[PowerPC] Improvements for BUILD_VECTOR Vol. 3

[nemanjai]

Summary

This is the third (and penultimate) patch that https://reviews.llvm.org/D25580 is split into.
The test case will be added with the final patch.

In this patch, support for converting a vector of loads into a single load if the loads are consecutive (in either direction).
There is currently a limitation in that consecutive loads are not recognized if one of them is a pre-increment load.

[kbarton]

I think the loop on line 10619 needs to be reworked a bit (or I don't understand what it's trying to do).

lib/Target/PowerPC/PPCISelLowering.cpp

10601

Please convert to doxygen-style comments.

10625

++i

10625

Haven't you already checked the first operand? Can this loop not start at 1 to ensure the remainder of the operands conform to the necessary conditions? That would eliminate the if (i==0) check below also.

[nemanjai]

lib/Target/PowerPC/PPCISelLowering.cpp

10601

OK.

10625

Good point. This is a remnant of a previous incarnation of this logic which got refactored but obviously incompletely. I'll simplify the loop.

[nemanjai]

Fixed the weird loop and added doxygen comments.

[kbarton]

One minor question, more for my own curiosity, that we can likely address offline. Aside from that, LGTM.

lib/Target/PowerPC/PPCISelLowering.cpp

10678

Any reason for a std::vector vs a SmallVector?

[nemanjai]

lib/Target/PowerPC/PPCISelLowering.cpp

10678

That's a good point. The SmallVector is a better choice here. I'll change it prior to comitting.

[nemanjai]

Committed revision 288219.

lib/Target/PowerPC/PPCISelLowering.cpp

if (FirstConversion == PPCISD::FCTIDZ ||
 
     EVT NewVT = TargetVT == MVT::v2i64 ? MVT::v2f64 : MVT::v4f32;
     SDValue BV = DAG.getBuildVector(NewVT, dl, Ops);
     return DAG.getNode(Opcode, dl, TargetVT, BV);
   }
   return SDValue();
 }
 
+/// \brief Reduce the number of loads when building a vector.
+///
+/// Building a vector out of multiple loads can be converted to a load
+/// of the vector type if the loads are consecutive. If the loads are
+/// consecutive but in descending order, a shuffle is added at the end
+/// to reorder the vector.
+static SDValue combineBVOfConsecutiveLoads(SDNode *N, SelectionDAG &DAG) {
+  assert(N->getOpcode() == ISD::BUILD_VECTOR &&
+         "Should be called with a BUILD_VECTOR node");
+
+  SDLoc dl(N);
+  bool InputsAreConsecutiveLoads = true;
+  bool InputsAreReverseConsecutive = true;
+  unsigned ElemSize = N->getValueType(0).getScalarSizeInBits() / 8;
+  SDValue FirstInput = N->getOperand(0);
+  bool IsRoundOfExtLoad = false;
+
+  if (FirstInput.getOpcode() == ISD::FP_ROUND &&
+      FirstInput.getOperand(0).getOpcode() == ISD::LOAD) {
+    LoadSDNode *LD = dyn_cast<LoadSDNode>(FirstInput.getOperand(0));
+    IsRoundOfExtLoad = LD->getExtensionType() == ISD::EXTLOAD;
+  }
+  // Not a build vector of (possibly fp_rounded) loads.
+  if (!IsRoundOfExtLoad && FirstInput.getOpcode() != ISD::LOAD)
+    return SDValue();
+
+  for (int i = 1, e = N->getNumOperands(); i < e; ++i) {
+    // If any inputs are fp_round(extload), they all must be.
+    if (IsRoundOfExtLoad && N->getOperand(i).getOpcode() != ISD::FP_ROUND)
+      return SDValue();
+
+    SDValue NextInput = IsRoundOfExtLoad ? N->getOperand(i).getOperand(0) :
+      N->getOperand(i);
+    if (NextInput.getOpcode() != ISD::LOAD)
+      return SDValue();
+
+    SDValue PreviousInput =
+      IsRoundOfExtLoad ? N->getOperand(i-1).getOperand(0) : N->getOperand(i-1);
+    LoadSDNode *LD1 = dyn_cast<LoadSDNode>(PreviousInput);
+    LoadSDNode *LD2 = dyn_cast<LoadSDNode>(NextInput);
+
+    // If any inputs are fp_round(extload), they all must be.
+    if (IsRoundOfExtLoad && LD2->getExtensionType() != ISD::EXTLOAD)
+      return SDValue();
+
+    if (!isConsecutiveLS(LD2, LD1, ElemSize, 1, DAG))
+      InputsAreConsecutiveLoads = false;
+    if (!isConsecutiveLS(LD1, LD2, ElemSize, 1, DAG))
+      InputsAreReverseConsecutive = false;
+
+    // Exit early if the loads are neither consecutive nor reverse consecutive.
+    if (!InputsAreConsecutiveLoads && !InputsAreReverseConsecutive)
+      return SDValue();
+  }
+
+  assert(!(InputsAreConsecutiveLoads && InputsAreReverseConsecutive) &&
+         "The loads cannot be both consecutive and reverse consecutive.");
+
+  SDValue FirstLoadOp =
+    IsRoundOfExtLoad ? FirstInput.getOperand(0) : FirstInput;
+  SDValue LastLoadOp =
+    IsRoundOfExtLoad ? N->getOperand(N->getNumOperands()-1).getOperand(0) :
+                       N->getOperand(N->getNumOperands()-1);
+
+  LoadSDNode *LD1 = dyn_cast<LoadSDNode>(FirstLoadOp);
+  LoadSDNode *LDL = dyn_cast<LoadSDNode>(LastLoadOp);
+  if (InputsAreConsecutiveLoads) {
+    assert(LD1 && "Input needs to be a LoadSDNode.");
+    return DAG.getLoad(N->getValueType(0), dl, LD1->getChain(),
+                       LD1->getBasePtr(), LD1->getPointerInfo(),
+                       LD1->getAlignment());
+  }
+  if (InputsAreReverseConsecutive) {
+    assert(LDL && "Input needs to be a LoadSDNode.");
+    SDValue Load = DAG.getLoad(N->getValueType(0), dl, LDL->getChain(),
+                               LDL->getBasePtr(), LDL->getPointerInfo(),
+                               LDL->getAlignment());
+    std::vector<int> Ops;
+    for (int i = N->getNumOperands() - 1; i >= 0; i--)
+      Ops.push_back(i);
+
+    return DAG.getVectorShuffle(N->getValueType(0), dl, Load,
+                                DAG.getUNDEF(N->getValueType(0)), Ops);
+  }
+  return SDValue();
+}
+
 SDValue PPCTargetLowering::DAGCombineBuildVector(SDNode *N,
                                                  DAGCombinerInfo &DCI) const {
   assert(N->getOpcode() == ISD::BUILD_VECTOR &&
          "Should be called with a BUILD_VECTOR node");
 
   SelectionDAG &DAG = DCI.DAG;
   SDLoc dl(N);
 
   if (!Subtarget.hasVSX())
     return SDValue();
 
   // The target independent DAG combiner will leave a build_vector of
   // float-to-int conversions intact. We can generate MUCH better code for
   // a float-to-int conversion of a vector of floats.
   SDValue FirstInput = N->getOperand(0);
   if (FirstInput.getOpcode() == PPCISD::MFVSR) {
     SDValue Reduced = combineElementTruncationToVectorTruncation(N, DCI);
     if (Reduced)
       return Reduced;
   }
+
+  // If we're building a vector out of consecutive loads, just load that
+  // vector type.
+  SDValue Reduced = combineBVOfConsecutiveLoads(N, DAG);
+  if (Reduced)
+    return Reduced;
+
   if (N->getValueType(0) != MVT::v2f64)
     return SDValue();
 
   // Looking for:
   // (build_vector ([su]int_to_fp (extractelt 0)), [su]int_to_fp (extractelt 1))
-  if (N->getOperand(0).getOpcode() != ISD::SINT_TO_FP &&
-      N->getOperand(0).getOpcode() != ISD::UINT_TO_FP)
+  if (FirstInput.getOpcode() != ISD::SINT_TO_FP &&
+      FirstInput.getOpcode() != ISD::UINT_TO_FP)
     return SDValue();
   if (N->getOperand(1).getOpcode() != ISD::SINT_TO_FP &&
       N->getOperand(1).getOpcode() != ISD::UINT_TO_FP)
     return SDValue();
-  if (N->getOperand(0).getOpcode() != N->getOperand(1).getOpcode())
+  if (FirstInput.getOpcode() != N->getOperand(1).getOpcode())
     return SDValue();
 
-  SDValue Ext1 = N->getOperand(0).getOperand(0);
+  SDValue Ext1 = FirstInput.getOperand(0);
   SDValue Ext2 = N->getOperand(1).getOperand(0);
   if(Ext1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
      Ext2.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
     return SDValue();
 
   ConstantSDNode *Ext1Op = dyn_cast<ConstantSDNode>(Ext1.getOperand(1));
   ConstantSDNode *Ext2Op = dyn_cast<ConstantSDNode>(Ext2.getOperand(1));
   if (!Ext1Op || !Ext2Op)