHS<1,1,0b0011,0, "vsubw", "u", int_arm_neon_vsubwu, 0>;
+defm VSUBWs   : N3VW_QHS<0,1,0b0011,0, "vsubw", "s", sub, sext, 0>;
+defm VSUBWu   : N3VW_QHS<1,1,0b0011,0, "vsubw", "u", sub, zext, 0>;
 //   VHSUB    : Vector Halving Subtract
 defm VHSUBs   : N3VInt_QHS<0, 0, 0b0010, 0, N3RegFrm,
                            IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q, IIC_VSUBi4Q,
@@ -2559,7 +2590,7 @@ def  VABDfq   : N3VQInt<1, 0, 0b10, 0b1101, 0, N3RegFrm, IIC_VBINQ,
 defm VABDLs   : N3VLInt_QHS<0,1,0b0111,0, IIC_VSUBi4Q, IIC_VSUBi4Q,
                             "vabdl", "s", int_arm_neon_vabdls, 0>;
 defm VABDLu   : N3VLInt_QHS<1,1,0b0111,0, IIC_VSUBi4Q, IIC_VSUBi4Q,
-                             "vabdl", "u", int_arm_neon_vabdlu, 0>;
+                            "vabdl", "u", int_arm_neon_vabdlu, 0>;
 
 //   VABA     : Vector Absolute Difference and Accumulate
 defm VABAs    : N3VInt3_QHS<0,0,0b0111,1, IIC_VABAD, IIC_VABAQ,
diff --git a/lib/VMCore/AutoUpgrade.cpp b/lib/VMCore/AutoUpgrade.cpp
index 052fd2d5b1..62a46259ac 100644
--- a/lib/VMCore/AutoUpgrade.cpp
+++ b/lib/VMCore/AutoUpgrade.cpp
@@ -79,8 +79,17 @@ static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
         return true;
       }
     } else if (Name.compare(5, 9, "arm.neon.", 9) == 0) {
-      if (Name.compare(14, 7, "vmovls.", 7) == 0 ||
-          Name.compare(14, 7, "vmovlu.", 7) == 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))) {
+
         // Calls to these are transformed into IR without intrinsics.
         NewFn = 0;
         return true;
@@ -371,6 +380,27 @@ void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
       } 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 {
         llvm_unreachable("Unknown arm.neon function for CallInst upgrade.");
       }
diff --git a/test/Bitcode/neon-intrinsics.ll b/test/Bitcode/neon-intrinsics.ll
index fe76514ee0..d6d03d0970 100644
--- a/test/Bitcode/neon-intrinsics.ll
+++ b/test/Bitcode/neon-intrinsics.ll
@@ -28,6 +28,54 @@
 ; CHECK-NOT: arm.neon.vmovlu.v2i64
 ; CHECK: zext <2 x i32>
 
+; vaddl/vaddw should be auto-upgraded to add with sext/zext
+
+; CHECK: vaddls16
+; CHECK-NOT: arm.neon.vaddls.v4i32
+; CHECK: sext <4 x i16>
+; CHECK-NEXT: sext <4 x i16>
+; CHECK-NEXT: add <4 x i32>
+
+; CHECK: vaddlu32
+; CHECK-NOT: arm.neon.vaddlu.v2i64
+; CHECK: zext <2 x i32>
+; CHECK-NEXT: zext <2 x i32>
+; CHECK-NEXT: add <2 x i64>
+
+; CHECK: vaddws8
+; CHECK-NOT: arm.neon.vaddws.v8i16
+; CHECK: sext <8 x i8>
+; CHECK-NEXT: add <8 x i16>
+
+; CHECK: vaddwu16
+; CHECK-NOT: arm.neon.vaddwu.v4i32
+; CHECK: zext <4 x i16>
+; CHECK-NEXT: add <4 x i32>
+
+; vsubl/vsubw should be auto-upgraded to sub with sext/zext
+
+; CHECK: vsubls16
+; CHECK-NOT: arm.neon.vsubls.v4i32
+; CHECK: sext <4 x i16>
+; CHECK-NEXT: sext <4 x i16>
+; CHECK-NEXT: sub <4 x i32>
+
+; CHECK: vsublu32
+; CHECK-NOT: arm.neon.vsublu.v2i64
+; CHECK: zext <2 x i32>
+; CHECK-NEXT: zext <2 x i32>
+; CHECK-NEXT: sub <2 x i64>
+
+; CHECK: vsubws8
+; CHECK-NOT: arm.neon.vsubws.v8i16
+; CHECK: sext <8 x i8>
+; CHECK-NEXT: sub <8 x i16>
+
+; CHECK: vsubwu16
+; CHECK-NOT: arm.neon.vsubwu.v4i32
+; CHECK: zext <4 x i16>
+; CHECK-NEXT: sub <4 x i32>
+
 ; vld* and vst* intrinsic calls need an alignment argument (defaulted to 1)
 
 ; CHECK: vld1i8
diff --git a/test/Bitcode/neon-intrinsics.ll.bc b/test/Bitcode/neon-intrinsics.ll.bc
index c324aeef18..bc01eb540e 100644
--- a/test/Bitcode/neon-intrinsics.ll.bc
+++ b/test/Bitcode/neon-intrinsics.ll.bc
diff --git a/test/CodeGen/ARM/vadd.ll b/test/CodeGen/ARM/vadd.ll
index 9bb8bf5610..a830e968ff 100644
--- a/test/CodeGen/ARM/vadd.ll
+++ b/test/CodeGen/ARM/vadd.ll
@@ -157,8 +157,10 @@ define <8 x i16> @vaddls8(<8 x i8>* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vaddl.s8
 	%tmp1 = load <8 x i8>* %A
 	%tmp2 = load <8 x i8>* %B
-	%tmp3 = call <8 x i16> @llvm.arm.neon.vaddls.v8i16(<8 x i8> %tmp1, <8 x i8> %tmp2)
-	ret <8 x i16> %tmp3
+	%tmp3 = sext <8 x i8> %tmp1 to <8 x i16>
+	%tmp4 = sext <8 x i8> %tmp2 to <8 x i16>
+	%tmp5 = add <8 x i16> %tmp3, %tmp4
+	ret <8 x i16> %tmp5
 }
 
 define <4 x i32> @vaddls16(<4 x i16>* %A, <4 x i16>* %B) nounwind {
@@ -166,8 +168,10 @@ define <4 x i32> @vaddls16(<4 x i16>* %A, <4 x i16>* %B) nounwind {
 ;CHECK: vaddl.s16
 	%tmp1 = load <4 x i16>* %A
 	%tmp2 = load <4 x i16>* %B
-	%tmp3 = call <4 x i32> @llvm.arm.neon.vaddls.v4i32(<4 x i16> %tmp1, <4 x i16> %tmp2)
-	ret <4 x i32> %tmp3
+	%tmp3 = sext <4 x i16> %tmp1 to <4 x i32>
+	%tmp4 = sext <4 x i16> %tmp2 to <4 x i32>
+	%tmp5 = add <4 x i32> %tmp3, %tmp4
+	ret <4 x i32> %tmp5
 }
 
 define <2 x i64> @vaddls32(<2 x i32>* %A, <2 x i32>* %B) nounwind {
@@ -175,8 +179,10 @@ define <2 x i64> @vaddls32(<2 x i32>* %A, <2 x i32>* %B) nounwind {
 ;CHECK: vaddl.s32
 	%tmp1 = load <2 x i32>* %A
 	%tmp2 = load <2 x i32>* %B
-	%tmp3 = call <2 x i64> @llvm.arm.neon.vaddls.v2i64(<2 x i32> %tmp1, <2 x i32> %tmp2)
-	ret <2 x i64> %tmp3
+	%tmp3 = sext <2 x i32> %tmp1 to <2 x i64>
+	%tmp4 = sext <2 x i32> %tmp2 to <2 x i64>
+	%tmp5 = add <2 x i64> %tmp3, %tmp4
+	ret <2 x i64> %tmp5
 }
 
 define <8 x i16> @vaddlu8(<8 x i8>* %A, <8 x i8>* %B) nounwind {
@@ -184,8 +190,10 @@ define <8 x i16> @vaddlu8(<8 x i8>* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vaddl.u8
 	%tmp1 = load <8 x i8>* %A
 	%tmp2 = load <8 x i8>* %B
-	%tmp3 = call <8 x i16> @llvm.arm.neon.vaddlu.v8i16(<8 x i8> %tmp1, <8 x i8> %tmp2)
-	ret <8 x i16> %tmp3
+	%tmp3 = zext <8 x i8> %tmp1 to <8 x i16>
+	%tmp4 = zext <8 x i8> %tmp2 to <8 x i16>
+	%tmp5 = add <8 x i16> %tmp3, %tmp4
+	ret <8 x i16> %tmp5
 }
 
 define <4 x i32> @vaddlu16(<4 x i16>* %A, <4 x i16>* %B) nounwind {
@@ -193,8 +201,10 @@ define <4 x i32> @vaddlu16(<4 x i16>* %A, <4 x i16>* %B) nounwind {
 ;CHECK: vaddl.u16
 	%tmp1 = load <4 x i16>* %A
 	%tmp2 = load <4 x i16>* %B
-	%tmp3 = call <4 x i32> @llvm.arm.neon.vaddlu.v4i32(<4 x i16> %tmp1, <4 x i16> %tmp2)
-	ret <4 x i32> %tmp3
+	%tmp3 = zext <4 x i16> %tmp1 to <4 x i32>
+	%tmp4 = zext <4 x i16> %tmp2 to <4 x i32>
+	%tmp5 = add <4 x i32> %tmp3, %tmp4
+	ret <4 x i32> %tmp5
 }
 
 define <2 x i64> @vaddlu32(<2 x i32>* %A, <2 x i32>* %B) nounwind {
@@ -202,25 +212,20 @@ define <2 x i64> @vaddlu32(<2 x i32>* %A, <2 x i32>* %B) nounwind {
 ;CHECK: vaddl.u32
 	%tmp1 = load <2 x i32>* %A
 	%tmp2 = load <2 x i32>* %B
-	%tmp3 = call <2 x i64> @llvm.arm.neon.vaddlu.v2i64(<2 x i32> %tmp1, <2 x i32> %tmp2)
-	ret <2 x i64> %tmp3
+	%tmp3 = zext <2 x i32> %tmp1 to <2 x i64>
+	%tmp4 = zext <2 x i32> %tmp2 to <2 x i64>
+	%tmp5 = add <2 x i64> %tmp3, %tmp4
+	ret <2 x i64> %tmp5
 }
 
-declare <8 x i16> @llvm.arm.neon.vaddls.v8i16(<8 x i8>, <8 x i8>) nounwind readnone
-declare <4 x i32> @llvm.arm.neon.vaddls.v4i32(<4 x i16>, <4 x i16>) nounwind readnone
-declare <2 x i64> @llvm.arm.neon.vaddls.v2i64(<2 x i32>, <2 x i32>) nounwind readnone
-
-declare <8 x i16> @llvm.arm.neon.vaddlu.v8i16(<8 x i8>, <8 x i8>) nounwind readnone
-declare <4 x i32> @llvm.arm.neon.vaddlu.v4i32(<4 x i16>, <4 x i16>) nounwind readnone
-declare <2 x i64> @llvm.arm.neon.vaddlu.v2i64(<2 x i32>, <2 x i32>) nounwind readnone
-
 define <8 x i16> @vaddws8(<8 x i16>* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vaddws8:
 ;CHECK: vaddw.s8
 	%tmp1 = load <8 x i16>* %A
 	%tmp2 = load <8 x i8>* %B
-	%tmp3 = call <8 x i16> @llvm.arm.neon.vaddws.v8i16(<8 x i16> %tmp1, <8 x i8> %tmp2)
-	ret <8 x i16> %tmp3
+	%tmp3 = sext <8 x i8> %tmp2 to <8 x i16>
+	%tmp4 = add <8 x i16> %tmp1, %tmp3
+	ret <8 x i16> %tmp4
 }
 
 define <4 x i32> @vaddws16(<4 x i32>* %A, <4 x i16>* %B) nounwind {
@@ -228,8 +233,9 @@ define <4 x i32> @vaddws16(<4 x i32>* %A, <4 x i16>* %B) nounwind {
 ;CHECK: vaddw.s16
 	%tmp1 = load <4 x i32>* %A
 	%tmp2 = load <4 x i16>* %B
-	%tmp3 = call <4 x i32> @llvm.arm.neon.vaddws.v4i32(<4 x i32> %tmp1, <4 x i16> %tmp2)
-	ret <4 x i32> %tmp3
+	%tmp3 = sext <4 x i16> %tmp2 to <4 x i32>
+	%tmp4 = add <4 x i32> %tmp1, %tmp3
+	ret <4 x i32> %tmp4
 }
 
 define <2 x i64> @vaddws32(<2 x i64>* %A, <2 x i32>* %B) nounwind {
@@ -237,8 +243,9 @@ define <2 x i64> @vaddws32(<2 x i64>* %A, <2 x i32>* %B) nounwind {
 ;CHECK: vaddw.s32
 	%tmp1 = load <2 x i64>* %A
 	%tmp2 = load <2 x i32>* %B
-	%tmp3 = call <2 x i64> @llvm.arm.neon.vaddws.v2i64(<2 x i64> %tmp1, <2 x i32> %tmp2)
-	ret <2 x i64> %tmp3
+	%tmp3 = sext <2 x i32> %tmp2 to <2 x i64>
+	%tmp4 = add <2 x i64> %tmp1, %tmp3
+	ret <2 x i64> %tmp4
 }
 
 define <8 x i16> @vaddwu8(<8 x i16>* %A, <8 x i8>* %B) nounwind {
@@ -246,8 +253,9 @@ define <8 x i16> @vaddwu8(<8 x i16>* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vaddw.u8
 	%tmp1 = load <8 x i16>* %A
 	%tmp2 = load <8 x i8>* %B
-	%tmp3 = call <8 x i16> @llvm.arm.neon.vaddwu.v8i16(<8 x i16> %tmp1, <8 x i8> %tmp2)
-	ret <8 x i16> %tmp3
+	%tmp3 = zext <8 x i8> %tmp2 to <8 x i16>
+	%tmp4 = add <8 x i16> %tmp1, %tmp3
+	ret <8 x i16> %tmp4
 }
 
 define <4 x i32> @vaddwu16(<4 x i32>* %A, <4 x i16>* %B) nounwind {
@@ -255,8 +263,9 @@ define <4 x i32> @vaddwu16(<4 x i32>* %A, <4 x i16>* %B) nounwind {
 ;CHECK: vaddw.u16
 	%tmp1 = load <4 x i32>* %A
 	%tmp2 = load <4 x i16>* %B
-	%tmp3 = call <4 x i32> @llvm.arm.neon.vaddwu.v4i32(<4 x i32> %tmp1, <4 x i16> %tmp2)
-	ret <4 x i32> %tmp3
+	%tmp3 = zext <4 x i16> %tmp2 to <4 x i32>
+	%tmp4 = add <4 x i32> %tmp1, %tmp3
+	ret <4 x i32> %tmp4
 }
 
 define <2 x i64> @vaddwu32(<2 x i64>* %A, <2 x i32>* %B) nounwind {
@@ -264,14 +273,7 @@ define <2 x i64> @vaddwu32(<2 x i64>* %A, <2 x i32>* %B) nounwind {
 ;CHECK: vaddw.u32
 	%tmp1 = load <2 x i64>* %A
 	%tmp2 = load <2 x i32>* %B
-	%tmp3 = call <2 x i64> @llvm.arm.neon.vaddwu.v2i64(<2 x i64> %tmp1, <2 x i32> %tmp2)
-	ret <2 x i64> %tmp3
+	%tmp3 = zext <2 x i32> %tmp2 to <2 x i64>
+	%tmp4 = add <2 x i64> %tmp1, %tmp3
+	ret <2 x i64> %tmp4
 }
-
-declare <8 x i16> @llvm.arm.neon.vaddws.v8i16(<8 x i16>, <8 x i8>) nounwind readnone
-declare <4 x i32> @llvm.arm.neon.vaddws.v4i32(<4 x i32>, <4 x i16>) nounwind readnone
-declare <2 x i64> @llvm.arm.neon.vaddws.v2i64(<2 x i64>, <2 x i32>) nounwind readnone
-
-declare <8 x i16> @llvm.arm.neon.vaddwu.v8i16(<8 x i16>, <8 x i8>) nounwind readnone
-declare <4 x i32> @llvm.arm.neon.vaddwu.v4i32(<4 x i32>, <4 x i16>) nounwind readnone
-declare <2 x i64> @llvm.arm.neon.vaddwu.v2i64(<2 x i64>, <2 x i32>) nounwind readnone
diff --git a/test/CodeGen/ARM/vsub.ll b/test/CodeGen/ARM/vsub.ll
index 3416de76f1..df77bb31fc 100644
--- a/test/CodeGen/ARM/vsub.ll
+++ b/test/CodeGen/ARM/vsub.ll
@@ -157,8 +157,10 @@ define <8 x i16> @vsubls8(<8 x i8>* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vsubl.s8
 	%tmp1 = load <8 x i8>* %A
 	%tmp2 = load <8 x i8>* %B
-	%tmp3 = call <8 x i16> @llvm.arm.neon.vsubls.v8i16(<8 x i8> %tmp1, <8 x i8> %tmp2)
-	ret <8 x i16> %tmp3
+	%tmp3 = sext <8 x i8> %tmp1 to <8 x i16>
+	%tmp4 = sext <8 x i8> %tmp2 to <8 x i16>
+	%tmp5 = sub <8 x i16> %tmp3, %tmp4
+	ret <8 x i16> %tmp5
 }
 
 define <4 x i32> @vsubls16(<4 x i16>* %A, <4 x i16>* %B) nounwind {
@@ -166,8 +168,10 @@ define <4 x i32> @vsubls16(<4 x i16>* %A, <4 x i16>* %B) nounwind {
 ;CHECK: vsubl.s16
 	%tmp1 = load <4 x i16>* %A
 	%tmp2 = load <4 x i16>* %B
-	%tmp3 = call <4 x i32> @llvm.arm.neon.vsubls.v4i32(<4 x i16> %tmp1, <4 x i16> %tmp2)
-	ret <4 x i32> %tmp3
+	%tmp3 = sext <4 x i16> %tmp1 to <4 x i32>
+	%tmp4 = sext <4 x i16> %tmp2 to <4 x i32>
+	%tmp5 = sub <4 x i32> %tmp3, %tmp4
+	ret <4 x i32> %tmp5
 }
 
 define <2 x i64> @vsubls32(<2 x i32>* %A, <2 x i32>* %B) nounwind {
@@ -175,8 +179,10 @@ define <2 x i64> @vsubls32(<2 x i32>* %A, <2 x i32>* %B) nounwind {
 ;CHECK: vsubl.s32
 	%tmp1 = load <2 x i32>* %A
 	%tmp2 = load <2 x i32>* %B
-	%tmp3 = call <2 x i64> @llvm.arm.neon.vsubls.v2i64(<2 x i32> %tmp1, <2 x i32> %tmp2)
-	ret <2 x i64> %tmp3
+	%tmp3 = sext <2 x i32> %tmp1 to <2 x i64>
+	%tmp4 = sext <2 x i32> %tmp2 to <2 x i64>
+	%tmp5 = sub <2 x i64> %tmp3, %tmp4
+	ret <2 x i64> %tmp5
 }
 
 define <8 x i16> @vsublu8(<8 x i8>* %A, <8 x i8>* %B) nounwind {
@@ -184,8 +190,10 @@ define <8 x i16> @vsublu8(<8 x i8>* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vsubl.u8
 	%tmp1 = load <8 x i8>* %A
 	%tmp2 = load <8 x i8>* %B
-	%tmp3 = call <8 x i16> @llvm.arm.neon.vsublu.v8i16(<8 x i8> %tmp1, <8 x i8> %tmp2)
-	ret <8 x i16> %tmp3
+	%tmp3 = zext <8 x i8> %tmp1 to <8 x i16>
+	%tmp4 = zext <8 x i8> %tmp2 to <8 x i16>
+	%tmp5 = sub <8 x i16> %tmp3, %tmp4
+	ret <8 x i16> %tmp5
 }
 
 define <4 x i32> @vsublu16(<4 x i16>* %A, <4 x i16>* %B) nounwind {
@@ -193,8 +201,10 @@ define <4 x i32> @vsublu16(<4 x i16>* %A, <4 x i16>* %B) nounwind {
 ;CHECK: vsubl.u16
 	%tmp1 = load <4 x i16>* %A
 	%tmp2 = load <4 x i16>* %B
-	%tmp3 = call <4 x i32> @llvm.arm.neon.vsublu.v4i32(<4 x i16> %tmp1, <4 x i16> %tmp2)
-	ret <4 x i32> %tmp3
+	%tmp3 = zext <4 x i16> %tmp1 to <4 x i32>
+	%tmp4 = zext <4 x i16> %tmp2 to <4 x i32>
+	%tmp5 = sub <4 x i32> %tmp3, %tmp4
+	ret <4 x i32> %tmp5
 }
 
 define <2 x i64> @vsublu32(<2 x i32>* %A, <2 x i32>* %B) nounwind {
@@ -202,25 +212,20 @@ define <2 x i64> @vsublu32(<2 x i32>* %A, <2 x i32>* %B) nounwind {
 ;CHECK: vsubl.u32
 	%tmp1 = load <2 x i32>* %A
 	%tmp2 = load <2 x i32>* %B
-	%tmp3 = call <2 x i64> @llvm.arm.neon.vsublu.v2i64(<2 x i32> %tmp1, <2 x i32> %tmp2)
-	ret <2 x i64> %tmp3
+	%tmp3 = zext <2 x i32> %tmp1 to <2 x i64>
+	%tmp4 = zext <2 x i32> %tmp2 to <2 x i64>
+	%tmp5 = sub <2 x i64> %tmp3, %tmp4
+	ret <2 x i64> %tmp5
 }
 
-declare <8 x i16> @llvm.arm.neon.vsubls.v8i16(<8 x i8>, <8 x i8>) nounwind readnone
-declare <4 x i32> @llvm.arm.neon.vsubls.v4i32(<4 x i16>, <4 x i16>) nounwind readnone
-declare <2 x i64> @llvm.arm.neon.vsubls.v2i64(<2 x i32>, <2 x i32>) nounwind readnone
-
-declare <8 x i16> @llvm.arm.neon.vsublu.v8i16(<8 x i8>, <8 x i8>) nounwind readnone
-declare <4 x i32> @llvm.arm.neon.vsublu.v4i32(<4 x i16>, <4 x i16>) nounwind readnone
-declare <2 x i64> @llvm.arm.neon.vsublu.v2i64(<2 x i32>, <2 x i32>) nounwind readnone
-
 define <8 x i16> @vsubws8(<8 x i16>* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vsubws8:
 ;CHECK: vsubw.s8
 	%tmp1 = load <8 x i16>* %A
 	%tmp2 = load <8 x i8>* %B
-	%tmp3 = call <8 x i16> @llvm.arm.neon.vsubws.v8i16(<8 x i16> %tmp1, <8 x i8> %tmp2)
-	ret <8 x i16> %tmp3
+	%tmp3 = sext <8 x i8> %tmp2 to <8 x i16>
+	%tmp4 = sub <8 x i16> %tmp1, %tmp3
+	ret <8 x i16> %tmp4
 }
 
 define <4 x i32> @vsubws16(<4 x i32>* %A, <4 x i16>* %B) nounwind {
@@ -228,8 +233,9 @@ define <4 x i32> @vsubws16(<4 x i32>* %A, <4 x i16>* %B) nounwind {
 ;CHECK: vsubw.s16
 	%tmp1 = load <4 x i32>* %A
 	%tmp2 = load <4 x i16>* %B
-	%tmp3 = call <4 x i32> @llvm.arm.neon.vsubws.v4i32(<4 x i32> %tmp1, <4 x i16> %tmp2)
-	ret <4 x i32> %tmp3
+	%tmp3 = sext <4 x i16> %tmp2 to <4 x i32>
+	%tmp4 = sub <4 x i32> %tmp1, %tmp3
+	ret <4 x i32> %tmp4
 }
 
 define <2 x i64> @vsubws32(<2 x i64>* %A, <2 x i32>* %B) nounwind {
@@ -237,8 +243,9 @@ define <2 x i64> @vsubws32(<2 x i64>* %A, <2 x i32>* %B) nounwind {
 ;CHECK: vsubw.s32
 	%tmp1 = load <2 x i64>* %A
 	%tmp2 = load <2 x i32>* %B
-	%tmp3 = call <2 x i64> @llvm.arm.neon.vsubws.v2i64(<2 x i64> %tmp1, <2 x i32> %tmp2)
-	ret <2 x i64> %tmp3
+	%tmp3 = sext <2 x i32> %tmp2 to <2 x i64>
+	%tmp4 = sub <2 x i64> %tmp1, %tmp3
+	ret <2 x i64> %tmp4
 }
 
 define <8 x i16> @vsubwu8(<8 x i16>* %A, <8 x i8>* %B) nounwind {
@@ -246,8 +253,9 @@ define <8 x i16> @vsubwu8(<8 x i16>* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vsubw.u8
 	%tmp1 = load <8 x i16>* %A
 	%tmp2 = load <8 x i8>* %B
-	%tmp3 = call <8 x i16> @llvm.arm.neon.vsubwu.v8i16(<8 x i16> %tmp1, <8 x i8> %tmp2)
-	ret <8 x i16> %tmp3
+	%tmp3 = zext <8 x i8> %tmp2 to <8 x i16>
+	%tmp4 = sub <8 x i16> %tmp1, %tmp3
+	ret <8 x i16> %tmp4
 }
 
 define <4 x i32> @vsubwu16(<4 x i32>* %A, <4 x i16>* %B) nounwind {
@@ -255,8 +263,9 @@ define <4 x i32> @vsubwu16(<4 x i32>* %A, <4 x i16>* %B) nounwind {
 ;CHECK: vsubw.u16
 	%tmp1 = load <4 x i32>* %A
 	%tmp2 = load <4 x i16>* %B
-	%tmp3 = call <4 x i32> @llvm.arm.neon.vsubwu.v4i32(<4 x i32> %tmp1, <4 x i16> %tmp2)
-	ret <4 x i32> %tmp3
+	%tmp3 = zext <4 x i16> %tmp2 to <4 x i32>
+	%tmp4 = sub <4 x i32> %tmp1, %tmp3
+	ret <4 x i32> %tmp4
 }
 
 define <2 x i64> @vsubwu32(<2 x i64>* %A, <2 x i32>* %B) nounwind {
@@ -264,14 +273,7 @@ define <2 x i64> @vsubwu32(<2 x i64>* %A, <2 x i32>* %B) nounwind {
 ;CHECK: vsubw.u32
 	%tmp1 = load <2 x i64>* %A
 	%tmp2 = load <2 x i32>* %B
-	%tmp3 = call <2 x i64> @llvm.arm.neon.vsubwu.v2i64(<2 x i64> %tmp1, <2 x i32> %tmp2)
-	ret <2 x i64> %tmp3
+	%tmp3 = zext <2 x i32> %tmp2 to <2 x i64>
+	%tmp4 = sub <2 x i64> %tmp1, %tmp3
+	ret <2 x i64> %tmp4
 }
-
-declare <8 x i16> @llvm.arm.neon.vsubws.v8i16(<8 x i16>, <8 x i8>) nounwind readnone
-declare <4 x i32> @llvm.arm.neon.vsubws.v4i32(<4 x i32>, <4 x i16>) nounwind readnone
-declare <2 x i64> @llvm.arm.neon.vsubws.v2i64(<2 x i64>, <2 x i32>) nounwind readnone
-
-declare <8 x i16> @llvm.arm.neon.vsubwu.v8i16(<8 x i16>, <8 x i8>) nounwind readnone
-declare <4 x i32> @llvm.arm.neon.vsubwu.v4i32(<4 x i32>, <4 x i16>) nounwind readnone
-declare <2 x i64> @llvm.arm.neon.vsubwu.v2i64(<2 x i64>, <2 x i32>) nounwind readnone