[AMDGPU] Support horizontal vectorization.

Author: FarhanaAleen

Reviewed By: rampitec, arsenm

Subscribers: llvm-commits, AMDGPU

Differential Revision: https://reviews.llvm.org/D46213

llvm-svn: 331313

3 files changed, 43 insertions, 0 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
index bf560a94b31..d9bee122ce7 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
@@ -468,6 +468,21 @@ unsigned AMDGPUTTIImpl::getCFInstrCost(unsigned Opcode) {
   }
 }
 
+int AMDGPUTTIImpl::getArithmeticReductionCost(unsigned Opcode, Type *Ty,
+                                              bool IsPairwise) {
+  EVT OrigTy = TLI->getValueType(DL, Ty);
+
+  // Computes cost on targets that have packed math instructions(which support
+  // 16-bit types only).
+  if (IsPairwise ||
+      !ST->hasVOP3PInsts() ||
+      OrigTy.getScalarSizeInBits() != 16)
+    return BaseT::getArithmeticReductionCost(Opcode, Ty, IsPairwise);
+
+  std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
+  return LT.first * getFullRateInstrCost();
+}
+
 int AMDGPUTTIImpl::getVectorInstrCost(unsigned Opcode, Type *ValTy,
                                       unsigned Index) {
   switch (Opcode) {
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h
index a112757173d..744093d038f 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h
@@ -172,6 +172,10 @@ public:
                            const Function *Callee) const;
 
   unsigned getInliningThresholdMultiplier() { return 9; }
+
+  int getArithmeticReductionCost(unsigned Opcode,
+                                 Type *Ty,
+                                 bool IsPairwise);
 };
 
 } // end namespace llvm
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 4186e221ede..1ed6dd963fa 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -6616,6 +6616,30 @@ SDValue SITargetLowering::performExtractVectorEltCombine(
     return DAG.getNode(Vec.getOpcode(), SL, EltVT, Elt);
   }
 
+  // ScalarRes = EXTRACT_VECTOR_ELT ((vector-BINOP Vec1, Vec2), Idx)
+  //    =>
+  // Vec1Elt = EXTRACT_VECTOR_ELT(Vec1, Idx)
+  // Vec2Elt = EXTRACT_VECTOR_ELT(Vec2, Idx)
+  // ScalarRes = scalar-BINOP Vec1Elt, Vec2Elt
+  if (Vec.hasOneUse()) {
+    SDLoc SL(N);
+    EVT EltVT = N->getValueType(0);
+    SDValue Idx = N->getOperand(1);
+    unsigned Opc = Vec.getOpcode();
+
+    switch(Opc) {
+    default:
+      return SDValue();
+      // TODO: Support other binary operations.
+    case ISD::FADD:
+    case ISD::ADD:
+      return DAG.getNode(Opc, SL, EltVT,
+                         DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
+                                     Vec.getOperand(0), Idx),
+                         DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
+                                     Vec.getOperand(1), Idx));
+    }
+  }
   return SDValue();
 }