From 55c790e29c334af76790f6ee01df6b0575f668ec Mon Sep 17 00:00:00 2001
From: Krzysztof Parzyszek <kparzysz@codeaurora.org>
Date: Thu, 3 Dec 2015 16:47:20 +0000
Subject: [PATCH] [Hexagon] Implement CONCAT_VECTORS for HVX using V6_vcombine

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@254617 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/Hexagon/HexagonISelLowering.cpp | 10 ++++++++++
 lib/Target/Hexagon/HexagonISelLowering.h   |  1 +
 lib/Target/Hexagon/HexagonInstrInfoV60.td  | 16 +++++++++++++++-
 3 files changed, 26 insertions(+), 1 deletion(-)

diff --git a/lib/Target/Hexagon/HexagonISelLowering.cpp b/lib/Target/Hexagon/HexagonISelLowering.cpp
index 04f5b664929..a75f391a4ee 100644
--- a/lib/Target/Hexagon/HexagonISelLowering.cpp
+++ b/lib/Target/Hexagon/HexagonISelLowering.cpp
@@ -2042,6 +2042,7 @@ const char* HexagonTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case HexagonISD::VCMPWEQ:       return "HexagonISD::VCMPWEQ";
   case HexagonISD::VCMPWGT:       return "HexagonISD::VCMPWGT";
   case HexagonISD::VCMPWGTU:      return "HexagonISD::VCMPWGTU";
+  case HexagonISD::VCOMBINE:      return "HexagonISD::VCOMBINE";
   case HexagonISD::VSHLH:         return "HexagonISD::VSHLH";
   case HexagonISD::VSHLW:         return "HexagonISD::VSHLW";
   case HexagonISD::VSPLATB:       return "HexagonISD::VSPLTB";
@@ -2346,6 +2347,7 @@ SDValue
 HexagonTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
                                            SelectionDAG &DAG) const {
   SDLoc dl(Op);
+  bool UseHVX = Subtarget.useHVXOps();
   EVT VT = Op.getValueType();
   unsigned NElts = Op.getNumOperands();
   SDValue Vec = Op.getOperand(0);
@@ -2376,6 +2378,14 @@ HexagonTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
     }
   }
 
+  if (UseHVX) {
+    SDValue Vec0 = Op.getOperand(1);
+    uint64_t VS = VecVT.getSizeInBits();
+    assert((VS == 64*8 && Subtarget.useHVXSglOps()) ||
+           (VS == 128*8 && Subtarget.useHVXDblOps()));
+    SDValue Combined = DAG.getNode(HexagonISD::VCOMBINE, dl, VT, Vec0, Vec);
+    return Combined;
+  }
   for (unsigned i = 0, e = NElts; i != e; ++i) {
     unsigned OpIdx = NElts - i - 1;
     SDValue Operand = Op.getOperand(OpIdx);
diff --git a/lib/Target/Hexagon/HexagonISelLowering.h b/lib/Target/Hexagon/HexagonISelLowering.h
index 64033d95ee3..b6d39fe9172 100644
--- a/lib/Target/Hexagon/HexagonISelLowering.h
+++ b/lib/Target/Hexagon/HexagonISelLowering.h
@@ -80,6 +80,7 @@ bool isPositiveHalfWord(SDNode *N);
       INSERTRP,
       EXTRACTU,
       EXTRACTURP,
+      VCOMBINE,
       TC_RETURN,
       EH_RETURN,
       DCFETCH,
diff --git a/lib/Target/Hexagon/HexagonInstrInfoV60.td b/lib/Target/Hexagon/HexagonInstrInfoV60.td
index 394df0fdd6e..897ada08153 100644
--- a/lib/Target/Hexagon/HexagonInstrInfoV60.td
+++ b/lib/Target/Hexagon/HexagonInstrInfoV60.td
@@ -1535,6 +1535,20 @@ let isRegSequence = 1, Itinerary = CVI_VA_DV, Type = TypeCVI_VA_DV in
 defm V6_vcombine :
      T_HVX_alu_WV <"$dst = vcombine($src1,$src2)">, V6_vcombine_enc;
 
+def SDTHexagonVCOMBINE: SDTypeProfile<1, 2, [SDTCisSameAs<1, 2>,
+      SDTCisSubVecOfVec<1, 0>]>;
+
+def HexagonVCOMBINE: SDNode<"HexagonISD::VCOMBINE", SDTHexagonVCOMBINE>;
+
+def: Pat<(v32i32 (HexagonVCOMBINE (v16i32 VectorRegs:$Vs),
+                                  (v16i32 VectorRegs:$Vt))),
+         (V6_vcombine VectorRegs:$Vs, VectorRegs:$Vt)>,
+         Requires<[UseHVXSgl]>;
+def: Pat<(v64i32 (HexagonVCOMBINE (v32i32 VecDblRegs:$Vs),
+                                  (v32i32 VecDblRegs:$Vt))),
+         (V6_vcombine_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>,
+         Requires<[UseHVXDbl]>;
+
 let Itinerary = CVI_VINLANESAT, Type = TypeCVI_VINLANESAT in {
 defm V6_vsathub :
      T_HVX_alu_VV <"$dst.ub = vsat($src1.h,$src2.h)">, V6_vsathub_enc;
@@ -1872,7 +1886,7 @@ defm V6_vasrhbrndsat :
      V6_vasrhbrndsat_enc;
 }
 
-// Assemlber mapped -- alias?
+// Assembler mapped -- alias?
 //defm V6_vtran2x2vdd : T_HVX_shift_VV <"">, V6_vtran2x2vdd_enc;
 let Itinerary = CVI_VP_VS_LONG, Type = TypeCVI_VP_VS in {
 defm V6_vshuffvdd :
-- 
2.34.1