Initial work to fold

2 files changed, 40 insertions, 3 deletions

diff --git a/lib/Target/R600/AMDILISelDAGToDAG.cpp b/lib/Target/R600/AMDILISelDAGToDAG.cpp
index 567b3e26cf..5b4b88f72a 100644
--- a/lib/Target/R600/AMDILISelDAGToDAG.cpp
+++ b/lib/Target/R600/AMDILISelDAGToDAG.cpp
@@ -23,7 +23,7 @@
 #include "llvm/CodeGen/SelectionDAG.h"
 #include <list>
 #include <queue>
-
+#include <iostream>
 using namespace llvm;
 
 //===----------------------------------------------------------------------===//
@@ -171,6 +171,32 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
     }
     break;
   }
+  case ISD::INTRINSIC_WO_CHAIN: {
+    unsigned Intrinsic = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
+    if (Intrinsic == AMDGPUIntrinsic::AMDIL_clamp) {
+      SDNode *SubOperand = NULL;
+      SDNode *PreviousSubOperand;
+      do {
+        PreviousSubOperand = SubOperand;
+        Select(N->getOperand(1).getNode());
+        SubOperand = N->getOperand(1).getNode();
+      } while (PreviousSubOperand != SubOperand);
+      unsigned SubResultOpNum = SubOperand->getNumOperands();
+      const R600InstrInfo *TII =
+          static_cast<const R600InstrInfo*>(TM.getInstrInfo());
+      if (!SubOperand->isMachineOpcode() || !TII->isALUInstr(SubOperand->getMachineOpcode()))
+        return CurDAG->SelectNodeTo(N, AMDGPU::CLAMP_R600, MVT::f32, SDValue(SubOperand,0));
+      unsigned ClampIdx =
+         TII->getOperandIdx(SubOperand->getMachineOpcode(), R600Operands::CLAMP);
+      std::vector<SDValue> Ops;
+      for (unsigned i = 0; i < SubResultOpNum; ++i) {
+        Ops.push_back(SubOperand->getOperand(i));
+      }
+      Ops[ClampIdx - 1] = CurDAG->getTargetConstant(1, MVT::i32);
+      return CurDAG->SelectNodeTo(N, SubOperand->getMachineOpcode(), SubOperand->getVTList(), Ops.data(), SubResultOpNum);
+    }
+    break;
+  }
   case ISD::ConstantFP:
   case ISD::Constant: {
     const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
@@ -303,6 +329,12 @@ bool AMDGPUDAGToDAGISel::FoldOperands(unsigned Opcode,
     TII->getOperandIdx(Opcode, R600Operands::SRC1_SEL),
     TII->getOperandIdx(Opcode, R600Operands::SRC2_SEL)
   };
+  int NegIdx[] = {
+    TII->getOperandIdx(Opcode, R600Operands::SRC0_NEG),
+    TII->getOperandIdx(Opcode, R600Operands::SRC1_NEG),
+    TII->getOperandIdx(Opcode, R600Operands::SRC2_NEG)
+  };
+
   for (unsigned i = 0; i < 3; i++) {
     if (OperandIdx[i] < 0)
       return false;
@@ -318,6 +350,11 @@ bool AMDGPUDAGToDAGISel::FoldOperands(unsigned Opcode,
       }
       }
       break;
+    case ISD::FNEG:
+      Ops[OperandIdx[i] - 1] = Operand.getOperand(0);
+      Ops[NegIdx[i] - 1] = CurDAG->getTargetConstant(1, MVT::i32);
+      return true;
+      break;
     default:
       break;
     }
diff --git a/lib/Target/R600/R600Instructions.td b/lib/Target/R600/R600Instructions.td
index d4709080af..a9aacd9336 100644
--- a/lib/Target/R600/R600Instructions.td
+++ b/lib/Target/R600/R600Instructions.td
@@ -1240,10 +1240,10 @@ let Predicates = [isEGorCayman] in {
   // which do not need to be truncated since the fp values are 0.0f or 1.0f.
   // We should look into handling these cases separately.
   def : Pat<(fp_to_sint R600_Reg32:$src0),
-    (FLT_TO_INT_eg (TRUNC R600_Reg32:$src0))>;
+    (FLT_TO_INT_eg R600_Reg32:$src0)>;
 
   def : Pat<(fp_to_uint R600_Reg32:$src0),
-    (FLT_TO_UINT_eg (TRUNC R600_Reg32:$src0))>;
+    (FLT_TO_UINT_eg R600_Reg32:$src0)>;
 
   def EG_ExportSwz : ExportSwzInst {
     let Word1{19-16} = 1; // BURST_COUNT