Replace the arm.neon.vmovls and vmovlu intrinsics with vector sign-extend and

zero-extend operations.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@111614 91177308-0d34-0410-b5e6-96231b3b80d8

8 files changed, 86 insertions, 46 deletions

diff --git a/include/llvm/IntrinsicsARM.td b/include/llvm/IntrinsicsARM.td
index f4a80bb9cb..37d813151c 100644
--- a/include/llvm/IntrinsicsARM.td
+++ b/include/llvm/IntrinsicsARM.td
@@ -60,9 +60,6 @@ let TargetPrefix = "arm" in {  // All intrinsics start with "llvm.arm.".
   class Neon_1Arg_Narrow_Intrinsic
     : Intrinsic<[llvm_anyvector_ty],
                 [LLVMExtendedElementVectorType<0>], [IntrNoMem]>;
-  class Neon_1Arg_Long_Intrinsic
-    : Intrinsic<[llvm_anyvector_ty],
-                [LLVMTruncatedElementVectorType<0>], [IntrNoMem]>;
   class Neon_2Arg_Intrinsic
     : Intrinsic<[llvm_anyvector_ty], [LLVMMatchType<0>, LLVMMatchType<0>],
                 [IntrNoMem]>;
@@ -322,8 +319,6 @@ def int_arm_neon_vmovn : Neon_1Arg_Narrow_Intrinsic;
 def int_arm_neon_vqmovns : Neon_1Arg_Narrow_Intrinsic;
 def int_arm_neon_vqmovnu : Neon_1Arg_Narrow_Intrinsic;
 def int_arm_neon_vqmovnsu : Neon_1Arg_Narrow_Intrinsic;
-def int_arm_neon_vmovls : Neon_1Arg_Long_Intrinsic;
-def int_arm_neon_vmovlu : Neon_1Arg_Long_Intrinsic;
 
 // Vector Table Lookup.
 // The first 1-4 arguments are the table.
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 9e9198b726..034114c680 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -125,12 +125,14 @@ void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
   setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
   setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
   setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
-  setOperationAction(ISD::ZERO_EXTEND, VT.getSimpleVT(), Expand);
   if (VT.isInteger()) {
     setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
     setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
     setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
+    setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
+    setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
   }
+  setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
 
   // Promote all bit-wise operations.
   if (VT.isInteger() && VT != PromotedBitwiseVT) {
@@ -320,6 +322,8 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
     setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
     setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
 
+    setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
+
     // Neon does not support some operations on v1i64 and v2i64 types.
     setOperationAction(ISD::MUL, MVT::v1i64, Expand);
     setOperationAction(ISD::MUL, MVT::v2i64, Expand);
@@ -3786,7 +3790,7 @@ SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
   case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
   case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
-  case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
+  case ISD::FLT_ROUNDS_:   return LowerFLT_ROUNDS_(Op, DAG);
   }
   return SDValue();
 }
diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 595a2fc795..9c01fcd184 100644
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -888,14 +888,14 @@ class N2VNInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
         (ins QPR:$src), itin, OpcodeStr, Dt, "$dst, $src", "",
         [(set DPR:$dst, (TyD (IntOp (TyQ QPR:$src))))]>;
 
-// Long 2-register intrinsics (currently only used for VMOVL).
-class N2VLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
-              bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
-              InstrItinClass itin, string OpcodeStr, string Dt,
-              ValueType TyQ, ValueType TyD, Intrinsic IntOp>
+// Long 2-register operations (currently only used for VMOVL).
+class N2VL<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
+           bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
+           InstrItinClass itin, string OpcodeStr, string Dt,
+           ValueType TyQ, ValueType TyD, SDNode OpNode>
   : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, op6, op4, (outs QPR:$dst),
         (ins DPR:$src), itin, OpcodeStr, Dt, "$dst, $src", "",
-        [(set QPR:$dst, (TyQ (IntOp (TyD DPR:$src))))]>;
+        [(set QPR:$dst, (TyQ (OpNode (TyD DPR:$src))))]>;
 
 // 2-register shuffles (VTRN/VZIP/VUZP), both double- and quad-register.
 class N2VDShuffle<bits<2> op19_18, bits<5> op11_7, string OpcodeStr, string Dt>
@@ -1508,14 +1508,14 @@ multiclass N2VNInt_HSD<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
 
 // Neon Lengthening 2-register vector intrinsic (currently specific to VMOVL).
 //   source operand element sizes of 16, 32 and 64 bits:
-multiclass N2VLInt_QHS<bits<2> op24_23, bits<5> op11_7, bit op6, bit op4,
-                       string OpcodeStr, string Dt, Intrinsic IntOp> {
-  def v8i16 : N2VLInt<op24_23, 0b00, 0b10, 0b00, op11_7, op6, op4, IIC_VQUNAiD,
-                      OpcodeStr, !strconcat(Dt, "8"), v8i16, v8i8, IntOp>;
-  def v4i32 : N2VLInt<op24_23, 0b01, 0b00, 0b00, op11_7, op6, op4, IIC_VQUNAiD,
-                      OpcodeStr, !strconcat(Dt, "16"), v4i32, v4i16, IntOp>;
-  def v2i64 : N2VLInt<op24_23, 0b10, 0b00, 0b00, op11_7, op6, op4, IIC_VQUNAiD,
-                      OpcodeStr, !strconcat(Dt, "32"), v2i64, v2i32, IntOp>;
+multiclass N2VL_QHS<bits<2> op24_23, bits<5> op11_7, bit op6, bit op4,
+                    string OpcodeStr, string Dt, SDNode OpNode> {
+  def v8i16 : N2VL<op24_23, 0b00, 0b10, 0b00, op11_7, op6, op4, IIC_VQUNAiD,
+                   OpcodeStr, !strconcat(Dt, "8"), v8i16, v8i8, OpNode>;
+  def v4i32 : N2VL<op24_23, 0b01, 0b00, 0b00, op11_7, op6, op4, IIC_VQUNAiD,
+                   OpcodeStr, !strconcat(Dt, "16"), v4i32, v4i16, OpNode>;
+  def v2i64 : N2VL<op24_23, 0b10, 0b00, 0b00, op11_7, op6, op4, IIC_VQUNAiD,
+                   OpcodeStr, !strconcat(Dt, "32"), v2i64, v2i32, OpNode>;
 }
 
 
@@ -3123,10 +3123,8 @@ defm VQMOVNu  : N2VNInt_HSD<0b11,0b11,0b10,0b00101,1,0, IIC_VQUNAiD,
 defm VQMOVNsu : N2VNInt_HSD<0b11,0b11,0b10,0b00100,1,0, IIC_VQUNAiD,
                             "vqmovun", "s", int_arm_neon_vqmovnsu>;
 //   VMOVL    : Vector Lengthening Move
-defm VMOVLs   : N2VLInt_QHS<0b01,0b10100,0,1, "vmovl", "s",
-                            int_arm_neon_vmovls>;
-defm VMOVLu   : N2VLInt_QHS<0b11,0b10100,0,1, "vmovl", "u",
-                            int_arm_neon_vmovlu>;
+defm VMOVLs   : N2VL_QHS<0b01,0b10100,0,1, "vmovl", "s", sext>;
+defm VMOVLu   : N2VL_QHS<0b11,0b10100,0,1, "vmovl", "u", zext>;
 
 // Vector Conversions.
 
diff --git a/lib/VMCore/AutoUpgrade.cpp b/lib/VMCore/AutoUpgrade.cpp
index 4f4daeba4d..f76d0d254d 100644
--- a/lib/VMCore/AutoUpgrade.cpp
+++ b/lib/VMCore/AutoUpgrade.cpp
@@ -78,6 +78,13 @@ static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
         NewFn = F;
         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) {
+        // Calls to these are transformed into IR without intrinsics.
+        NewFn = 0;
+        return true;
+      }
     }
     break;
   case 'b':
@@ -320,6 +327,28 @@ void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
   assert(F && "CallInst has no function associated with it.");
 
   if (!NewFn) {
+    // Get the Function's name.
+    const std::string& Name = F->getName();
+
+    // Upgrade ARM NEON intrinsics.
+    if (Name.compare(5, 9, "arm.neon.", 9) == 0) {
+      Instruction *NewI;
+      if (Name.compare(14, 7, "vmovls.", 7) == 0) {
+        NewI = new SExtInst(CI->getArgOperand(0), CI->getType(),
+                            "upgraded." + CI->getName(), CI);
+      } else if (Name.compare(14, 7, "vmovlu.", 7) == 0) {
+        NewI = new ZExtInst(CI->getArgOperand(0), CI->getType(),
+                            "upgraded." + CI->getName(), CI);
+      } else {
+        llvm_unreachable("Unknown arm.neon function for CallInst upgrade.");
+      }
+      // Replace any uses of the old CallInst.
+      if (!CI->use_empty())
+        CI->replaceAllUsesWith(NewI);
+      CI->eraseFromParent();
+      return;
+    }
+
     bool isLoadH = false, isLoadL = false, isMovL = false;
     bool isMovSD = false, isShufPD = false;
     bool isUnpckhPD = false, isUnpcklPD = false;
diff --git a/test/Bitcode/neon-intrinsics.ll b/test/Bitcode/neon-intrinsics.ll
new file mode 100644
index 0000000000..73ca7075d8
--- /dev/null
+++ b/test/Bitcode/neon-intrinsics.ll
@@ -0,0 +1,29 @@
+; RUN: llvm-dis < %s.bc | FileCheck %s
+
+; vmovls should be auto-upgraded to sext
+
+; CHECK: vmovls8
+; CHECK-NOT: arm.neon.vmovls.v8i16
+; CHECK: sext <8 x i8>
+
+; CHECK: vmovls16
+; CHECK-NOT: arm.neon.vmovls.v4i32
+; CHECK: sext <4 x i16>
+
+; CHECK: vmovls32
+; CHECK-NOT: arm.neon.vmovls.v2i64
+; CHECK: sext <2 x i32>
+
+; vmovlu should be auto-upgraded to zext
+
+; CHECK: vmovlu8
+; CHECK-NOT: arm.neon.vmovlu.v8i16
+; CHECK: zext <8 x i8>
+
+; CHECK: vmovlu16
+; CHECK-NOT: arm.neon.vmovlu.v4i32
+; CHECK: zext <4 x i16>
+
+; CHECK: vmovlu32
+; CHECK-NOT: arm.neon.vmovlu.v2i64
+; CHECK: zext <2 x i32>
diff --git a/test/Bitcode/neon-intrinsics.ll.bc b/test/Bitcode/neon-intrinsics.ll.bc
new file mode 100644
index 0000000000..93eeabc822
--- /dev/null
+++ b/test/Bitcode/neon-intrinsics.ll.bc
diff --git a/test/CodeGen/ARM/neon-ops.ll b/test/CodeGen/ARM/neon-ops.ll
deleted file mode 100644
index d7e893f64f..0000000000
--- a/test/CodeGen/ARM/neon-ops.ll
+++ /dev/null
@@ -1,7 +0,0 @@
-; RUN: llc -march=arm -mattr=+neon -O2 -o /dev/null
-
-; This used to crash.
-define <4 x i32> @test1(<4 x i16> %a) {
-  %A = zext <4 x i16> %a to <4 x i32>
-  ret <4 x i32> %A
-}
diff --git a/test/CodeGen/ARM/vmov.ll b/test/CodeGen/ARM/vmov.ll
index 5e872ab6d0..f863bf8a09 100644
--- a/test/CodeGen/ARM/vmov.ll
+++ b/test/CodeGen/ARM/vmov.ll
@@ -192,7 +192,7 @@ define <8 x i16> @vmovls8(<8 x i8>* %A) nounwind {
 ;CHECK: vmovls8:
 ;CHECK: vmovl.s8
 	%tmp1 = load <8 x i8>* %A
-	%tmp2 = call <8 x i16> @llvm.arm.neon.vmovls.v8i16(<8 x i8> %tmp1)
+	%tmp2 = sext <8 x i8> %tmp1 to <8 x i16>
 	ret <8 x i16> %tmp2
 }
 
@@ -200,7 +200,7 @@ define <4 x i32> @vmovls16(<4 x i16>* %A) nounwind {
 ;CHECK: vmovls16:
 ;CHECK: vmovl.s16
 	%tmp1 = load <4 x i16>* %A
-	%tmp2 = call <4 x i32> @llvm.arm.neon.vmovls.v4i32(<4 x i16> %tmp1)
+	%tmp2 = sext <4 x i16> %tmp1 to <4 x i32>
 	ret <4 x i32> %tmp2
 }
 
@@ -208,7 +208,7 @@ define <2 x i64> @vmovls32(<2 x i32>* %A) nounwind {
 ;CHECK: vmovls32:
 ;CHECK: vmovl.s32
 	%tmp1 = load <2 x i32>* %A
-	%tmp2 = call <2 x i64> @llvm.arm.neon.vmovls.v2i64(<2 x i32> %tmp1)
+	%tmp2 = sext <2 x i32> %tmp1 to <2 x i64>
 	ret <2 x i64> %tmp2
 }
 
@@ -216,7 +216,7 @@ define <8 x i16> @vmovlu8(<8 x i8>* %A) nounwind {
 ;CHECK: vmovlu8:
 ;CHECK: vmovl.u8
 	%tmp1 = load <8 x i8>* %A
-	%tmp2 = call <8 x i16> @llvm.arm.neon.vmovlu.v8i16(<8 x i8> %tmp1)
+	%tmp2 = zext <8 x i8> %tmp1 to <8 x i16>
 	ret <8 x i16> %tmp2
 }
 
@@ -224,7 +224,7 @@ define <4 x i32> @vmovlu16(<4 x i16>* %A) nounwind {
 ;CHECK: vmovlu16:
 ;CHECK: vmovl.u16
 	%tmp1 = load <4 x i16>* %A
-	%tmp2 = call <4 x i32> @llvm.arm.neon.vmovlu.v4i32(<4 x i16> %tmp1)
+	%tmp2 = zext <4 x i16> %tmp1 to <4 x i32>
 	ret <4 x i32> %tmp2
 }
 
@@ -232,18 +232,10 @@ define <2 x i64> @vmovlu32(<2 x i32>* %A) nounwind {
 ;CHECK: vmovlu32:
 ;CHECK: vmovl.u32
 	%tmp1 = load <2 x i32>* %A
-	%tmp2 = call <2 x i64> @llvm.arm.neon.vmovlu.v2i64(<2 x i32> %tmp1)
+	%tmp2 = zext <2 x i32> %tmp1 to <2 x i64>
 	ret <2 x i64> %tmp2
 }
 
-declare <8 x i16> @llvm.arm.neon.vmovls.v8i16(<8 x i8>) nounwind readnone
-declare <4 x i32> @llvm.arm.neon.vmovls.v4i32(<4 x i16>) nounwind readnone
-declare <2 x i64> @llvm.arm.neon.vmovls.v2i64(<2 x i32>) nounwind readnone
-
-declare <8 x i16> @llvm.arm.neon.vmovlu.v8i16(<8 x i8>) nounwind readnone
-declare <4 x i32> @llvm.arm.neon.vmovlu.v4i32(<4 x i16>) nounwind readnone
-declare <2 x i64> @llvm.arm.neon.vmovlu.v2i64(<2 x i32>) nounwind readnone
-
 define <8 x i8> @vmovni16(<8 x i16>* %A) nounwind {
 ;CHECK: vmovni16:
 ;CHECK: vmovn.i16