From 64c7fa3201f2db0348e1a48ec4a2f9caaa681f38 Mon Sep 17 00:00:00 2001
From: Benjamin Kramer <benny.kra@googlemail.com>
Date: Fri, 20 Jul 2018 20:59:46 +0000
Subject: [PATCH] Revert "[X86][AVX] Convert X86ISD::VBROADCAST demanded elts
 combine to use SimplifyDemandedVectorElts"

This reverts commit r337547. It triggers an infinite loop.

llvm-svn: 337617
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 58 ++++++++++-----------------------
 llvm/lib/Target/X86/X86ISelLowering.h   |  7 ----
 2 files changed, 17 insertions(+), 48 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 92a095a7..f17e4bf 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -30634,13 +30634,24 @@ static SDValue combineTargetShuffle(SDValue N, SelectionDAG &DAG,
 
   switch (Opcode) {
   case X86ISD::VBROADCAST: {
+    // If broadcasting from another shuffle, attempt to simplify it.
     // TODO - we really need a general SimplifyDemandedVectorElts mechanism.
-    APInt KnownUndef, KnownZero;
-    APInt DemandedMask(APInt::getAllOnesValue(VT.getVectorNumElements()));
-    const TargetLowering &TLI = DAG.getTargetLoweringInfo();
-    if (TLI.SimplifyDemandedVectorElts(N, DemandedMask, KnownUndef, KnownZero,
-                                       DCI)) {
-      return SDValue(N.getNode(), 0);
+    SDValue Src = N.getOperand(0);
+    SDValue BC = peekThroughBitcasts(Src);
+    EVT SrcVT = Src.getValueType();
+    EVT BCVT = BC.getValueType();
+    if (isTargetShuffle(BC.getOpcode()) &&
+        VT.getScalarSizeInBits() % BCVT.getScalarSizeInBits() == 0) {
+      unsigned Scale = VT.getScalarSizeInBits() / BCVT.getScalarSizeInBits();
+      SmallVector<int, 16> DemandedMask(BCVT.getVectorNumElements(),
+                                        SM_SentinelUndef);
+      for (unsigned i = 0; i != Scale; ++i)
+        DemandedMask[i] = i;
+      if (SDValue Res = combineX86ShufflesRecursively(
+              {BC}, 0, BC, DemandedMask, {}, /*Depth*/ 1,
+              /*HasVarMask*/ false, DAG, Subtarget))
+        return DAG.getNode(X86ISD::VBROADCAST, DL, VT,
+                           DAG.getBitcast(SrcVT, Res));
     }
     return SDValue();
   }
@@ -31271,41 +31282,6 @@ static SDValue combineShuffle(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
-bool X86TargetLowering::SimplifyDemandedVectorEltsForTargetNode(
-    SDValue Op, const APInt &DemandedElts, APInt &KnownUndef, APInt &KnownZero,
-    TargetLoweringOpt &TLO, unsigned Depth) const {
-
-  if (X86ISD::VBROADCAST != Op.getOpcode())
-    return false;
-
-  EVT VT = Op.getValueType();
-  SDValue Src = Op.getOperand(0);
-  SDValue BC = peekThroughBitcasts(Src);
-  EVT SrcVT = Src.getValueType();
-  EVT BCVT = BC.getValueType();
-
-  if (!isTargetShuffle(BC.getOpcode()) ||
-      (VT.getScalarSizeInBits() % BCVT.getScalarSizeInBits()) != 0)
-    return false;
-
-  unsigned Scale = VT.getScalarSizeInBits() / BCVT.getScalarSizeInBits();
-  SmallVector<int, 16> DemandedMask(BCVT.getVectorNumElements(),
-                                    SM_SentinelUndef);
-  for (unsigned i = 0; i != Scale; ++i)
-    DemandedMask[i] = i;
-
-  if (SDValue Res = combineX86ShufflesRecursively(
-          {BC}, 0, BC, DemandedMask, {}, Depth + 1, /*HasVarMask*/ false,
-          TLO.DAG, Subtarget)) {
-    SDLoc DL(Op);
-    Res = TLO.DAG.getNode(X86ISD::VBROADCAST, DL, VT,
-                          TLO.DAG.getBitcast(SrcVT, Res));
-    return TLO.CombineTo(Op, Res);
-  }
-
-  return false;
-}
-
 /// Check if a vector extract from a target-specific shuffle of a load can be
 /// folded into a single element load.
 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
diff --git a/llvm/lib/Target/X86/X86ISelLowering.h b/llvm/lib/Target/X86/X86ISelLowering.h
index 623b95b..32215b1 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.h
+++ b/llvm/lib/Target/X86/X86ISelLowering.h
@@ -866,13 +866,6 @@ namespace llvm {
                                              const SelectionDAG &DAG,
                                              unsigned Depth) const override;
 
-    bool SimplifyDemandedVectorEltsForTargetNode(SDValue Op,
-                                                 const APInt &DemandedElts,
-                                                 APInt &KnownUndef,
-                                                 APInt &KnownZero,
-                                                 TargetLoweringOpt &TLO,
-                                                 unsigned Depth) const override;
-
     SDValue unwrapAddress(SDValue N) const override;
 
     bool isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
-- 
2.7.4