[WIP][RISCV][CostModel] Add cost model for shuffle

[luke957]

Summary

In order to make auto-vectorizer work efficiently, correct cost estimate(cost model) is needed. This patch is still a WIP, not only as all kinds of TTI::ShuffleKind need to be handled, but also for the related processes(eg. TargetLoweringBase::getTypeLegalizationCost()).

[craig.topper]

llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp

178

Should this consider LMUL? The ALU likely needs more cycles for LMUL 2 than LMUL 1.

llvm/test/Analysis/CostModel/RISCV/shuffle-broadcast.ll

1

Doesn’t this need -riscv-v-vector-bits-min?

[luke957]

llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp

178

Em, I have not much knowledge in this area. Is ALU(as well as cache, pipeline) described in the RISCV spec? Or the design of ALU-like moduls is target independent?

llvm/test/Analysis/CostModel/RISCV/shuffle-broadcast.ll

1

Oh yes, I'll fix this.

llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h

   unsigned getMaxInterleaveFactor(unsigned VF) {
     return ST->getMaxInterleaveFactor();
   }
+
+  InstructionCost getShuffleCost(TTI::ShuffleKind Kind, VectorType *Tp,
+                                 ArrayRef<int> Mask, int Index,
+                                 VectorType *SubTp);
 };
 
 } // end namespace llvm

llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp

       getMemoryOpCost(Opcode, VTy->getElementType(), Alignment, 0, CostKind, I);
   return NumLoads * MemOpCost;
 }
+
+InstructionCost RISCVTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
+                                             VectorType *Tp, ArrayRef<int> Mask,
+                                             int Index, VectorType *SubTp) {
+  if (!ST->hasStdExtV() ||
+      (Tp->getElementType()->isHalfTy() && !ST->hasStdExtZfh()) ||
+      (Tp->getElementType()->isFloatTy() && !ST->hasStdExtF()) ||
+      (Tp->getElementType()->isDoubleTy() && !ST->hasStdExtD()))
+    return BaseT::getShuffleCost(Kind, Tp, Mask, Index, SubTp);
+
+  Kind = improveShuffleKindFromMask(Kind, Mask);
+  // Todo: handle other TTI::ShuffleKind kinds
+  if (Kind == TTI::SK_Broadcast) {
+    return TLI->getTypeLegalizationCost(DL, Tp).first;
+  }
+
+  return BaseT::getShuffleCost(Kind, Tp, Mask, Index, SubTp);
+}

llvm/test/Analysis/CostModel/RISCV/shuffle-broadcast.ll

+; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py
+; RUN: opt < %s -mtriple=riscv64--linux-gnu -mattr=+experimental-v,+f -cost-model -analyze | FileCheck %s
+
+define void @broadcast() {
+; CHECK-LABEL: 'broadcast'
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v7 = shufflevector <2 x i8> undef, <2 x i8> undef, <2 x i32> zeroinitializer
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v8 = shufflevector <4 x i8> undef, <4 x i8> undef, <4 x i32> zeroinitializer
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v9 = shufflevector <8 x i8> undef, <8 x i8> undef, <8 x i32> zeroinitializer
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v10 = shufflevector <16 x i8> undef, <16 x i8> undef, <16 x i32> zeroinitializer
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v11 = shufflevector <2 x i16> undef, <2 x i16> undef, <2 x i32> zeroinitializer
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v12 = shufflevector <4 x i16> undef, <4 x i16> undef, <4 x i32> zeroinitializer
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v13 = shufflevector <8 x i16> undef, <8 x i16> undef, <8 x i32> zeroinitializer
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v14 = shufflevector <2 x i32> undef, <2 x i32> undef, <2 x i32> zeroinitializer
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v15 = shufflevector <4 x i32> undef, <4 x i32> undef, <4 x i32> zeroinitializer
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v16 = shufflevector <2 x float> undef, <2 x float> undef, <2 x i32> zeroinitializer
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v17 = shufflevector <4 x float> undef, <4 x float> undef, <4 x i32> zeroinitializer
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
+;
+  %v7 = shufflevector <2 x i8> undef, <2 x i8>undef, <2 x i32> zeroinitializer
+  %v8 = shufflevector <4 x i8> undef, <4 x i8>undef, <4 x i32> zeroinitializer
+  %v9 = shufflevector <8 x i8> undef, <8 x i8>undef, <8 x i32> zeroinitializer
+  %v10 = shufflevector <16 x i8> undef, <16 x i8>undef, <16 x i32> zeroinitializer
+
+  %v11 = shufflevector <2 x i16> undef, <2 x i16>undef, <2 x i32> zeroinitializer
+  %v12 = shufflevector <4 x i16> undef, <4 x i16>undef, <4 x i32> zeroinitializer
+  %v13 = shufflevector <8 x i16> undef, <8 x i16>undef, <8 x i32> zeroinitializer
+
+  %v14 = shufflevector <2 x i32> undef, <2 x i32>undef, <2 x i32> zeroinitializer
+  %v15 = shufflevector <4 x i32> undef, <4 x i32>undef, <4 x i32> zeroinitializer
+
+  %v16 = shufflevector <2 x float> undef, <2 x float>undef, <2 x i32> zeroinitializer
+  %v17 = shufflevector <4 x float> undef, <4 x float>undef, <4 x i32> zeroinitializer
+
+  ret void
+}