[X86][XOP] Added VPERMIL2PD/VPERMIL2PS raw mask decoding for target shuffle combines

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@271834 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/Utils/X86ShuffleDecode.cpp b/lib/Target/X86/Utils/X86ShuffleDecode.cpp
index 472474d..4ec4b78 100644 (file)
--- a/lib/Target/X86/Utils/X86ShuffleDecode.cpp
+++ b/lib/Target/X86/Utils/X86ShuffleDecode.cpp
@@ -512,6 +512,46 @@ void DecodeVPERMILPMask(MVT VT, ArrayRef<uint64_t> RawMask,
   }
 }
 
+void DecodeVPERMIL2PMask(MVT VT, unsigned M2Z, ArrayRef<uint64_t> RawMask,
+                         SmallVectorImpl<int> &ShuffleMask) {
+  unsigned VecSize = VT.getSizeInBits();
+  unsigned EltSize = VT.getScalarSizeInBits();
+  unsigned NumLanes = VecSize / 128;
+  unsigned NumEltsPerLane = VT.getVectorNumElements() / NumLanes;
+  assert((VecSize == 128 || VecSize == 256) &&
+         "Unexpected vector size");
+  assert((EltSize == 32 || EltSize == 64) && "Unexpected element size");
+
+  for (unsigned i = 0, e = RawMask.size(); i < e; ++i) {
+    // VPERMIL2 Operation.
+    // Bits[3] - Match Bit.
+    // Bits[2:1] - (Per Lane) PD Shuffle Mask.
+    // Bits[2:0] - (Per Lane) PS Shuffle Mask.
+    uint64_t Selector = RawMask[i];
+    int MatchBit = (Selector >> 3) & 0x1;
+
+    // M2Z[0:1]     MatchBit
+    //   0Xb           X        Source selected by Selector index.
+    //   10b           0        Source selected by Selector index.
+    //   10b           1        Zero.
+    //   11b           0        Zero.
+    //   11b           1        Source selected by Selector index.
+    if ((M2Z & 0x2) != 0 && MatchBit != (M2Z & 0x1)) {
+      ShuffleMask.push_back(SM_SentinelZero);
+      continue;
+    }
+
+    unsigned Index = i & ~(NumEltsPerLane - 1);
+    if (EltSize == 64)
+      Index += (Selector >> 1) & 0x1;
+    else
+      Index += Selector & 0x3;
+
+    unsigned SrcOffset = (Selector >> 2) & 1;
+    ShuffleMask.push_back((int)(SrcOffset + Index));
+  }
+}
+
 void DecodeVPERMVMask(ArrayRef<uint64_t> RawMask,
                       SmallVectorImpl<int> &ShuffleMask) {
   for (int i = 0, e = RawMask.size(); i < e; ++i) {

diff --git a/lib/Target/X86/Utils/X86ShuffleDecode.h b/lib/Target/X86/Utils/X86ShuffleDecode.h
index ec05a66..f1caa56 100644 (file)
--- a/lib/Target/X86/Utils/X86ShuffleDecode.h
+++ b/lib/Target/X86/Utils/X86ShuffleDecode.h
@@ -138,6 +138,10 @@ void DecodeINSERTQIMask(int Len, int Idx,
 void DecodeVPERMILPMask(MVT VT, ArrayRef<uint64_t> RawMask,
                         SmallVectorImpl<int> &ShuffleMask);
 
+/// Decode a VPERMIL2PD/VPERMIL2PS variable mask from a raw array of constants.
+void DecodeVPERMIL2PMask(MVT VT, unsigned M2Z, ArrayRef<uint64_t> RawMask,
+                         SmallVectorImpl<int> &ShuffleMask);
+
 /// Decode a VPERM W/D/Q/PS/PD mask from a raw array of constants.
 void DecodeVPERMVMask(ArrayRef<uint64_t> RawMask,
                       SmallVectorImpl<int> &ShuffleMask);

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 12e8f60..f861a0c 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -4937,6 +4937,11 @@ static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero,
     SDValue CtrlNode = N->getOperand(3);
     if (ConstantSDNode *CtrlOp = dyn_cast<ConstantSDNode>(CtrlNode)) {
       unsigned CtrlImm = CtrlOp->getZExtValue();
+      SmallVector<uint64_t, 32> RawMask;
+      if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask)) {
+        DecodeVPERMIL2PMask(VT, CtrlImm, RawMask, Mask);
+        break;
+      }
       if (auto *C = getTargetShuffleMaskConstant(MaskNode)) {
         DecodeVPERMIL2PMask(C, CtrlImm, MaskEltSize, Mask);
         break;

diff --git a/test/CodeGen/X86/vector-shuffle-combining-xop.ll b/test/CodeGen/X86/vector-shuffle-combining-xop.ll
index 0eb59d6..8ad8e13 100644 (file)
--- a/test/CodeGen/X86/vector-shuffle-combining-xop.ll
+++ b/test/CodeGen/X86/vector-shuffle-combining-xop.ll
@@ -13,10 +13,7 @@ declare <16 x i8> @llvm.x86.xop.vpperm(<16 x i8>, <16 x i8>, <16 x i8>) nounwind
 define <2 x double> @combine_vpermil2pd_identity(<2 x double> %a0, <2 x double> %a1) {
 ; CHECK-LABEL: combine_vpermil2pd_identity:
 ; CHECK:       # BB#0:
-; CHECK-NEXT:    movl $2, %eax
-; CHECK-NEXT:    vmovq %rax, %xmm2
-; CHECK-NEXT:    vpermil2pd $0, %xmm2, %xmm0, %xmm1, %xmm0
-; CHECK-NEXT:    vpermil2pd $0, %xmm2, %xmm0, %xmm0, %xmm0
+; CHECK-NEXT:    vmovaps %xmm1, %xmm0
 ; CHECK-NEXT:    retq
   %res0 = call <2 x double> @llvm.x86.xop.vpermil2pd(<2 x double> %a1, <2 x double> %a0, <2 x i64> <i64 2, i64 0>, i8 0)
   %res1 = call <2 x double> @llvm.x86.xop.vpermil2pd(<2 x double> %res0, <2 x double> undef, <2 x i64> <i64 2, i64 0>, i8 0)