Reintroduce "[AMDGPU][CostModel] Refine cost model for half- and quar…

…ter-rate instructions."

This patch seems to be fine now underlying issue with discards has been fixed.

This reverts commit 1a8972a.

Change-Id: I5e8f647473d1e3e0f0f6c63c36355d707612e616

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

@@ -472,9 +472,50 @@ int GCNTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
     // FIXME: We're having to query the throughput cost so that the basic
     // implementation tries to generate legalize and scalarization costs. Maybe
     // we could hoist the scalarization code here?
-    return BaseT::getArithmeticInstrCost(Opcode, Ty, TTI::TCK_RecipThroughput,
-                                         Opd1Info, Opd2Info, Opd1PropInfo,
-                                         Opd2PropInfo, Args, CxtI);
+    if (CostKind != TTI::TCK_CodeSize)
+      return BaseT::getArithmeticInstrCost(Opcode, Ty, TTI::TCK_RecipThroughput,
+                                           Opd1Info, Opd2Info, Opd1PropInfo,
+                                           Opd2PropInfo, Args, CxtI);
+    // Scalarization
+
+    // Check if any of the operands are vector operands.
+    int ISD = TLI->InstructionOpcodeToISD(Opcode);
+    assert(ISD && "Invalid opcode");
+
+    std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
+
+    bool IsFloat = Ty->isFPOrFPVectorTy();
+    // Assume that floating point arithmetic operations cost twice as much as
+    // integer operations.
+    unsigned OpCost = (IsFloat ? 2 : 1);
+
+    if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
+      // The operation is legal. Assume it costs 1.
+      // TODO: Once we have extract/insert subvector cost we need to use them.
+      return LT.first * OpCost;
+    }
+
+    if (!TLI->isOperationExpand(ISD, LT.second)) {
+      // If the operation is custom lowered, then assume that the code is twice
+      // as expensive.
+      return LT.first * 2 * OpCost;
+    }
+
+    // Else, assume that we need to scalarize this op.
+    // TODO: If one of the types get legalized by splitting, handle this
+    // similarly to what getCastInstrCost() does.
+    if (auto *VTy = dyn_cast<VectorType>(Ty)) {
+      unsigned Num = cast<FixedVectorType>(VTy)->getNumElements();
+      unsigned Cost = getArithmeticInstrCost(
+          Opcode, VTy->getScalarType(), CostKind, Opd1Info, Opd2Info,
+          Opd1PropInfo, Opd2PropInfo, Args, CxtI);
+      // Return the cost of multiple scalar invocation plus the cost of
+      // inserting and extracting the values.
+      return getScalarizationOverhead(VTy, Args) + Num * Cost;
+    }
+
+    // We don't know anything about this scalar instruction.
+    return OpCost;
   }
 
   // Legalize the type.
@@ -493,7 +534,7 @@ int GCNTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
   case ISD::SRL:
   case ISD::SRA:
     if (SLT == MVT::i64)
-      return get64BitInstrCost() * LT.first * NElts;
+      return get64BitInstrCost(CostKind) * LT.first * NElts;
 
     if (ST->has16BitInsts() && SLT == MVT::i16)
       NElts = (NElts + 1) / 2;
@@ -515,7 +556,7 @@ int GCNTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
 
     return LT.first * NElts * getFullRateInstrCost();
   case ISD::MUL: {
-    const int QuarterRateCost = getQuarterRateInstrCost();
+    const int QuarterRateCost = getQuarterRateInstrCost(CostKind);
     if (SLT == MVT::i64) {
       const int FullRateCost = getFullRateInstrCost();
       return (4 * QuarterRateCost + (2 * 2) * FullRateCost) * LT.first * NElts;
@@ -552,7 +593,7 @@ int GCNTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
   case ISD::FADD:
   case ISD::FSUB:
     if (SLT == MVT::f64)
-      return LT.first * NElts * get64BitInstrCost();
+      return LT.first * NElts * get64BitInstrCost(CostKind);
 
     if (ST->has16BitInsts() && SLT == MVT::f16)
       NElts = (NElts + 1) / 2;
@@ -565,7 +606,9 @@ int GCNTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
     // FIXME: frem should be handled separately. The fdiv in it is most of it,
     // but the current lowering is also not entirely correct.
     if (SLT == MVT::f64) {
-      int Cost = 4 * get64BitInstrCost() + 7 * getQuarterRateInstrCost();
+      int Cost = 7 * get64BitInstrCost(CostKind) +
+                 getQuarterRateInstrCost(CostKind) +
+                 3 * getHalfRateInstrCost(CostKind);
       // Add cost of workaround.
       if (!ST->hasUsableDivScaleConditionOutput())
         Cost += 3 * getFullRateInstrCost();
@@ -577,7 +620,7 @@ int GCNTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
       // TODO: This is more complicated, unsafe flags etc.
       if ((SLT == MVT::f32 && !HasFP32Denormals) ||
           (SLT == MVT::f16 && ST->has16BitInsts())) {
-        return LT.first * getQuarterRateInstrCost() * NElts;
+        return LT.first * getQuarterRateInstrCost(CostKind) * NElts;
       }
     }
 
@@ -587,12 +630,15 @@ int GCNTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
       // f32 fmul
       // v_cvt_f16_f32
       // f16 div_fixup
-      int Cost = 4 * getFullRateInstrCost() + 2 * getQuarterRateInstrCost();
+      int Cost =
+          4 * getFullRateInstrCost() + 2 * getQuarterRateInstrCost(CostKind);
       return LT.first * Cost * NElts;
     }
 
     if (SLT == MVT::f32 || SLT == MVT::f16) {
-      int Cost = 7 * getFullRateInstrCost() + 1 * getQuarterRateInstrCost();
+      // 4 more v_cvt_* insts without f16 insts support
+      int Cost = (SLT == MVT::f16 ? 14 : 10) * getFullRateInstrCost() +
+                 1 * getQuarterRateInstrCost(CostKind);
 
       if (!HasFP32Denormals) {
         // FP mode switches.
@@ -642,7 +688,48 @@ int GCNTTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
   Type *RetTy = ICA.getReturnType();
   EVT OrigTy = TLI->getValueType(DL, RetTy);
   if (!OrigTy.isSimple()) {
-    return BaseT::getIntrinsicInstrCost(ICA, CostKind);
+    if (CostKind != TTI::TCK_CodeSize)
+      return BaseT::getIntrinsicInstrCost(ICA, CostKind);
+
+    // TODO: Combine these two logic paths.
+    if (ICA.isTypeBasedOnly())
+      return getTypeBasedIntrinsicInstrCost(ICA, CostKind);
+
+    Type *RetTy = ICA.getReturnType();
+    unsigned VF = ICA.getVectorFactor();
+    unsigned RetVF =
+        (RetTy->isVectorTy() ? cast<FixedVectorType>(RetTy)->getNumElements()
+                             : 1);
+    assert((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type");
+    const IntrinsicInst *I = ICA.getInst();
+    const SmallVectorImpl<const Value *> &Args = ICA.getArgs();
+    FastMathFlags FMF = ICA.getFlags();
+    // Assume that we need to scalarize this intrinsic.
+    SmallVector<Type *, 4> Types;
+    for (const Value *Op : Args) {
+      Type *OpTy = Op->getType();
+      assert(VF == 1 || !OpTy->isVectorTy());
+      Types.push_back(VF == 1 ? OpTy : FixedVectorType::get(OpTy, VF));
+    }
+
+    if (VF > 1 && !RetTy->isVoidTy())
+      RetTy = FixedVectorType::get(RetTy, VF);
+
+    // Compute the scalarization overhead based on Args for a vector
+    // intrinsic. A vectorizer will pass a scalar RetTy and VF > 1, while
+    // CostModel will pass a vector RetTy and VF is 1.
+    unsigned ScalarizationCost = std::numeric_limits<unsigned>::max();
+    if (RetVF > 1 || VF > 1) {
+      ScalarizationCost = 0;
+      if (!RetTy->isVoidTy())
+        ScalarizationCost +=
+            getScalarizationOverhead(cast<VectorType>(RetTy), true, false);
+      ScalarizationCost += getOperandsScalarizationOverhead(Args, VF);
+    }
+
+    IntrinsicCostAttributes Attrs(ICA.getID(), RetTy, Types, FMF,
+                                  ScalarizationCost, I);
+    return getIntrinsicInstrCost(Attrs, CostKind);
   }
 
   // Legalize the type.
@@ -654,16 +741,16 @@ int GCNTTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
   MVT::SimpleValueType SLT = LT.second.getScalarType().SimpleTy;
 
   if (SLT == MVT::f64)
-    return LT.first * NElts * get64BitInstrCost();
+    return LT.first * NElts * get64BitInstrCost(CostKind);
 
   if (ST->has16BitInsts() && SLT == MVT::f16)
     NElts = (NElts + 1) / 2;
 
   // TODO: Get more refined intrinsic costs?
-  unsigned InstRate = getQuarterRateInstrCost();
+  unsigned InstRate = getQuarterRateInstrCost(CostKind);
   if (ICA.getID() == Intrinsic::fma) {
-    InstRate = ST->hasFastFMAF32() ? getHalfRateInstrCost()
-                                   : getQuarterRateInstrCost();
+    InstRate = ST->hasFastFMAF32() ? getHalfRateInstrCost(CostKind)
+                                   : getQuarterRateInstrCost(CostKind);
   }
 
   return LT.first * NElts * InstRate;
@@ -714,7 +801,7 @@ int GCNTTIImpl::getMinMaxReductionCost(VectorType *Ty, VectorType *CondTy,
                                          CostKind);
 
   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
-  return LT.first * getHalfRateInstrCost();
+  return LT.first * getHalfRateInstrCost(CostKind);
 }
 
 int GCNTTIImpl::getVectorInstrCost(unsigned Opcode, Type *ValTy,

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h

@@ -115,21 +115,26 @@ class GCNTTIImpl final : public BasicTTIImplBase<GCNTTIImpl> {
     return TargetTransformInfo::TCC_Basic;
   }
 
-  static inline int getHalfRateInstrCost() {
-    return 2 * TargetTransformInfo::TCC_Basic;
+  static inline int getHalfRateInstrCost(
+      TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput) {
+    return CostKind == TTI::TCK_CodeSize ? 2
+                                         : 2 * TargetTransformInfo::TCC_Basic;
   }
 
   // TODO: The size is usually 8 bytes, but takes 4x as many cycles. Maybe
   // should be 2 or 4.
-  static inline int getQuarterRateInstrCost() {
-    return 3 * TargetTransformInfo::TCC_Basic;
+  static inline int getQuarterRateInstrCost(
+      TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput) {
+    return CostKind == TTI::TCK_CodeSize ? 2
+                                         : 4 * TargetTransformInfo::TCC_Basic;
   }
 
-   // On some parts, normal fp64 operations are half rate, and others
-   // quarter. This also applies to some integer operations.
-  inline int get64BitInstrCost() const {
-    return ST->hasHalfRate64Ops() ?
-      getHalfRateInstrCost() : getQuarterRateInstrCost();
+  // On some parts, normal fp64 operations are half rate, and others
+  // quarter. This also applies to some integer operations.
+  inline int get64BitInstrCost(
+      TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput) const {
+    return ST->hasHalfRate64Ops() ? getHalfRateInstrCost(CostKind)
+                                  : getQuarterRateInstrCost(CostKind);
   }
 
 public:

llvm/test/Analysis/CostModel/AMDGPU/fadd.ll

@@ -1,9 +1,9 @@
 ; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900  -mattr=+half-rate-64-ops < %s | FileCheck -check-prefixes=FASTF64,FASTF16,ALL %s
 ; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=-half-rate-64-ops < %s | FileCheck -check-prefixes=SLOWF64,SLOWF16,ALL %s
-; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 -mattr=+half-rate-64-ops < %s | FileCheck -check-prefixes=FASTF64,FASTF16,ALL %s
-; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=-half-rate-64-ops < %s | FileCheck -check-prefixes=SLOWF64,SLOWF16,ALL %s
+; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900  -mattr=+half-rate-64-ops < %s | FileCheck -check-prefixes=FASTF16,SIZEALL,ALL %s
+; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=-half-rate-64-ops < %s | FileCheck -check-prefixes=SLOWF16,SIZEALL,ALL %s
 
-; ALL: 'fadd_f32'
+; ALL-LABEL: 'fadd_f32'
 ; ALL: estimated cost of 1 for {{.*}} fadd float
 define amdgpu_kernel void @fadd_f32(float addrspace(1)* %out, float addrspace(1)* %vaddr, float %b) #0 {
   %vec = load float, float addrspace(1)* %vaddr
@@ -12,7 +12,7 @@ define amdgpu_kernel void @fadd_f32(float addrspace(1)* %out, float addrspace(1)
   ret void
 }
 
-; ALL: 'fadd_v2f32'
+; ALL-LABEL: 'fadd_v2f32'
 ; ALL: estimated cost of 2 for {{.*}} fadd <2 x float>
 define amdgpu_kernel void @fadd_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %vaddr, <2 x float> %b) #0 {
   %vec = load <2 x float>, <2 x float> addrspace(1)* %vaddr
@@ -21,59 +21,58 @@ define amdgpu_kernel void @fadd_v2f32(<2 x float> addrspace(1)* %out, <2 x float
   ret void
 }
 
-; ALL: 'fadd_v3f32'
-; Allow for 4 when v3f32 is illegal and TargetLowering thinks it needs widening,
-; and 3 when it is legal.
-; ALL: estimated cost of {{[34]}} for {{.*}} fadd <3 x float>
+; ALL-LABEL: 'fadd_v3f32'
+; ALL: estimated cost of 3 for {{.*}} fadd <3 x float>
 define amdgpu_kernel void @fadd_v3f32(<3 x float> addrspace(1)* %out, <3 x float> addrspace(1)* %vaddr, <3 x float> %b) #0 {
   %vec = load <3 x float>, <3 x float> addrspace(1)* %vaddr
   %add = fadd <3 x float> %vec, %b
   store <3 x float> %add, <3 x float> addrspace(1)* %out
   ret void
 }
 
-; ALL: 'fadd_v5f32'
-; Allow for 8 when v5f32 is illegal and TargetLowering thinks it needs widening,
-; and 5 when it is legal.
-; ALL: estimated cost of {{[58]}} for {{.*}} fadd <5 x float>
+; ALL-LABEL: 'fadd_v5f32'
+; ALL: estimated cost of 5 for {{.*}} fadd <5 x float>
 define amdgpu_kernel void @fadd_v5f32(<5 x float> addrspace(1)* %out, <5 x float> addrspace(1)* %vaddr, <5 x float> %b) #0 {
   %vec = load <5 x float>, <5 x float> addrspace(1)* %vaddr
   %add = fadd <5 x float> %vec, %b
   store <5 x float> %add, <5 x float> addrspace(1)* %out
   ret void
 }
 
-; ALL: 'fadd_f64'
+; ALL-LABEL: 'fadd_f64'
 ; FASTF64: estimated cost of 2 for {{.*}} fadd double
-; SLOWF64: estimated cost of 3 for {{.*}} fadd double
+; SLOWF64: estimated cost of 4 for {{.*}} fadd double
+; SIZEALL: estimated cost of 2 for {{.*}} fadd double
 define amdgpu_kernel void @fadd_f64(double addrspace(1)* %out, double addrspace(1)* %vaddr, double %b) #0 {
   %vec = load double, double addrspace(1)* %vaddr
   %add = fadd double %vec, %b
   store double %add, double addrspace(1)* %out
   ret void
 }
 
-; ALL: 'fadd_v2f64'
+; ALL-LABEL: 'fadd_v2f64'
 ; FASTF64: estimated cost of 4 for {{.*}} fadd <2 x double>
-; SLOWF64: estimated cost of 6 for {{.*}} fadd <2 x double>
+; SLOWF64: estimated cost of 8 for {{.*}} fadd <2 x double>
+; SIZEALL: estimated cost of 4 for {{.*}} fadd <2 x double>
 define amdgpu_kernel void @fadd_v2f64(<2 x double> addrspace(1)* %out, <2 x double> addrspace(1)* %vaddr, <2 x double> %b) #0 {
   %vec = load <2 x double>, <2 x double> addrspace(1)* %vaddr
   %add = fadd <2 x double> %vec, %b
   store <2 x double> %add, <2 x double> addrspace(1)* %out
   ret void
 }
 
-; ALL: 'fadd_v3f64'
+; ALL-LABEL: 'fadd_v3f64'
 ; FASTF64: estimated cost of 6 for {{.*}} fadd <3 x double>
-; SLOWF64: estimated cost of 9 for {{.*}} fadd <3 x double>
+; SLOWF64: estimated cost of 12 for {{.*}} fadd <3 x double>
+; SIZEALL: estimated cost of 6 for {{.*}} fadd <3 x double>
 define amdgpu_kernel void @fadd_v3f64(<3 x double> addrspace(1)* %out, <3 x double> addrspace(1)* %vaddr, <3 x double> %b) #0 {
   %vec = load <3 x double>, <3 x double> addrspace(1)* %vaddr
   %add = fadd <3 x double> %vec, %b
   store <3 x double> %add, <3 x double> addrspace(1)* %out
   ret void
 }
 
-; ALL: 'fadd_f16'
+; ALL-LABEL: 'fadd_f16'
 ; ALL: estimated cost of 1 for {{.*}} fadd half
 define amdgpu_kernel void @fadd_f16(half addrspace(1)* %out, half addrspace(1)* %vaddr, half %b) #0 {
   %vec = load half, half addrspace(1)* %vaddr
@@ -82,7 +81,7 @@ define amdgpu_kernel void @fadd_f16(half addrspace(1)* %out, half addrspace(1)*
   ret void
 }
 
-; ALL: 'fadd_v2f16'
+; ALL-LABEL: 'fadd_v2f16'
 ; SLOWF16: estimated cost of 2 for {{.*}} fadd <2 x half>
 ; FASTF16: estimated cost of 1 for {{.*}} fadd <2 x half>
 define amdgpu_kernel void @fadd_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %vaddr, <2 x half> %b) #0 {
@@ -92,7 +91,7 @@ define amdgpu_kernel void @fadd_v2f16(<2 x half> addrspace(1)* %out, <2 x half>
   ret void
 }
 
-; ALL: 'fadd_v3f16'
+; ALL-LABEL: 'fadd_v3f16'
 ; SLOWF16: estimated cost of 4 for {{.*}} fadd <3 x half>
 ; FASTF16: estimated cost of 2 for {{.*}} fadd <3 x half>
 define amdgpu_kernel void @fadd_v3f16(<3 x half> addrspace(1)* %out, <3 x half> addrspace(1)* %vaddr, <3 x half> %b) #0 {
@@ -102,7 +101,7 @@ define amdgpu_kernel void @fadd_v3f16(<3 x half> addrspace(1)* %out, <3 x half>
   ret void
 }
 
-; ALL: 'fadd_v4f16'
+; ALL-LABEL: 'fadd_v4f16'
 ; SLOWF16: estimated cost of 4 for {{.*}} fadd <4 x half>
 ; FASTF16: estimated cost of 2 for {{.*}} fadd <4 x half>
 define amdgpu_kernel void @fadd_v4f16(<4 x half> addrspace(1)* %out, <4 x half> addrspace(1)* %vaddr, <4 x half> %b) #0 {