AVX-512: Handled extractelement from mask vector;

Added VMOSHDUP/VMOVSLDUP shuffle instructions.

llvm-svn: 194691

lib/Target/X86/X86ISelLowering.cpp

@@ -16323,6 +16323,44 @@ static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
                      EltNo);
 }
 
+/// Extract one bit from mask vector, like v16i1 or v8i1.
+/// AVX-512 feature.
+static SDValue ExtractBitFromMaskVector(SDNode *N, SelectionDAG &DAG) {
+  SDValue Vec = N->getOperand(0);
+  SDLoc dl(Vec);
+  MVT VecVT = Vec.getSimpleValueType();
+  SDValue Idx = N->getOperand(1);
+  MVT EltVT = N->getSimpleValueType(0);
+  
+  assert((VecVT.getVectorElementType() == MVT::i1 && EltVT == MVT::i8) ||
+         "Unexpected operands in ExtractBitFromMaskVector");
+
+  // variable index
+  if (!isa<ConstantSDNode>(Idx)) {
+    MVT ExtVT = (VecVT == MVT::v8i1 ?  MVT::v8i64 : MVT::v16i32);
+    SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
+    SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
+                              ExtVT.getVectorElementType(), Ext);
+    return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
+  }
+
+  unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+
+  MVT ScalarVT = MVT::getIntegerVT(VecVT.getSizeInBits());
+  unsigned MaxShift = VecVT.getSizeInBits() - 1;
+  Vec = DAG.getNode(ISD::BITCAST, dl, ScalarVT, Vec);
+  Vec = DAG.getNode(ISD::SHL, dl, ScalarVT, Vec, 
+              DAG.getConstant(MaxShift - IdxVal, ScalarVT));
+  Vec = DAG.getNode(ISD::SRL, dl, ScalarVT, Vec,
+    DAG.getConstant(MaxShift, ScalarVT));
+
+  if (VecVT == MVT::v16i1) {
+    Vec = DAG.getNode(ISD::BITCAST, dl, MVT::i16, Vec);
+    return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Vec);
+  }
+  return DAG.getNode(ISD::BITCAST, dl, MVT::i8, Vec);
+}
+
 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
 /// generation and convert it from being a bunch of shuffles and extracts
 /// to a simple store and scalar loads to extract the elements.
@@ -16333,6 +16371,11 @@ static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
     return NewOp;
 
   SDValue InputVector = N->getOperand(0);
+
+  if (InputVector.getValueType().getVectorElementType() == MVT::i1 &&
+      !DCI.isBeforeLegalize())
+    return ExtractBitFromMaskVector(N, DAG);
+
   // Detect whether we are trying to convert from mmx to i32 and the bitcast
   // from mmx to v2i32 has a single usage.
   if (InputVector.getNode()->getOpcode() == llvm::ISD::BITCAST &&

lib/Target/X86/X86InstrAVX512.td

@@ -2075,6 +2075,38 @@ defm VMOVDDUPZ : avx512_movddup<"vmovddup", VR512, v8f64, f512mem, memopv8f64>,
 def : Pat<(X86Movddup (v8f64 (scalar_to_vector (loadf64 addr:$src)))),
           (VMOVDDUPZrm addr:$src)>;
 
+//===---------------------------------------------------------------------===//
+// Replicate Single FP - MOVSHDUP and MOVSLDUP
+//===---------------------------------------------------------------------===//
+multiclass avx512_replicate_sfp<bits<8> op, SDNode OpNode, string OpcodeStr,
+                              ValueType vt, RegisterClass RC, PatFrag mem_frag,
+                              X86MemOperand x86memop> {
+  def rr : AVX512XSI<op, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
+                    !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
+                      [(set RC:$dst, (vt (OpNode RC:$src)))]>, EVEX;
+  let mayLoad = 1 in
+  def rm : AVX512XSI<op, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
+                    !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
+                      [(set RC:$dst, (OpNode (mem_frag addr:$src)))]>, EVEX;
+}
+
+defm VMOVSHDUPZ  : avx512_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
+                       v16f32, VR512, memopv16f32, f512mem>, EVEX_V512,
+                       EVEX_CD8<32, CD8VF>;
+defm VMOVSLDUPZ  : avx512_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
+                       v16f32, VR512, memopv16f32, f512mem>, EVEX_V512,
+                       EVEX_CD8<32, CD8VF>;
+
+def : Pat<(v16i32 (X86Movshdup VR512:$src)), (VMOVSHDUPZrr VR512:$src)>;
+def : Pat<(v16i32 (X86Movshdup (memopv16i32 addr:$src))),
+           (VMOVSHDUPZrm addr:$src)>;
+def : Pat<(v16i32 (X86Movsldup VR512:$src)), (VMOVSLDUPZrr VR512:$src)>;
+def : Pat<(v16i32 (X86Movsldup (memopv16i32 addr:$src))),
+           (VMOVSLDUPZrm addr:$src)>;
+
+//===----------------------------------------------------------------------===//
+// Move Low to High and High to Low packed FP Instructions
+//===----------------------------------------------------------------------===//
 def VMOVLHPSZrr : AVX512PSI<0x16, MRMSrcReg, (outs VR128X:$dst),
           (ins VR128X:$src1, VR128X:$src2),
           "vmovlhps{z}\t{$src2, $src1, $dst|$dst, $src1, $src2}",
@@ -3385,6 +3417,7 @@ multiclass avx512_alignr<string OpcodeStr, RegisterClass RC,
                      !strconcat(OpcodeStr,
                      "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
                      []>, EVEX_4V;
+  let mayLoad = 1 in
   def rmi : AVX512AIi8<0x03, MRMSrcMem, (outs RC:$dst),
                      (ins RC:$src1, x86memop:$src2, i8imm:$src3),
                      !strconcat(OpcodeStr,

lib/Target/X86/X86InstrInfo.cpp

@@ -3165,7 +3165,8 @@ static unsigned getLoadStoreRegOpcode(unsigned Reg,
     assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
     return load ? X86::LD_Fp80m : X86::ST_FpP80m;
   case 16: {
-    assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
+    assert((X86::VR128RegClass.hasSubClassEq(RC) ||
+            X86::VR128XRegClass.hasSubClassEq(RC))&& "Unknown 16-byte regclass");
     // If stack is realigned we can use aligned stores.
     if (isStackAligned)
       return load ?
@@ -3177,7 +3178,8 @@ static unsigned getLoadStoreRegOpcode(unsigned Reg,
         (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
   }
   case 32:
-    assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
+    assert((X86::VR256RegClass.hasSubClassEq(RC) ||
+            X86::VR256XRegClass.hasSubClassEq(RC)) && "Unknown 32-byte regclass");
     // If stack is realigned we can use aligned stores.
     if (isStackAligned)
       return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;

test/CodeGen/X86/avx512-insert-extract.ll

@@ -98,3 +98,28 @@ define i32 @test10(<16 x i32> %x, i32 %ind) nounwind {
   ret i32 %e
 }
 
+;CHECK-LABEL: test11
+;CHECK: movl    $260
+;CHECK: bextrl
+;CHECK: movl    $268
+;CHECK: bextrl
+;CHECK: ret
+define <16 x i32> @test11(<16 x i32>%a, <16 x i32>%b) {
+  %cmp_res = icmp ult <16 x i32> %a, %b
+  %ia = extractelement <16 x i1> %cmp_res, i32 4
+  %ib = extractelement <16 x i1> %cmp_res, i32 12
+
+  br i1 %ia, label %A, label %B
+
+  A:
+    ret <16 x i32>%b
+  B:
+   %c = add <16 x i32>%b, %a
+  br i1 %ib, label %C, label %D
+  C:
+   %c1 = sub <16 x i32>%c, %a
+   ret <16 x i32>%c1
+  D:
+   %c2 = mul <16 x i32>%c, %a
+   ret <16 x i32>%c2
+}

test/CodeGen/X86/avx512-shuffle.ll

@@ -215,4 +215,12 @@ define <16 x i32> @test24(<16 x i32> %a, <16 x i32> %b) nounwind {
 define <16 x i32> @test25(<16 x i32> %a, <16 x i32> %b) nounwind {
   %c = shufflevector <16 x i32> %a, <16 x i32> %b, <16 x i32> <i32 0, i32 1, i32 19, i32 undef, i32 4, i32 5, i32 23, i32 undef, i32 8, i32 9, i32 27, i32 undef, i32 12, i32 13, i32 undef, i32 undef>
   ret <16 x i32> %c
+}
+
+; CHECK-LABEL: @test26
+; CHECK: vmovshdup
+; CHECK: ret
+define <16 x i32> @test26(<16 x i32> %a) nounwind {
+  %c = shufflevector <16 x i32> %a, <16 x i32> undef, <16 x i32> <i32 1, i32 1, i32 3, i32 3, i32 5, i32 5, i32 7, i32 undef, i32 9, i32 9, i32 undef, i32 11, i32 13, i32 undef, i32 undef, i32 undef>
+  ret <16 x i32> %c
 }