Add support for NEON VMVN immediate instructions.

llvm-svn: 108324

4 files changed, 96 insertions, 7 deletions

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 9e9192bc8a0..0091df753eb 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -625,6 +625,7 @@ const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case ARMISD::VGETLANEu:     return "ARMISD::VGETLANEu";
   case ARMISD::VGETLANEs:     return "ARMISD::VGETLANEs";
   case ARMISD::VMOVIMM:       return "ARMISD::VMOVIMM";
+  case ARMISD::VMVNIMM:       return "ARMISD::VMVNIMM";
   case ARMISD::VDUP:          return "ARMISD::VDUP";
   case ARMISD::VDUPLANE:      return "ARMISD::VDUPLANE";
   case ARMISD::VEXT:          return "ARMISD::VEXT";
@@ -2925,6 +2926,8 @@ static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
 
   switch (SplatBitSize) {
   case 8:
+    if (!isVMOV)
+      return SDValue();
     // Any 1-byte value is OK.  Op=0, Cmode=1110.
     assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
     OpCmode = 0xe;
@@ -3006,9 +3009,9 @@ static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
     return SDValue();
 
   case 64: {
-    // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
     if (!isVMOV)
       return SDValue();
+    // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
     uint64_t BitMask = 0xff;
     uint64_t Val = 0;
     unsigned ImmMask = 1;
@@ -3248,6 +3251,17 @@ static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
         SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
         return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
       }
+
+      // Try an immediate VMVN.
+      uint64_t NegatedImm = (SplatBits.getZExtValue() ^
+                             ((1LL << SplatBitSize) - 1));
+      Val = isNEONModifiedImm(NegatedImm,
+                                      SplatUndef.getZExtValue(), SplatBitSize,
+                                      DAG, VmovVT, VT.is128BitVector(), false);
+      if (Val.getNode()) {
+        SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
+        return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
+      }
     }
   }
 
@@ -4232,14 +4246,15 @@ static SDValue PerformVMOVRRDCombine(SDNode *N,
 /// ARMISD::VDUPLANE.
 static SDValue PerformVDUPLANECombine(SDNode *N,
                                       TargetLowering::DAGCombinerInfo &DCI) {
-  // If the source is already a VMOVIMM splat, the VDUPLANE is redundant.
+  // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
+  // redundant.
   SDValue Op = N->getOperand(0);
   EVT VT = N->getValueType(0);
 
   // Ignore bit_converts.
   while (Op.getOpcode() == ISD::BIT_CONVERT)
     Op = Op.getOperand(0);
-  if (Op.getOpcode() != ARMISD::VMOVIMM)
+  if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
     return SDValue();
 
   // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
diff --git a/llvm/lib/Target/ARM/ARMISelLowering.h b/llvm/lib/Target/ARM/ARMISelLowering.h
index 81120c8d994..128b72e1e74 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.h
+++ b/llvm/lib/Target/ARM/ARMISelLowering.h
@@ -124,8 +124,11 @@ namespace llvm {
       VGETLANEu,    // zero-extend vector extract element
       VGETLANEs,    // sign-extend vector extract element
 
-      // Vector duplicate:
+      // Vector move immediate and move negated immediate:
       VMOVIMM,
+      VMVNIMM,
+
+      // Vector duplicate:
       VDUP,
       VDUPLANE,
 
diff --git a/llvm/lib/Target/ARM/ARMInstrNEON.td b/llvm/lib/Target/ARM/ARMInstrNEON.td
index bde806d5248..7f7eb980abe 100644
--- a/llvm/lib/Target/ARM/ARMInstrNEON.td
+++ b/llvm/lib/Target/ARM/ARMInstrNEON.td
@@ -65,9 +65,10 @@ def SDTARMVGETLN  : SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisInt<1>,
 def NEONvgetlaneu : SDNode<"ARMISD::VGETLANEu", SDTARMVGETLN>;
 def NEONvgetlanes : SDNode<"ARMISD::VGETLANEs", SDTARMVGETLN>;
 
-def NEONvmovImm   : SDNode<"ARMISD::VMOVIMM",
-                           SDTypeProfile<1, 1, [SDTCisVec<0>,
-                                                SDTCisVT<1, i32>]>>;
+def SDTARMVMOVIMM : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVT<1, i32>]>;
+def NEONvmovImm   : SDNode<"ARMISD::VMOVIMM", SDTARMVMOVIMM>;
+def NEONvmvnImm   : SDNode<"ARMISD::VMVNIMM", SDTARMVMOVIMM>;
+
 def NEONvdup      : SDNode<"ARMISD::VDUP", SDTypeProfile<1, 1, [SDTCisVec<0>]>>;
 
 // VDUPLANE can produce a quad-register result from a double-register source,
@@ -2383,6 +2384,28 @@ def  VORNq    : N3VX<0, 0, 0b11, 0b0001, 1, 1, (outs QPR:$dst),
                      [(set QPR:$dst, (v4i32 (or QPR:$src1,
                                                 (vnotq QPR:$src2))))]>;
 
+//   VMVN     : Vector Bitwise NOT (Immediate)
+
+let isReMaterializable = 1 in {
+def VMVNv4i16 : N1ModImm<1, 0b000, {1,0,?,0}, 0, 0, 1, 1, (outs DPR:$dst),
+                         (ins nModImm:$SIMM), IIC_VMOVImm,
+                         "vmvn", "i16", "$dst, $SIMM", "",
+                         [(set DPR:$dst, (v4i16 (NEONvmvnImm timm:$SIMM)))]>;
+def VMVNv8i16 : N1ModImm<1, 0b000, {1,0,?,0}, 0, 1, 1, 1, (outs QPR:$dst),
+                         (ins nModImm:$SIMM), IIC_VMOVImm,
+                         "vmvn", "i16", "$dst, $SIMM", "",
+                         [(set QPR:$dst, (v8i16 (NEONvmvnImm timm:$SIMM)))]>;
+
+def VMVNv2i32 : N1ModImm<1, 0b000, {?,?,?,?}, 0, 0, 1, 1, (outs DPR:$dst),
+                         (ins nModImm:$SIMM), IIC_VMOVImm,
+                         "vmvn", "i32", "$dst, $SIMM", "",
+                         [(set DPR:$dst, (v2i32 (NEONvmvnImm timm:$SIMM)))]>;
+def VMVNv4i32 : N1ModImm<1, 0b000, {?,?,?,?}, 0, 1, 1, 1, (outs QPR:$dst),
+                         (ins nModImm:$SIMM), IIC_VMOVImm,
+                         "vmvn", "i32", "$dst, $SIMM", "",
+                         [(set QPR:$dst, (v4i32 (NEONvmvnImm timm:$SIMM)))]>;
+}
+
 //   VMVN     : Vector Bitwise NOT
 def  VMVNd    : N2VX<0b11, 0b11, 0b00, 0b00, 0b01011, 0, 0,
                      (outs DPR:$dst), (ins DPR:$src), IIC_VSUBiD,
diff --git a/llvm/test/CodeGen/ARM/vmov.ll b/llvm/test/CodeGen/ARM/vmov.ll
index f80301863d7..5e872ab6d0b 100644
--- a/llvm/test/CodeGen/ARM/vmov.ll
+++ b/llvm/test/CodeGen/ARM/vmov.ll
@@ -18,6 +18,18 @@ define <4 x i16> @v_movi16b() nounwind {
 	ret <4 x i16> < i16 4096, i16 4096, i16 4096, i16 4096 >
 }
 
+define <4 x i16> @v_mvni16a() nounwind {
+;CHECK: v_mvni16a:
+;CHECK: vmvn.i16 d0, #0x10
+	ret <4 x i16> < i16 65519, i16 65519, i16 65519, i16 65519 >
+}
+
+define <4 x i16> @v_mvni16b() nounwind {
+;CHECK: v_mvni16b:
+;CHECK: vmvn.i16 d0, #0x1000
+	ret <4 x i16> < i16 61439, i16 61439, i16 61439, i16 61439 >
+}
+
 define <2 x i32> @v_movi32a() nounwind {
 ;CHECK: v_movi32a:
 ;CHECK: vmov.i32 d0, #0x20
@@ -54,6 +66,42 @@ define <2 x i32> @v_movi32f() nounwind {
 	ret <2 x i32> < i32 2162687, i32 2162687 >
 }
 
+define <2 x i32> @v_mvni32a() nounwind {
+;CHECK: v_mvni32a:
+;CHECK: vmvn.i32 d0, #0x20
+	ret <2 x i32> < i32 4294967263, i32 4294967263 >
+}
+
+define <2 x i32> @v_mvni32b() nounwind {
+;CHECK: v_mvni32b:
+;CHECK: vmvn.i32 d0, #0x2000
+	ret <2 x i32> < i32 4294959103, i32 4294959103 >
+}
+
+define <2 x i32> @v_mvni32c() nounwind {
+;CHECK: v_mvni32c:
+;CHECK: vmvn.i32 d0, #0x200000
+	ret <2 x i32> < i32 4292870143, i32 4292870143 >
+}
+
+define <2 x i32> @v_mvni32d() nounwind {
+;CHECK: v_mvni32d:
+;CHECK: vmvn.i32 d0, #0x20000000
+	ret <2 x i32> < i32 3758096383, i32 3758096383 >
+}
+
+define <2 x i32> @v_mvni32e() nounwind {
+;CHECK: v_mvni32e:
+;CHECK: vmvn.i32 d0, #0x20FF
+	ret <2 x i32> < i32 4294958848, i32 4294958848 >
+}
+
+define <2 x i32> @v_mvni32f() nounwind {
+;CHECK: v_mvni32f:
+;CHECK: vmvn.i32 d0, #0x20FFFF
+	ret <2 x i32> < i32 4292804608, i32 4292804608 >
+}
+
 define <1 x i64> @v_movi64() nounwind {
 ;CHECK: v_movi64:
 ;CHECK: vmov.i64 d0, #0xFF0000FF0000FFFF