AVX-512: Fixed extract_vector_elt for v8i1 vector

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@202624 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 14 insertions, 6 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 71aee4ff3d..3ae4147a79 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -7697,7 +7697,8 @@ static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
 
 /// Extract one bit from mask vector, like v16i1 or v8i1.
 /// AVX-512 feature.
-static SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) {
+SDValue
+X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
   SDValue Vec = Op.getOperand(0);
   SDLoc dl(Vec);
   MVT VecVT = Vec.getSimpleValueType();
@@ -7717,7 +7718,8 @@ static SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) {
   }
 
   unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
-  unsigned MaxSift = VecVT.getSizeInBits() - 1;
+  const TargetRegisterClass* rc = getRegClassFor(VecVT);
+  unsigned MaxSift = rc->getSize()*8 - 1;
   Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
                     DAG.getConstant(MaxSift - IdxVal, MVT::i8));
   Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index af97b15cc1..9bf2b90f8f 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -867,6 +867,7 @@ namespace llvm {
     SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const;
     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 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 328d74f8d7..825ea09cfe 100644
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -1213,6 +1213,11 @@ def : Pat<(v16i1 (insert_subvector undef, (v8i1 VK8:$src), (iPTR 0))),
 def : Pat<(v8i1 (extract_subvector (v16i1 VK16:$src), (iPTR 8))),
           (v8i1 (COPY_TO_REGCLASS (KSHIFTRWri VK16:$src, (i8 8)), VK8))>;
 
+def : Pat<(v8i1 (X86vshli VK8:$src, (i8 imm:$imm))),
+          (v8i1 (COPY_TO_REGCLASS (KSHIFTLWri (COPY_TO_REGCLASS VK8:$src, VK16), (I8Imm $imm)), VK8))>;
+
+def : Pat<(v8i1 (X86vsrli VK8:$src, (i8 imm:$imm))),
+          (v8i1 (COPY_TO_REGCLASS (KSHIFTRWri (COPY_TO_REGCLASS VK8:$src, VK16), (I8Imm $imm)), VK8))>;
 //===----------------------------------------------------------------------===//
 // AVX-512 - Aligned and unaligned load and store
 //
diff --git a/lib/Target/X86/X86RegisterInfo.td b/lib/Target/X86/X86RegisterInfo.td
index a88b2bb1f1..33c402b69a 100644
--- a/lib/Target/X86/X86RegisterInfo.td
+++ b/lib/Target/X86/X86RegisterInfo.td
@@ -466,10 +466,10 @@ def VR256X : RegisterClass<"X86", [v32i8, v16i16, v8i32, v4i64, v8f32, v4f64],
 // The size of the all masked registers is 16 bit because we have only one
 // KMOVW istruction that can store this register in memory, and it writes 2 bytes
 def VK1     : RegisterClass<"X86", [i1],    16, (sequence "K%u", 0, 7)>;
-def VK8     : RegisterClass<"X86", [v8i1],  16, (sequence "K%u", 0, 7)>;
-def VK16    : RegisterClass<"X86", [v16i1], 16, (add VK8)>;
+def VK8     : RegisterClass<"X86", [v8i1],  16, (add VK1)> {let Size = 16;}
+def VK16    : RegisterClass<"X86", [v16i1], 16, (add VK8)> {let Size = 16;}
 
-def VK1WM   : RegisterClass<"X86", [i1],    16, (sub VK1, K0)>;
-def VK8WM   : RegisterClass<"X86", [v8i1],  16, (sub VK8, K0)>;
+def VK1WM   : RegisterClass<"X86", [i1],    16, (sub VK1, K0)> {let Size = 16;}
+def VK8WM   : RegisterClass<"X86", [v8i1],  16, (sub VK8, K0)> {let Size = 16;}
 def VK16WM  : RegisterClass<"X86", [v16i1], 16, (add VK8WM)>;