Implement feedback from Bruno on making pblendvb an x86-specific ISD node in addition to being an intrinsic, and convert

lowering to use it.  Hopefully the pattern fragment is doing the right thing with XMM0, looks correct in testing.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122277 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 14 insertions, 11 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 55c83a5..e4561aa 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -8398,9 +8398,7 @@ SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
     M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
                     DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
                     DAG.getConstant(4, MVT::i32));
-    R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
-                    DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
-                    R, M, Op);
+    R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
     // a += a
     Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
 
@@ -8415,15 +8413,12 @@ SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
     M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
                     DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
                     DAG.getConstant(2, MVT::i32));
-    R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
-                    DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
-                    R, M, Op);
+    R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
     // a += a
     Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
 
     // return pblendv(r, r+r, a);
-    R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
-                    DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
+    R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, 
                     R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
     return R;
   }
@@ -8897,6 +8892,7 @@ const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
   case X86ISD::PSIGNB:             return "X86ISD::PSIGNB";
   case X86ISD::PSIGNW:             return "X86ISD::PSIGNW";
   case X86ISD::PSIGND:             return "X86ISD::PSIGND";
+  case X86ISD::PBLENDVB:           return "X86ISD::PBLENDVB";
   case X86ISD::FMAX:               return "X86ISD::FMAX";
   case X86ISD::FMIN:               return "X86ISD::FMIN";
   case X86ISD::FRSQRT:             return "X86ISD::FRSQRT";
@@ -11209,12 +11205,10 @@ static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
         if (!Subtarget->hasSSE41())
           return SDValue();
         
-        unsigned IID = Intrinsic::x86_sse41_pblendvb;
         X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
         Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
         Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
-        Mask = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, MVT::v16i8,
-                           DAG.getConstant(IID, MVT::i32), X, Y, Mask);
+        Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
         return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
       }
     }
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 3b59ffa..1c33fee 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -166,6 +166,9 @@ namespace llvm {
       /// PSIGNB/W/D - Copy integer sign.
       PSIGNB, PSIGNW, PSIGND, 
       
+      /// PBLENDVB - Variable blend
+      PBLENDVB,
+      
       /// FMAX, FMIN - Floating point max and min.
       ///
       FMAX, FMIN,
diff --git a/lib/Target/X86/X86InstrFragmentsSIMD.td b/lib/Target/X86/X86InstrFragmentsSIMD.td
index 954dbbe..16c9bcd 100644
--- a/lib/Target/X86/X86InstrFragmentsSIMD.td
+++ b/lib/Target/X86/X86InstrFragmentsSIMD.td
@@ -55,6 +55,9 @@ def X86psignw  : SDNode<"X86ISD::PSIGNW",
 def X86psignd  : SDNode<"X86ISD::PSIGND", 
                  SDTypeProfile<1, 2, [SDTCisVT<0, v4i32>, SDTCisSameAs<0,1>,
                                       SDTCisSameAs<0,2>]>>;
+def X86pblendv : SDNode<"X86ISD::PBLENDVB", 
+                 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
+                                      SDTCisSameAs<0,2>, SDTCisSameAs<0,3>]>>;
 def X86pextrb  : SDNode<"X86ISD::PEXTRB",
                  SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
 def X86pextrw  : SDNode<"X86ISD::PEXTRW",
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 058b79b..ad80860 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -4855,6 +4855,9 @@ defm BLENDVPD     : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
 defm BLENDVPS     : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
 defm PBLENDVB     : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
 
+def : Pat<(X86pblendv VR128:$src1, VR128:$src2, XMM0),
+          (PBLENDVBrr0 VR128:$src1, VR128:$src2)>;
+
 let isAsmParserOnly = 1, Predicates = [HasAVX] in
 def VMOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
                        "vmovntdqa\t{$src, $dst|$dst, $src}",