[AVX512] Handle valign masking intrinsic via C++ lowering

I think that this will scale better in most cases than adding a Pat<> for each
mapping from the intrinsic DAG to the intruction (i.e. rri, rrik, rrikz).  We
can just lower to the SDNode and have the resulting DAG be matches by the DAG
patterns.

Alternatively (long term), we could keep the Pat<>s but generate them via the
new AVX512_masking multiclass.  The difficulty is that in order to formulate
that we would have to concatenate DAGs.  Currently this is only supported if
the operators of the input DAGs are identical.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215473 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index f7e4ac9..2a85805 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -14412,6 +14412,22 @@ static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
   case Intrinsic::x86_avx_sqrt_pd_256:
     return DAG.getNode(ISD::FSQRT, dl, Op.getValueType(), Op.getOperand(1));
 
+  case Intrinsic::x86_avx512_mask_valign_q_512:
+  case Intrinsic::x86_avx512_mask_valign_d_512: {
+    EVT VT = Op.getValueType();
+    EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
+                                  MVT::i1, VT.getVectorNumElements());
+    assert(MaskVT.isSimple() && "invalid valign mask type");
+    // Vector source operands are swapped.
+    return DAG.getNode(ISD::VSELECT, dl, VT,
+                       DAG.getNode(ISD::BITCAST, dl, MaskVT,
+                                   Op.getOperand(5)),
+                       DAG.getNode(X86ISD::VALIGN, dl, VT,
+                                   Op.getOperand(2), Op.getOperand(1),
+                                   Op.getOperand(3)),
+                       Op.getOperand(4));
+  }
+
   // ptest and testp intrinsics. The intrinsic these come from are designed to
   // return an integer value, not just an instruction so lower it to the ptest
   // or testp pattern and a setcc for the result.

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index 3383654..02be95f 100644 (file)
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -4509,22 +4509,6 @@ multiclass avx512_valign<string Suffix, RegisterClass RC, RegisterClass KRC,
   def : Pat<(FloatVT (X86VAlign RC:$src1, RC:$src2, (i8 imm:$imm))),
             (!cast<Instruction>(NAME##rri) RC:$src2, RC:$src1, imm:$imm)>;
 
-  // Non-masking intrinsic call.
-  def : Pat<(IntVT
-               (!cast<Intrinsic>("int_x86_avx512_mask_valign_"##Suffix##"_512")
-                     RC:$src1, RC:$src2, imm:$src3,
-                     (IntVT (bitconvert (v16i32 immAllZerosV))), -1)),
-            (!cast<Instruction>(NAME#rri) RC:$src1, RC:$src2, imm:$src3)>;
-
-  // Masking intrinsic call.
-  def : Pat<(IntVT
-               (!cast<Intrinsic>("int_x86_avx512_mask_valign_"##Suffix##"_512")
-                     RC:$src1, RC:$src2, imm:$src3,
-                     RC:$src4, MRC:$mask)),
-            (!cast<Instruction>(NAME#rrik) RC:$src4,
-               (COPY_TO_REGCLASS MRC:$mask, KRC), RC:$src1,
-               RC:$src2, imm:$src3)>;
-
   let mayLoad = 1 in
   def rmi : AVX512AIi8<0x03, MRMSrcMem, (outs RC:$dst),
                      (ins RC:$src1, x86memop:$src2, i8imm:$src3),

diff --git a/test/CodeGen/X86/avx512-intrinsics.ll b/test/CodeGen/X86/avx512-intrinsics.ll
index 48ec3c0..a586f32 100644 (file)
--- a/test/CodeGen/X86/avx512-intrinsics.ll
+++ b/test/CodeGen/X86/avx512-intrinsics.ll
@@ -627,3 +627,12 @@ define <8 x i64> @test_mask_valign_q(<8 x i64> %a, <8 x i64> %b, <8 x i64> %src,
 }
 
 declare <8 x i64> @llvm.x86.avx512.mask.valign.q.512(<8 x i64>, <8 x i64>, i8, <8 x i64>, i8)
+
+define <16 x i32> @test_maskz_valign_d(<16 x i32> %a, <16 x i32> %b, i16 %mask) {
+; CHECK-LABEL: test_maskz_valign_d:
+; CHECK: valignd $5, %zmm1, %zmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf3,0x7d,0xc9,0x03,0xc1,0x05]
+  %res = call <16 x i32> @llvm.x86.avx512.mask.valign.d.512(<16 x i32> %a, <16 x i32> %b, i8 5, <16 x i32> zeroinitializer, i16 %mask)
+  ret <16 x i32> %res
+}
+
+declare <16 x i32> @llvm.x86.avx512.mask.valign.d.512(<16 x i32>, <16 x i32>, i8, <16 x i32>, i16)