R600/SI: Use v_cvt_f32_ubyte* instructions

This eliminates extra extract instructions when loading an i8 vector to
a float vector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@210666 91177308-0d34-0410-b5e6-96231b3b80d8

7 files changed, 170 insertions, 5 deletions

diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index 0f1a77a7d6..92ab174695 100644
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -786,6 +786,18 @@ SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
       return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
                          Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
 
+    case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
+      return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
+
+    case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
+      return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
+
+    case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
+      return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
+
+    case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
+      return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
+
     case AMDGPUIntrinsic::AMDGPU_bfe_i32:
       return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
                          Op.getOperand(1),
@@ -1256,7 +1268,6 @@ SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
   FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
                         DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
   return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
-
 }
 
 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
@@ -1582,6 +1593,10 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(SAMPLEB)
   NODE_NAME_CASE(SAMPLED)
   NODE_NAME_CASE(SAMPLEL)
+  NODE_NAME_CASE(CVT_F32_UBYTE0)
+  NODE_NAME_CASE(CVT_F32_UBYTE1)
+  NODE_NAME_CASE(CVT_F32_UBYTE2)
+  NODE_NAME_CASE(CVT_F32_UBYTE3)
   NODE_NAME_CASE(STORE_MSKOR)
   NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
   }
diff --git a/lib/Target/R600/AMDGPUISelLowering.h b/lib/Target/R600/AMDGPUISelLowering.h
index ee98b3bfd3..f50afababc 100644
--- a/lib/Target/R600/AMDGPUISelLowering.h
+++ b/lib/Target/R600/AMDGPUISelLowering.h
@@ -205,6 +205,12 @@ enum {
   SAMPLEB,
   SAMPLED,
   SAMPLEL,
+
+  // These cvt_f32_ubyte* nodes need to remain consecutive and in order.
+  CVT_F32_UBYTE0,
+  CVT_F32_UBYTE1,
+  CVT_F32_UBYTE2,
+  CVT_F32_UBYTE3,
   FIRST_MEM_OPCODE_NUMBER = ISD::FIRST_TARGET_MEMORY_OPCODE,
   STORE_MSKOR,
   LOAD_CONSTANT,
diff --git a/lib/Target/R600/AMDGPUInstrInfo.td b/lib/Target/R600/AMDGPUInstrInfo.td
index f96dbb4d8a..03137800d6 100644
--- a/lib/Target/R600/AMDGPUInstrInfo.td
+++ b/lib/Target/R600/AMDGPUInstrInfo.td
@@ -59,6 +59,17 @@ def AMDGPUumin : SDNode<"AMDGPUISD::UMIN", SDTIntBinOp,
   [SDNPCommutative, SDNPAssociative]
 >;
 
+
+def AMDGPUcvt_f32_ubyte0 : SDNode<"AMDGPUISD::CVT_F32_UBYTE0",
+  SDTIntToFPOp, []>;
+def AMDGPUcvt_f32_ubyte1 : SDNode<"AMDGPUISD::CVT_F32_UBYTE1",
+  SDTIntToFPOp, []>;
+def AMDGPUcvt_f32_ubyte2 : SDNode<"AMDGPUISD::CVT_F32_UBYTE2",
+  SDTIntToFPOp, []>;
+def AMDGPUcvt_f32_ubyte3 : SDNode<"AMDGPUISD::CVT_F32_UBYTE3",
+  SDTIntToFPOp, []>;
+
+
 // urecip - This operation is a helper for integer division, it returns the
 // result of 1 / a as a fractional unsigned integer.
 // out = (2^32 / a) + e
diff --git a/lib/Target/R600/AMDGPUIntrinsics.td b/lib/Target/R600/AMDGPUIntrinsics.td
index 9ad5e72d3f..5790357bf1 100644
--- a/lib/Target/R600/AMDGPUIntrinsics.td
+++ b/lib/Target/R600/AMDGPUIntrinsics.td
@@ -53,6 +53,10 @@ let TargetPrefix = "AMDGPU", isTarget = 1 in {
   def int_AMDGPU_imul24 : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
   def int_AMDGPU_imad24 : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
   def int_AMDGPU_umad24 : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_cvt_f32_ubyte0 : Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_cvt_f32_ubyte1 : Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_cvt_f32_ubyte2 : Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_cvt_f32_ubyte3 : Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>;
   def int_AMDGPU_cube : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty], [IntrNoMem]>;
   def int_AMDGPU_bfi : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
   def int_AMDGPU_bfe_i32 : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index 608aad2d3c..846aeb6309 100644
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -24,6 +24,7 @@
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/SelectionDAG.h"
 #include "llvm/IR/Function.h"
+#include "llvm/ADT/SmallString.h"
 
 using namespace llvm;
 
@@ -214,6 +215,8 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
   setTargetDAGCombine(ISD::SELECT_CC);
   setTargetDAGCombine(ISD::SETCC);
 
+  setTargetDAGCombine(ISD::UINT_TO_FP);
+
   setSchedulingPreference(Sched::RegPressure);
 }
 
@@ -979,6 +982,96 @@ SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
 // Custom DAG optimizations
 //===----------------------------------------------------------------------===//
 
+SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
+                                                     DAGCombinerInfo &DCI) {
+  EVT VT = N->getValueType(0);
+  EVT ScalarVT = VT.getScalarType();
+  if (ScalarVT != MVT::f32)
+    return SDValue();
+
+  SelectionDAG &DAG = DCI.DAG;
+  SDLoc DL(N);
+
+  SDValue Src = N->getOperand(0);
+  EVT SrcVT = Src.getValueType();
+
+  // TODO: We could try to match extracting the higher bytes, which would be
+  // easier if i8 vectors weren't promoted to i32 vectors, particularly after
+  // types are legalized. v4i8 -> v4f32 is probably the only case to worry
+  // about in practice.
+  if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) {
+    if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) {
+      SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src);
+      DCI.AddToWorklist(Cvt.getNode());
+      return Cvt;
+    }
+  }
+
+  // We are primarily trying to catch operations on illegal vector types
+  // before they are expanded.
+  // For scalars, we can use the more flexible method of checking masked bits
+  // after legalization.
+  if (!DCI.isBeforeLegalize() ||
+      !SrcVT.isVector() ||
+      SrcVT.getVectorElementType() != MVT::i8) {
+    return SDValue();
+  }
+
+  assert(DCI.isBeforeLegalize() && "Unexpected legal type");
+
+  // Weird sized vectors are a pain to handle, but we know 3 is really the same
+  // size as 4.
+  unsigned NElts = SrcVT.getVectorNumElements();
+  if (!SrcVT.isSimple() && NElts != 3)
+    return SDValue();
+
+  // Handle v4i8 -> v4f32 extload. Replace the v4i8 with a legal i32 load to
+  // prevent a mess from expanding to v4i32 and repacking.
+  if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse()) {
+    EVT LoadVT = getEquivalentMemType(*DAG.getContext(), SrcVT);
+    EVT RegVT = getEquivalentLoadRegType(*DAG.getContext(), SrcVT);
+    EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f32, NElts);
+
+    LoadSDNode *Load = cast<LoadSDNode>(Src);
+    SDValue NewLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegVT,
+                                     Load->getChain(),
+                                     Load->getBasePtr(),
+                                     LoadVT,
+                                     Load->getMemOperand());
+
+    // Make sure successors of the original load stay after it by updating
+    // them to use the new Chain.
+    DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), NewLoad.getValue(1));
+
+    SmallVector<SDValue, 4> Elts;
+    if (RegVT.isVector())
+      DAG.ExtractVectorElements(NewLoad, Elts);
+    else
+      Elts.push_back(NewLoad);
+
+    SmallVector<SDValue, 4> Ops;
+
+    unsigned EltIdx = 0;
+    for (SDValue Elt : Elts) {
+      unsigned ComponentsInElt = std::min(4u, NElts - 4 * EltIdx);
+      for (unsigned I = 0; I < ComponentsInElt; ++I) {
+        unsigned Opc = AMDGPUISD::CVT_F32_UBYTE0 + I;
+        SDValue Cvt = DAG.getNode(Opc, DL, MVT::f32, Elt);
+        DCI.AddToWorklist(Cvt.getNode());
+        Ops.push_back(Cvt);
+      }
+
+      ++EltIdx;
+    }
+
+    assert(Ops.size() == NElts);
+
+    return DAG.getNode(ISD::BUILD_VECTOR, DL, FloatVT, Ops);
+  }
+
+  return SDValue();
+}
+
 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
                                             DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
@@ -1020,6 +1113,31 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
       }
       break;
     }
+
+  case AMDGPUISD::CVT_F32_UBYTE0:
+  case AMDGPUISD::CVT_F32_UBYTE1:
+  case AMDGPUISD::CVT_F32_UBYTE2:
+  case AMDGPUISD::CVT_F32_UBYTE3: {
+    unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
+
+    SDValue Src = N->getOperand(0);
+    APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);
+
+    APInt KnownZero, KnownOne;
+    TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
+                                          !DCI.isBeforeLegalizeOps());
+    const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+    if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
+        TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {
+      DCI.CommitTargetLoweringOpt(TLO);
+    }
+
+    break;
+  }
+
+  case ISD::UINT_TO_FP: {
+    return performUCharToFloatCombine(N, DCI);
+  }
   }
 
   return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
diff --git a/lib/Target/R600/SIISelLowering.h b/lib/Target/R600/SIISelLowering.h
index d9bad4e4ac..2f97a9ada8 100644
--- a/lib/Target/R600/SIISelLowering.h
+++ b/lib/Target/R600/SIISelLowering.h
@@ -43,6 +43,9 @@ class SITargetLowering : public AMDGPUTargetLowering {
   void adjustWritemask(MachineSDNode *&N, SelectionDAG &DAG) const;
   MachineSDNode *AdjustRegClass(MachineSDNode *N, SelectionDAG &DAG) const;
 
+  static SDValue performUCharToFloatCombine(SDNode *N,
+                                            DAGCombinerInfo &DCI);
+
 public:
   SITargetLowering(TargetMachine &tm);
   bool allowsUnalignedMemoryAccesses(EVT VT, unsigned AS,
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 7d37d10567..9af1a70617 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -978,10 +978,18 @@ defm V_CVT_F32_F64 : VOP1_32_64 <0x0000000f, "V_CVT_F32_F64",
 defm V_CVT_F64_F32 : VOP1_64_32 <0x00000010, "V_CVT_F64_F32",
   [(set f64:$dst, (fextend f32:$src0))]
 >;
-//defm V_CVT_F32_UBYTE0 : VOP1_32 <0x00000011, "V_CVT_F32_UBYTE0", []>;
-//defm V_CVT_F32_UBYTE1 : VOP1_32 <0x00000012, "V_CVT_F32_UBYTE1", []>;
-//defm V_CVT_F32_UBYTE2 : VOP1_32 <0x00000013, "V_CVT_F32_UBYTE2", []>;
-//defm V_CVT_F32_UBYTE3 : VOP1_32 <0x00000014, "V_CVT_F32_UBYTE3", []>;
+defm V_CVT_F32_UBYTE0 : VOP1_32 <0x00000011, "V_CVT_F32_UBYTE0",
+  [(set f32:$dst, (AMDGPUcvt_f32_ubyte0 i32:$src0))]
+>;
+defm V_CVT_F32_UBYTE1 : VOP1_32 <0x00000012, "V_CVT_F32_UBYTE1",
+  [(set f32:$dst, (AMDGPUcvt_f32_ubyte1 i32:$src0))]
+>;
+defm V_CVT_F32_UBYTE2 : VOP1_32 <0x00000013, "V_CVT_F32_UBYTE2",
+  [(set f32:$dst, (AMDGPUcvt_f32_ubyte2 i32:$src0))]
+>;
+defm V_CVT_F32_UBYTE3 : VOP1_32 <0x00000014, "V_CVT_F32_UBYTE3",
+  [(set f32:$dst, (AMDGPUcvt_f32_ubyte3 i32:$src0))]
+>;
 defm V_CVT_U32_F64 : VOP1_32_64 <0x00000015, "V_CVT_U32_F64",
   [(set i32:$dst, (fp_to_uint f64:$src0))]
 >;