AVX-512: insert element to mask vector; store i1 data

Implemented INSERT_VECTOR_ELT operation for v16i1 and v8i1 vectors;
Implemented "store" for i1 type


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205850 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 124 insertions, 1 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 5426656f25..bc2c4e9f66 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1394,6 +1394,8 @@ void X86TargetLowering::resetOperationActions() {
 
     setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1,  Custom);
     setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
+    setOperationAction(ISD::INSERT_VECTOR_ELT,  MVT::v16i1, Custom);
+    setOperationAction(ISD::INSERT_VECTOR_ELT,  MVT::v8i1, Custom);
     setOperationAction(ISD::BUILD_VECTOR,       MVT::v8i1, Custom);
     setOperationAction(ISD::BUILD_VECTOR,       MVT::v16i1, Custom);
     setOperationAction(ISD::SELECT,             MVT::v8f64, Custom);
@@ -5811,6 +5813,8 @@ X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
   uint64_t Immediate = 0;
   int NonConstIdx = -1;
   bool IsSplat = true;
+  unsigned NumNonConsts = 0;
+  unsigned NumConsts = 0;
   for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
     SDValue In = Op.getOperand(idx);
     if (In.getOpcode() == ISD::UNDEF)
@@ -5818,9 +5822,13 @@ X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
     if (!isa<ConstantSDNode>(In)) {
       AllContants = false;
       NonConstIdx = idx;
+      NumNonConsts++;
     }
-    else if (cast<ConstantSDNode>(In)->getZExtValue())
+    else {
+      NumConsts++;
+      if (cast<ConstantSDNode>(In)->getZExtValue())
       Immediate |= (1ULL << idx);
+    }
     if (In != Op.getOperand(0))
       IsSplat = false;
   }
@@ -5832,6 +5840,19 @@ X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
                        DAG.getIntPtrConstant(0));
   }
 
+  if (NumNonConsts == 1 && NonConstIdx != 0) {
+    SDValue DstVec;
+    if (NumConsts) {
+      SDValue VecAsImm = DAG.getConstant(Immediate,
+                                         MVT::getIntegerVT(VT.getSizeInBits()));
+      DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
+    }
+    else 
+      DstVec = DAG.getUNDEF(VT);
+    return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
+                       Op.getOperand(NonConstIdx),
+                       DAG.getIntPtrConstant(NonConstIdx));
+  }
   if (!IsSplat && (NonConstIdx != 0))
     llvm_unreachable("Unsupported BUILD_VECTOR operation");
   MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
@@ -7948,10 +7969,47 @@ static SDValue LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
   return SDValue();
 }
 
+/// Insert one bit to mask vector, like v16i1 or v8i1.
+/// AVX-512 feature.
+SDValue 
+X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
+  SDLoc dl(Op);
+  SDValue Vec = Op.getOperand(0);
+  SDValue Elt = Op.getOperand(1);
+  SDValue Idx = Op.getOperand(2);
+  MVT VecVT = Vec.getSimpleValueType();
+
+  if (!isa<ConstantSDNode>(Idx)) {
+    // Non constant index. Extend source and destination,
+    // insert element and then truncate the result.
+    MVT ExtVecVT = (VecVT == MVT::v8i1 ?  MVT::v8i64 : MVT::v16i32);
+    MVT ExtEltVT = (VecVT == MVT::v8i1 ?  MVT::i64 : MVT::i32);
+    SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT, 
+      DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
+      DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
+    return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
+  }
+
+  unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+  SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
+  if (Vec.getOpcode() == ISD::UNDEF)
+    return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
+                       DAG.getConstant(IdxVal, MVT::i8));
+  const TargetRegisterClass* rc = getRegClassFor(VecVT);
+  unsigned MaxSift = rc->getSize()*8 - 1;
+  EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
+                    DAG.getConstant(MaxSift, MVT::i8));
+  EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec,
+                    DAG.getConstant(MaxSift - IdxVal, MVT::i8));
+  return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
+}
 SDValue
 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
   MVT VT = Op.getSimpleValueType();
   MVT EltVT = VT.getVectorElementType();
+  
+  if (EltVT == MVT::i1)
+    return InsertBitToMaskVector(Op, DAG);
 
   SDLoc dl(Op);
   SDValue N0 = Op.getOperand(0);
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 3d4cf342a8..c430f7e2ae 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -874,6 +874,8 @@ namespace llvm {
     SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
     SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const;
+    SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const;
+
     SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index 1bc58fe663..34d517fb72 100644
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -984,6 +984,10 @@ let Predicates = [HasAVX512] in {
             (EXTRACT_SUBREG
              (AND32ri (KMOVWrk (COPY_TO_REGCLASS VK1:$src, VK16)), (i32 1)),
               sub_16bit)>;
+  def : Pat<(v16i1 (scalar_to_vector VK1:$src)),
+            (COPY_TO_REGCLASS VK1:$src, VK16)>;
+  def : Pat<(v8i1 (scalar_to_vector VK1:$src)),
+            (COPY_TO_REGCLASS VK1:$src, VK8)>;
 }
 // With AVX-512 only, 8-bit mask is promoted to 16-bit mask.
 let Predicates = [HasAVX512] in {
@@ -1356,6 +1360,14 @@ defm VMOVDQU64: avx512_load<0x6F, VR512, VK8WM, i512mem, load,
                               "vmovdqu64", SSEPackedInt, v8i64>,
                                XS, VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
 
+def: Pat<(v16i32 (int_x86_avx512_mask_loadu_d_512 addr:$ptr,
+                 (v16i32 immAllZerosV), GR16:$mask)),
+       (VMOVDQU32rmkz (v16i1 (COPY_TO_REGCLASS GR16:$mask, VK16WM)), addr:$ptr)>;
+
+def: Pat<(v8i64 (int_x86_avx512_mask_loadu_q_512 addr:$ptr,
+                 (bc_v8i64 (v16i32 immAllZerosV)), GR8:$mask)),
+       (VMOVDQU64rmkz (v8i1 (COPY_TO_REGCLASS GR8:$mask, VK8WM)), addr:$ptr)>;
+
 let AddedComplexity = 20 in {
 def : Pat<(v8i64 (vselect VK8WM:$mask, (v8i64 VR512:$src),
                            (bc_v8i64 (v16i32 immAllZerosV)))),
@@ -4211,3 +4223,7 @@ def : Pat<(int_x86_avx512_mask_conflict_q_512 VR512:$src2, VR512:$src1,
                                               GR8:$mask),
           (VPCONFLICTQrrk VR512:$src1,
            (v8i1 (COPY_TO_REGCLASS GR8:$mask, VK8WM)), VR512:$src2)>;
+
+def : Pat<(store (i1 -1), addr:$dst), (MOV8mi addr:$dst, (i8 1))>;
+def : Pat<(store (i1  1), addr:$dst), (MOV8mi addr:$dst, (i8 1))>;
+def : Pat<(store (i1  0), addr:$dst), (MOV8mi addr:$dst, (i8 0))>;
diff --git a/test/CodeGen/X86/avx512-insert-extract.ll b/test/CodeGen/X86/avx512-insert-extract.ll
index 6557ac3493..b360c716b0 100644
--- a/test/CodeGen/X86/avx512-insert-extract.ll
+++ b/test/CodeGen/X86/avx512-insert-extract.ll
@@ -158,3 +158,41 @@ define i64 @test14(<8 x i64>%a, <8 x i64>%b, i64 %a1, i64 %b1) {
   %res = select i1 %extract24vector_func.i, i64 %a1, i64 %b1
   ret i64 %res
 }
+
+;CHECK-LABEL: test15
+;CHECK: kshiftlw
+;CHECK: kmovw
+;CHECK: ret
+define i16 @test15(i1 *%addr) {
+  %x = load i1 * %addr, align 128
+  %x1 = insertelement <16 x i1> undef, i1 %x, i32 10
+  %x2 = bitcast <16 x i1>%x1 to i16
+  ret i16 %x2
+}
+
+;CHECK-LABEL: test16
+;CHECK: kshiftlw
+;CHECK: kshiftrw
+;CHECK: korw
+;CHECK: ret
+define i16 @test16(i1 *%addr, i16 %a) {
+  %x = load i1 * %addr, align 128
+  %a1 = bitcast i16 %a to <16 x i1>
+  %x1 = insertelement <16 x i1> %a1, i1 %x, i32 10
+  %x2 = bitcast <16 x i1>%x1 to i16
+  ret i16 %x2
+}
+
+;CHECK-LABEL: test17
+;CHECK: kshiftlw
+;CHECK: kshiftrw
+;CHECK: korw
+;CHECK: ret
+define i8 @test17(i1 *%addr, i8 %a) {
+  %x = load i1 * %addr, align 128
+  %a1 = bitcast i8 %a to <8 x i1>
+  %x1 = insertelement <8 x i1> %a1, i1 %x, i32 10
+  %x2 = bitcast <8 x i1>%x1 to i8
+  ret i8 %x2
+}
+
diff --git a/test/CodeGen/X86/avx512-mov.ll b/test/CodeGen/X86/avx512-mov.ll
index 13e6843244..a769eb61d9 100644
--- a/test/CodeGen/X86/avx512-mov.ll
+++ b/test/CodeGen/X86/avx512-mov.ll
@@ -153,3 +153,12 @@ define void @test18(i8 * %addr, <8 x i64> %data) {
   ret void
 }
 
+; CHECK-LABEL: store_i1
+; CHECK: movb
+; CHECK: movb
+; CHECK: ret
+define void @store_i1() {
+  store i1 true, i1 addrspace(3)* undef, align 128
+  store i1 false, i1 addrspace(2)* undef, align 128
+  ret void
+}
\ No newline at end of file