R600: Move CONST_ADDRESS folding into AMDGPUDAGToDAGISel::Select()

This increases the number of opportunites we have for folding.  With the
previous implementation we were unable to fold into any instructions
other than the first when multiple instructions were selected from a
single SDNode.

Reviewed-by: Vincent Lejeune <vljn at ovi.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@186919 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/R600/AMDILISelDAGToDAG.cpp

@@ -50,7 +50,7 @@ public:
 private:
   inline SDValue getSmallIPtrImm(unsigned Imm);
   bool FoldOperand(SDValue &Src, SDValue &Sel, SDValue &Neg, SDValue &Abs,
-                   const R600InstrInfo *TII, std::vector<unsigned> Cst);
+                   const R600InstrInfo *TII);
   bool FoldOperands(unsigned, const R600InstrInfo *, std::vector<SDValue> &);
   bool FoldDotOperands(unsigned, const R600InstrInfo *, std::vector<SDValue> &);
 
@@ -158,12 +158,100 @@ bool AMDGPUDAGToDAGISel::SelectADDR64(SDValue Addr, SDValue& R1, SDValue& R2) {
 }
 
 SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
+  const R600InstrInfo *TII =
+                      static_cast<const R600InstrInfo*>(TM.getInstrInfo());
   unsigned int Opc = N->getOpcode();
   if (N->isMachineOpcode()) {
     return NULL;   // Already selected.
   }
   switch (Opc) {
   default: break;
+  case AMDGPUISD::CONST_ADDRESS: {
+    for (SDNode::use_iterator I = N->use_begin(), Next = llvm::next(I);
+                              I != SDNode::use_end(); I = Next) {
+      Next = llvm::next(I);
+      if (!I->isMachineOpcode()) {
+        continue;
+      }
+      unsigned Opcode = I->getMachineOpcode();
+      bool HasDst = TII->getOperandIdx(Opcode, AMDGPU::OpName::dst) > -1;
+      int SrcIdx = I.getOperandNo();
+      int SelIdx;
+      // Unlike MachineInstrs, SDNodes do not have results in their operand
+      // list, so we need to increment the SrcIdx, since
+      // R600InstrInfo::getOperandIdx is based on the MachineInstr indices.
+      if (HasDst) {
+        SrcIdx++;
+      }
+
+      SelIdx = TII->getSelIdx(I->getMachineOpcode(), SrcIdx);
+      if (SelIdx < 0) {
+        continue;
+      }
+
+      SDValue CstOffset;
+      if (N->getValueType(0).isVector() ||
+          !SelectGlobalValueConstantOffset(N->getOperand(0), CstOffset))
+        continue;
+
+      // Gather constants values
+      int SrcIndices[] = {
+        TII->getOperandIdx(Opcode, AMDGPU::OpName::src0),
+        TII->getOperandIdx(Opcode, AMDGPU::OpName::src1),
+        TII->getOperandIdx(Opcode, AMDGPU::OpName::src2),
+        TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_X),
+        TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_Y),
+        TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_Z),
+        TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_W),
+        TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_X),
+        TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_Y),
+        TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_Z),
+        TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_W)
+      };
+      std::vector<unsigned> Consts;
+      for (unsigned i = 0; i < sizeof(SrcIndices) / sizeof(int); i++) {
+        int OtherSrcIdx = SrcIndices[i];
+        int OtherSelIdx = TII->getSelIdx(Opcode, OtherSrcIdx);
+        if (OtherSrcIdx < 0 || OtherSelIdx < 0) {
+          continue;
+        }
+        if (HasDst) {
+          OtherSrcIdx--;
+          OtherSelIdx--;
+        }
+        if (RegisterSDNode *Reg =
+                         dyn_cast<RegisterSDNode>(I->getOperand(OtherSrcIdx))) {
+          if (Reg->getReg() == AMDGPU::ALU_CONST) {
+            ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(I->getOperand(OtherSelIdx));
+            Consts.push_back(Cst->getZExtValue());
+          }
+        }
+      }
+
+      ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(CstOffset);
+      Consts.push_back(Cst->getZExtValue());
+      if (!TII->fitsConstReadLimitations(Consts))
+        continue;
+
+      // Convert back to SDNode indices
+      if (HasDst) {
+        SrcIdx--;
+        SelIdx--;
+      }
+      std::vector<SDValue> Ops;
+      for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
+        if (i == SrcIdx) {
+          Ops.push_back(CurDAG->getRegister(AMDGPU::ALU_CONST, MVT::f32));
+        } else if (i == SelIdx) {
+          Ops.push_back(CstOffset);
+        } else {
+          Ops.push_back(I->getOperand(i));
+        }
+      }
+      CurDAG->UpdateNodeOperands(*I, Ops.data(), Ops.size());
+    }
+    break;
+  }
   case ISD::BUILD_VECTOR: {
     const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
     if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
@@ -224,7 +312,6 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
     if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
       break;
     }
-    const R600InstrInfo *TII = static_cast<const R600InstrInfo*>(TM.getInstrInfo());
 
     uint64_t ImmValue = 0;
     unsigned ImmReg = AMDGPU::ALU_LITERAL_X;
@@ -342,7 +429,7 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
     if (Result && Result->isMachineOpcode() &&
         !(TII->get(Result->getMachineOpcode()).TSFlags & R600_InstFlag::VECTOR)
         && TII->hasInstrModifiers(Result->getMachineOpcode())) {
-      // Fold FNEG/FABS/CONST_ADDRESS
+      // Fold FNEG/FABS
       // TODO: Isel can generate multiple MachineInst, we need to recursively
       // parse Result
       bool IsModified = false;
@@ -382,24 +469,8 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
 }
 
 bool AMDGPUDAGToDAGISel::FoldOperand(SDValue &Src, SDValue &Sel, SDValue &Neg,
-                                     SDValue &Abs, const R600InstrInfo *TII,
-                                     std::vector<unsigned> Consts) {
+                                     SDValue &Abs, const R600InstrInfo *TII) {
   switch (Src.getOpcode()) {
-  case AMDGPUISD::CONST_ADDRESS: {
-    SDValue CstOffset;
-    if (Src.getValueType().isVector() ||
-        !SelectGlobalValueConstantOffset(Src.getOperand(0), CstOffset))
-      return false;
-
-    ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(CstOffset);
-    Consts.push_back(Cst->getZExtValue());
-    if (!TII->fitsConstReadLimitations(Consts))
-      return false;
-
-    Src = CurDAG->getRegister(AMDGPU::ALU_CONST, MVT::f32);
-    Sel = CstOffset;
-    return true;
-    }
   case ISD::FNEG:
     Src = Src.getOperand(0);
     Neg = CurDAG->getTargetConstant(1, MVT::i32);
@@ -441,19 +512,6 @@ bool AMDGPUDAGToDAGISel::FoldOperands(unsigned Opcode,
     -1
   };
 
-  // Gather constants values
-  std::vector<unsigned> Consts;
-  for (unsigned j = 0; j < 3; j++) {
-    int SrcIdx = OperandIdx[j];
-    if (SrcIdx < 0)
-      break;
-    if (RegisterSDNode *Reg = dyn_cast<RegisterSDNode>(Ops[SrcIdx - 1])) {
-      if (Reg->getReg() == AMDGPU::ALU_CONST) {
-        ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Ops[SelIdx[j] - 1]);
-        Consts.push_back(Cst->getZExtValue());
-      }
-    }
-  }
 
   for (unsigned i = 0; i < 3; i++) {
     if (OperandIdx[i] < 0)
@@ -463,7 +521,7 @@ bool AMDGPUDAGToDAGISel::FoldOperands(unsigned Opcode,
     SDValue &Neg = Ops[NegIdx[i] - 1];
     SDValue FakeAbs;
     SDValue &Abs = (AbsIdx[i] > -1) ? Ops[AbsIdx[i] - 1] : FakeAbs;
-    if (FoldOperand(Src, Sel, Neg, Abs, TII, Consts))
+    if (FoldOperand(Src, Sel, Neg, Abs, TII))
       return true;
   }
   return false;
@@ -512,20 +570,6 @@ bool AMDGPUDAGToDAGISel::FoldDotOperands(unsigned Opcode,
     TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_abs_W)
   };
 
-  // Gather constants values
-  std::vector<unsigned> Consts;
-  for (unsigned j = 0; j < 8; j++) {
-    int SrcIdx = OperandIdx[j];
-    if (SrcIdx < 0)
-      break;
-    if (RegisterSDNode *Reg = dyn_cast<RegisterSDNode>(Ops[SrcIdx - 1])) {
-      if (Reg->getReg() == AMDGPU::ALU_CONST) {
-        ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Ops[SelIdx[j] - 1]);
-        Consts.push_back(Cst->getZExtValue());
-      }
-    }
-  }
-
   for (unsigned i = 0; i < 8; i++) {
     if (OperandIdx[i] < 0)
       return false;
@@ -533,7 +577,7 @@ bool AMDGPUDAGToDAGISel::FoldDotOperands(unsigned Opcode,
     SDValue &Sel = Ops[SelIdx[i] - 1];
     SDValue &Neg = Ops[NegIdx[i] - 1];
     SDValue &Abs = Ops[AbsIdx[i] - 1];
-    if (FoldOperand(Src, Sel, Neg, Abs, TII, Consts))
+    if (FoldOperand(Src, Sel, Neg, Abs, TII))
       return true;
   }
   return false;

lib/Target/R600/R600ISelLowering.cpp

@@ -1154,6 +1154,30 @@ SDValue R600TargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const
     return DAG.getMergeValues(MergedValues, 2, DL);
   }
 
+  // For most operations returning SDValue() will result int he node being
+  // expanded by the DAG Legalizer.  This is not the case for ISD::LOAD, so
+  // we need to manually expand loads that may be legal in some address spaces
+  // and illegal in others.  SEXT loads from CONSTANT_BUFFER_0 are supported
+  // for compute shaders, since the data is sign extended when it is uploaded
+  // to the buffer.  Howerver SEXT loads from other addresspaces are not
+  // supported, so we need to expand them here.
+  if (LoadNode->getExtensionType() == ISD::SEXTLOAD) {
+    EVT MemVT = LoadNode->getMemoryVT();
+    assert(!MemVT.isVector() && (MemVT == MVT::i16 || MemVT == MVT::i8));
+    SDValue ShiftAmount =
+          DAG.getConstant(VT.getSizeInBits() - MemVT.getSizeInBits(), MVT::i32);
+    SDValue NewLoad = DAG.getExtLoad(ISD::EXTLOAD, DL, VT, Chain, Ptr,
+                                  LoadNode->getPointerInfo(), MemVT,
+                                  LoadNode->isVolatile(),
+                                  LoadNode->isNonTemporal(),
+                                  LoadNode->getAlignment());
+    SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, NewLoad, ShiftAmount);
+    SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, Shl, ShiftAmount);
+
+    SDValue MergedValues[2] = { Sra, Chain };
+    return DAG.getMergeValues(MergedValues, 2, DL);
+  }
+
   if (LoadNode->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
     return SDValue();
   }

lib/Target/R600/R600InstrInfo.cpp

@@ -186,6 +186,42 @@ bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const {
   }
 }
 
+int R600InstrInfo::getSrcIdx(unsigned Opcode, unsigned SrcNum) const {
+  static const unsigned OpTable[] = {
+    AMDGPU::OpName::src0,
+    AMDGPU::OpName::src1,
+    AMDGPU::OpName::src2
+  };
+
+  assert (SrcNum < 3);
+  return getOperandIdx(Opcode, OpTable[SrcNum]);
+}
+
+#define SRC_SEL_ROWS 11
+int R600InstrInfo::getSelIdx(unsigned Opcode, unsigned SrcIdx) const {
+  static const unsigned SrcSelTable[SRC_SEL_ROWS][2] = {
+    {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
+    {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
+    {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
+    {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
+    {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
+    {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
+    {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
+    {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
+    {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
+    {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
+    {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W}
+  };
+
+  for (unsigned i = 0; i < SRC_SEL_ROWS; ++i) {
+    if (getOperandIdx(Opcode, SrcSelTable[i][0]) == (int)SrcIdx) {
+      return getOperandIdx(Opcode, SrcSelTable[i][1]);
+    }
+  }
+  return -1;
+}
+#undef SRC_SEL_ROWS
+
 SmallVector<std::pair<MachineOperand *, int64_t>, 3>
 R600InstrInfo::getSrcs(MachineInstr *MI) const {
   SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;

lib/Target/R600/R600InstrInfo.h

@@ -76,6 +76,13 @@ namespace llvm {
 
   bool mustBeLastInClause(unsigned Opcode) const;
 
+  /// \returns The operand index for the given source number.  Legal values
+  /// for SrcNum are 0, 1, and 2.
+  int getSrcIdx(unsigned Opcode, unsigned SrcNum) const;
+  /// \returns The operand Index for the Sel operand given an index to one
+  /// of the instruction's src operands.
+  int getSelIdx(unsigned Opcode, unsigned SrcIdx) const;
+
   /// \returns a pair for each src of an ALU instructions.
   /// The first member of a pair is the register id.
   /// If register is ALU_CONST, second member is SEL.

test/CodeGen/R600/bfi_int.ll

@@ -36,7 +36,7 @@ entry:
 ; SHA-256 Ma function
 ; ((x & z) | (y & (x | z)))
 ; R600-CHECK: @bfi_sha256_ma
-; R600-CHECK: XOR_INT * [[DST:T[0-9]+\.[XYZW]]],
+; R600-CHECK: XOR_INT * [[DST:T[0-9]+\.[XYZW]]], KC0[2].Z, KC0[2].W
 ; R600-CHECK: BFI_INT * {{T[0-9]+\.[XYZW]}}, {{[[DST]]|PV\.[XYZW]}}, KC0[3].X, KC0[2].W
 ; SI-CHECK: V_XOR_B32_e64 [[DST:VGPR[0-9]+]], {{[SV]GPR[0-9]+, [SV]GPR[0-9]+}}
 ; SI-CHECK: V_BFI_B32 {{VGPR[0-9]+}}, [[DST]], {{[SV]GPR[0-9]+, [SV]GPR[0-9]+}}