[X86][AVX512DQ] add scalar fpclass

Differential Revision: http://reviews.llvm.org/D13769

llvm-svn: 250650

4 files changed, 61 insertions, 6 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index b1bafba047c..a1137cf8bef 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -16012,6 +16012,8 @@ static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
 
   if (Op.getOpcode() == X86ISD::FSETCC)
     return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
+  if (Op.getOpcode() == X86ISD::VFPCLASS)
+    return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
 
   if (PreservedSrc.getOpcode() == ISD::UNDEF)
     PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
@@ -16357,6 +16359,15 @@ static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget
                                  DAG.getIntPtrConstant(0, dl));
        return DAG.getBitcast(Op.getValueType(), Res);
     }
+    case FPCLASSS: {
+      SDValue Src1 = Op.getOperand(1);
+      SDValue Imm = Op.getOperand(2);
+      SDValue Mask = Op.getOperand(3);
+      SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
+      SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
+        DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
+      return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
+    }
     case CMP_MASK:
     case CMP_MASK_CC: {
       // Comparison intrinsics with masks.
diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 31a7e95df15..a5a904873e5 100644
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -1803,6 +1803,42 @@ def : Pat<(v8i1 (X86cmpmu (v8i32 VR256X:$src1), (v8i32 VR256X:$src2), imm:$cc)),
 
 // ----------------------------------------------------------------
 // FPClass
+//handle fpclass instruction  mask =  op(reg_scalar,imm)
+//                                    op(mem_scalar,imm)
+multiclass avx512_scalar_fpclass<bits<8> opc, string OpcodeStr, SDNode OpNode,
+                                 X86VectorVTInfo _, Predicate prd> {
+  let Predicates = [prd] in {
+      def rr : AVX512<opc, MRMSrcReg, (outs _.KRC:$dst),//_.KRC:$dst),
+                      (ins _.RC:$src1, i32u8imm:$src2),
+                      OpcodeStr##_.Suffix#"\t{$src2, $src1, $dst | $dst, $src1, $src2}",
+                      [(set _.KRC:$dst,(OpNode (_.VT _.RC:$src1),
+                              (i32 imm:$src2)))], NoItinerary>;
+      def rrk : AVX512<opc, MRMSrcReg, (outs _.KRC:$dst),
+                      (ins _.KRCWM:$mask, _.RC:$src1, i32u8imm:$src2),
+                      OpcodeStr##_.Suffix#
+                      "\t{$src2, $src1, $dst {${mask}} | $dst {${mask}}, $src1, $src2}",
+                      [(set _.KRC:$dst,(or _.KRCWM:$mask, 
+                                      (OpNode (_.VT _.RC:$src1),
+                                      (i32 imm:$src2))))], NoItinerary>, EVEX_K;
+    let mayLoad = 1, AddedComplexity = 20 in {
+      def rm : AVX512<opc, MRMSrcMem, (outs _.KRC:$dst),
+                      (ins _.MemOp:$src1, i32u8imm:$src2),
+                      OpcodeStr##_.Suffix##
+                                "\t{$src2, $src1, $dst | $dst, $src1, $src2}",
+                      [(set _.KRC:$dst,
+                            (OpNode (_.VT (bitconvert (_.LdFrag addr:$src1))),
+                                    (i32 imm:$src2)))], NoItinerary>;
+      def rmk : AVX512<opc, MRMSrcMem, (outs _.KRC:$dst),
+                      (ins _.KRCWM:$mask, _.MemOp:$src1, i32u8imm:$src2),
+                      OpcodeStr##_.Suffix##
+                      "\t{$src2, $src1, $dst {${mask}} | $dst {${mask}}, $src1, $src2}",
+                      [(set _.KRC:$dst,(or _.KRCWM:$mask, 
+                          (OpNode (_.VT (bitconvert (_.LdFrag addr:$src1))),
+                              (i32 imm:$src2))))], NoItinerary>, EVEX_K;
+    }
+  }
+}
+
 //handle fpclass instruction mask = fpclass(reg_vec, reg_vec, imm)
 //                                  fpclass(reg_vec, mem_vec, imm)
 //                                  fpclass(reg_vec, broadcast(eltVt), imm)
@@ -1873,15 +1909,19 @@ multiclass avx512_vector_fpclass_all<string OpcodeStr,
 }
 
 multiclass avx512_fp_fpclass_all<string OpcodeStr, bits<8> opcVec,
-                                 SDNode OpNode, Predicate prd>{
+             bits<8> opcScalar, SDNode VecOpNode, SDNode ScalarOpNode, Predicate prd>{
   defm PS : avx512_vector_fpclass_all<OpcodeStr,  avx512vl_f32_info, opcVec, 
-                                      OpNode, prd, "{l}">, EVEX_CD8<32, CD8VF>;
+                                      VecOpNode, prd, "{l}">, EVEX_CD8<32, CD8VF>;
   defm PD : avx512_vector_fpclass_all<OpcodeStr,  avx512vl_f64_info, opcVec, 
-                                      OpNode, prd, "{q}">,EVEX_CD8<64, CD8VF> , VEX_W;
+                                      VecOpNode, prd, "{q}">,EVEX_CD8<64, CD8VF> , VEX_W;
+  defm SS : avx512_scalar_fpclass<opcScalar, OpcodeStr, ScalarOpNode,
+                                      f32x_info, prd>, EVEX_CD8<32, CD8VT1>;
+  defm SD : avx512_scalar_fpclass<opcScalar, OpcodeStr, ScalarOpNode,
+                                      f64x_info, prd>, EVEX_CD8<64, CD8VT1>, VEX_W;
 }
 
-defm VFPCLASS : avx512_fp_fpclass_all<"vfpclass", 0x66, X86Vfpclass, HasDQI>,
-                                      AVX512AIi8Base,EVEX;
+defm VFPCLASS : avx512_fp_fpclass_all<"vfpclass", 0x66, 0x67, X86Vfpclass,
+                                      X86Vfpclasss, HasDQI>, AVX512AIi8Base,EVEX;
 
 //-----------------------------------------------------------------
 // Mask register copy, including
diff --git a/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td b/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td
index ca67a2cff72..3eb5a6e5281 100644
--- a/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td
+++ b/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td
@@ -376,6 +376,8 @@ def X86VGetMant    : SDNode<"X86ISD::VGETMANT",  SDTFPUnaryOpImmRound>;
 def X86Vfpclass    : SDNode<"X86ISD::VFPCLASS", 
                        SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCVecEltisVT<0, i1>,
                                             SDTCisVec<1>, SDTCisInt<2>]>, []>;
+def X86Vfpclasss   : SDNode<"X86ISD::VFPCLASS", SDTypeProfile<1, 2, [SDTCisInt<0>,
+                              SDTCisFP<1>, SDTCisInt<2>]>,[]>;
 
 def X86SubVBroadcast : SDNode<"X86ISD::SUBV_BROADCAST",
                     SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
diff --git a/llvm/lib/Target/X86/X86IntrinsicsInfo.h b/llvm/lib/Target/X86/X86IntrinsicsInfo.h
index 74d6e4e6d73..2d8d9364c86 100644
--- a/llvm/lib/Target/X86/X86IntrinsicsInfo.h
+++ b/llvm/lib/Target/X86/X86IntrinsicsInfo.h
@@ -18,7 +18,7 @@ namespace llvm {
 
 enum IntrinsicType {
   INTR_NO_TYPE,
-  GATHER, SCATTER, PREFETCH, RDSEED, RDRAND, RDPMC, RDTSC, XTEST, ADX, FPCLASS,
+  GATHER, SCATTER, PREFETCH, RDSEED, RDRAND, RDPMC, RDTSC, XTEST, ADX, FPCLASS, FPCLASSS,
   INTR_TYPE_1OP, INTR_TYPE_2OP, INTR_TYPE_2OP_IMM8, INTR_TYPE_3OP, INTR_TYPE_4OP,
   CMP_MASK, CMP_MASK_CC,CMP_MASK_SCALAR_CC, VSHIFT, VSHIFT_MASK, COMI,
   INTR_TYPE_1OP_MASK, INTR_TYPE_1OP_MASK_RM,
@@ -688,6 +688,8 @@ static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx512_mask_fpclass_ps_128, FPCLASS, X86ISD::VFPCLASS, 0), 
   X86_INTRINSIC_DATA(avx512_mask_fpclass_ps_256, FPCLASS, X86ISD::VFPCLASS, 0),
   X86_INTRINSIC_DATA(avx512_mask_fpclass_ps_512, FPCLASS, X86ISD::VFPCLASS, 0), 
+  X86_INTRINSIC_DATA(avx512_mask_fpclass_sd, FPCLASSS, X86ISD::VFPCLASS, 0),

+  X86_INTRINSIC_DATA(avx512_mask_fpclass_ss, FPCLASSS, X86ISD::VFPCLASS, 0),
   X86_INTRINSIC_DATA(avx512_mask_getexp_pd_128, INTR_TYPE_1OP_MASK_RM,
                      X86ISD::FGETEXP_RND, 0),
   X86_INTRINSIC_DATA(avx512_mask_getexp_pd_256, INTR_TYPE_1OP_MASK_RM,