[AArch64] Add better fcmp costs for expanded predicates

Certain fcmp predicates need to be expanded into multiple operations and or'd
together. This adds some more accurate core modelling for them based on the
predicate. Unsupported operations are given the cost of a libcall and the
latency is set to 2 as that seemed to be fairly common between different CPUs.

Author: davemgreen

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

@@ -4315,10 +4315,9 @@ InstructionCost AArch64TTIImpl::getCmpSelInstrCost(
     unsigned Opcode, Type *ValTy, Type *CondTy, CmpInst::Predicate VecPred,
     TTI::TargetCostKind CostKind, TTI::OperandValueInfo Op1Info,
     TTI::OperandValueInfo Op2Info, const Instruction *I) const {
-  int ISD = TLI->InstructionOpcodeToISD(Opcode);
   // We don't lower some vector selects well that are wider than the register
   // width. TODO: Improve this with different cost kinds.
-  if (isa<FixedVectorType>(ValTy) && ISD == ISD::SELECT) {
+  if (isa<FixedVectorType>(ValTy) && Opcode == Instruction::Select) {
     // We would need this many instructions to hide the scalarization happening.
     const int AmortizationCost = 20;
 
@@ -4348,63 +4347,80 @@ InstructionCost AArch64TTIImpl::getCmpSelInstrCost(
         return LT.first;
     }
 
-    static const TypeConversionCostTblEntry
-    VectorSelectTbl[] = {
-      { ISD::SELECT, MVT::v2i1, MVT::v2f32, 2 },
-      { ISD::SELECT, MVT::v2i1, MVT::v2f64, 2 },
-      { ISD::SELECT, MVT::v4i1, MVT::v4f32, 2 },
-      { ISD::SELECT, MVT::v4i1, MVT::v4f16, 2 },
-      { ISD::SELECT, MVT::v8i1, MVT::v8f16, 2 },
-      { ISD::SELECT, MVT::v16i1, MVT::v16i16, 16 },
-      { ISD::SELECT, MVT::v8i1, MVT::v8i32, 8 },
-      { ISD::SELECT, MVT::v16i1, MVT::v16i32, 16 },
-      { ISD::SELECT, MVT::v4i1, MVT::v4i64, 4 * AmortizationCost },
-      { ISD::SELECT, MVT::v8i1, MVT::v8i64, 8 * AmortizationCost },
-      { ISD::SELECT, MVT::v16i1, MVT::v16i64, 16 * AmortizationCost }
-    };
+    static const TypeConversionCostTblEntry VectorSelectTbl[] = {
+        {Instruction::Select, MVT::v2i1, MVT::v2f32, 2},
+        {Instruction::Select, MVT::v2i1, MVT::v2f64, 2},
+        {Instruction::Select, MVT::v4i1, MVT::v4f32, 2},
+        {Instruction::Select, MVT::v4i1, MVT::v4f16, 2},
+        {Instruction::Select, MVT::v8i1, MVT::v8f16, 2},
+        {Instruction::Select, MVT::v16i1, MVT::v16i16, 16},
+        {Instruction::Select, MVT::v8i1, MVT::v8i32, 8},
+        {Instruction::Select, MVT::v16i1, MVT::v16i32, 16},
+        {Instruction::Select, MVT::v4i1, MVT::v4i64, 4 * AmortizationCost},
+        {Instruction::Select, MVT::v8i1, MVT::v8i64, 8 * AmortizationCost},
+        {Instruction::Select, MVT::v16i1, MVT::v16i64, 16 * AmortizationCost}};
 
     EVT SelCondTy = TLI->getValueType(DL, CondTy);
     EVT SelValTy = TLI->getValueType(DL, ValTy);
     if (SelCondTy.isSimple() && SelValTy.isSimple()) {
-      if (const auto *Entry = ConvertCostTableLookup(VectorSelectTbl, ISD,
+      if (const auto *Entry = ConvertCostTableLookup(VectorSelectTbl, Opcode,
                                                      SelCondTy.getSimpleVT(),
                                                      SelValTy.getSimpleVT()))
         return Entry->Cost;
     }
   }
 
-  if (isa<FixedVectorType>(ValTy) && ISD == ISD::SETCC) {
-    Type *ValScalarTy = ValTy->getScalarType();
-    if ((ValScalarTy->isHalfTy() && !ST->hasFullFP16()) ||
-        ValScalarTy->isBFloatTy()) {
-      auto *ValVTy = cast<FixedVectorType>(ValTy);
-
-      // Without dedicated instructions we promote [b]f16 compares to f32.
-      auto *PromotedTy =
-          VectorType::get(Type::getFloatTy(ValTy->getContext()), ValVTy);
-
-      InstructionCost Cost = 0;
-      // Promote operands to float vectors.
-      Cost += 2 * getCastInstrCost(Instruction::FPExt, PromotedTy, ValTy,
-                                   TTI::CastContextHint::None, CostKind);
-      // Compare float vectors.
+  if (Opcode == Instruction::FCmp) {
+    // Without dedicated instructions we promote f16 + bf16 compares to f32.
+    if ((!ST->hasFullFP16() && ValTy->getScalarType()->isHalfTy()) ||
+        ValTy->getScalarType()->isBFloatTy()) {
+      Type *PromotedTy =
+          ValTy->getWithNewType(Type::getFloatTy(ValTy->getContext()));
+      InstructionCost Cost =
+          getCastInstrCost(Instruction::FPExt, PromotedTy, ValTy,
+                           TTI::CastContextHint::None, CostKind);
+      if (!Op1Info.isConstant() && !Op2Info.isConstant())
+        Cost *= 2;
       Cost += getCmpSelInstrCost(Opcode, PromotedTy, CondTy, VecPred, CostKind,
                                  Op1Info, Op2Info);
-      // During codegen we'll truncate the vector result from i32 to i16.
-      Cost +=
-          getCastInstrCost(Instruction::Trunc, VectorType::getInteger(ValVTy),
-                           VectorType::getInteger(PromotedTy),
-                           TTI::CastContextHint::None, CostKind);
+      if (ValTy->isVectorTy())
+        Cost += getCastInstrCost(
+            Instruction::Trunc, VectorType::getInteger(cast<VectorType>(ValTy)),
+            VectorType::getInteger(cast<VectorType>(PromotedTy)),
+            TTI::CastContextHint::None, CostKind);
       return Cost;
     }
+
+    auto LT = getTypeLegalizationCost(ValTy);
+    // Model unknown fp compares as a libcall.
+    if (LT.second.getScalarType() != MVT::f64 &&
+        LT.second.getScalarType() != MVT::f32 &&
+        LT.second.getScalarType() != MVT::f16)
+      return LT.first * getCallInstrCost(/*Function*/ nullptr, ValTy,
+                                         {ValTy, ValTy}, CostKind);
+
+    // Some comparison operators require expanding to multiple compares + or.
+    unsigned Factor = 1;
+    if (!CondTy->isVectorTy() &&
+        (VecPred == FCmpInst::FCMP_ONE || VecPred == FCmpInst::FCMP_UEQ))
+      Factor = 2; // fcmp with 2 selects
+    else if (isa<FixedVectorType>(ValTy) &&
+             (VecPred == FCmpInst::FCMP_ONE || VecPred == FCmpInst::FCMP_UEQ ||
+              VecPred == FCmpInst::FCMP_ORD || VecPred == FCmpInst::FCMP_UNO))
+      Factor = 3; // fcmxx+fcmyy+or
+    else if (isa<ScalableVectorType>(ValTy) &&
+             (VecPred == FCmpInst::FCMP_ONE || VecPred == FCmpInst::FCMP_UEQ))
+      Factor = 3; // fcmxx+fcmyy+or
+
+    return Factor * (CostKind == TTI::TCK_Latency ? 2 : LT.first);
   }
 
   // Treat the icmp in icmp(and, 0) or icmp(and, -1/1) when it can be folded to
   // icmp(and, 0) as free, as we can make use of ands, but only if the
   // comparison is not unsigned. FIXME: Enable for non-throughput cost kinds
   // providing it will not cause performance regressions.
   if (CostKind == TTI::TCK_RecipThroughput && ValTy->isIntegerTy() &&
-      ISD == ISD::SETCC && I && !CmpInst::isUnsigned(VecPred) &&
+      Opcode == Instruction::ICmp && I && !CmpInst::isUnsigned(VecPred) &&
       TLI->isTypeLegal(TLI->getValueType(DL, ValTy)) &&
       match(I->getOperand(0), m_And(m_Value(), m_Value()))) {
     if (match(I->getOperand(1), m_Zero()))

llvm/test/Analysis/CostModel/AArch64/cmp.ll

@@ -13,13 +13,13 @@ define void @cmps() {
 ; CHECK-NEXT:  Cost Model: Found costs of 1 for: %cv16i8 = icmp slt <16 x i8> undef, undef
 ; CHECK-NEXT:  Cost Model: Found costs of 1 for: %cv8i16 = icmp ult <8 x i16> undef, undef
 ; CHECK-NEXT:  Cost Model: Found costs of 1 for: %cv4i32 = icmp sge <4 x i32> undef, undef
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %cf16 = fcmp oge half undef, undef
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %cf32 = fcmp ogt float undef, undef
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %cf64 = fcmp ogt double undef, undef
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %cbf64 = fcmp ogt bfloat undef, undef
+; CHECK-NEXT:  Cost Model: Found costs of RThru:3 CodeSize:3 Lat:4 SizeLat:3 for: %cf16 = fcmp oge half undef, undef
+; CHECK-NEXT:  Cost Model: Found costs of RThru:1 CodeSize:1 Lat:2 SizeLat:1 for: %cf32 = fcmp ogt float undef, undef
+; CHECK-NEXT:  Cost Model: Found costs of RThru:1 CodeSize:1 Lat:2 SizeLat:1 for: %cf64 = fcmp ogt double undef, undef
+; CHECK-NEXT:  Cost Model: Found costs of RThru:3 CodeSize:3 Lat:4 SizeLat:3 for: %cbf64 = fcmp ogt bfloat undef, undef
 ; CHECK-NEXT:  Cost Model: Found costs of RThru:7 CodeSize:5 Lat:5 SizeLat:5 for: %cfv816 = fcmp olt <8 x half> undef, undef
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %cfv432 = fcmp oge <4 x float> undef, undef
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %cfv264 = fcmp oge <2 x double> undef, undef
+; CHECK-NEXT:  Cost Model: Found costs of RThru:1 CodeSize:1 Lat:2 SizeLat:1 for: %cfv432 = fcmp oge <4 x float> undef, undef
+; CHECK-NEXT:  Cost Model: Found costs of RThru:1 CodeSize:1 Lat:2 SizeLat:1 for: %cfv264 = fcmp oge <2 x double> undef, undef
 ; CHECK-NEXT:  Cost Model: Found costs of RThru:7 CodeSize:5 Lat:5 SizeLat:5 for: %cbfv816 = fcmp olt <8 x bfloat> undef, undef
 ; CHECK-NEXT:  Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret void
 ;