1 //===- HexagonTargetTransformInfo.cpp - Hexagon specific TTI pass ---------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 /// \file
8 /// This file implements a TargetTransformInfo analysis pass specific to the
9 /// Hexagon target machine. It uses the target's detailed information to provide
10 /// more precise answers to certain TTI queries, while letting the target
11 /// independent and default TTI implementations handle the rest.
12 ///
13 //===----------------------------------------------------------------------===//
14 
15 #include "HexagonTargetTransformInfo.h"
16 #include "HexagonSubtarget.h"
17 #include "llvm/Analysis/TargetTransformInfo.h"
18 #include "llvm/CodeGen/ValueTypes.h"
19 #include "llvm/IR/InstrTypes.h"
20 #include "llvm/IR/Instructions.h"
21 #include "llvm/IR/User.h"
22 #include "llvm/Support/Casting.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Transforms/Utils/UnrollLoop.h"
25 
26 using namespace llvm;
27 
28 #define DEBUG_TYPE "hexagontti"
29 
30 static cl::opt<bool> HexagonAutoHVX("hexagon-autohvx", cl::init(false),
31   cl::Hidden, cl::desc("Enable loop vectorizer for HVX"));
32 
33 static cl::opt<bool> EmitLookupTables("hexagon-emit-lookup-tables",
34   cl::init(true), cl::Hidden,
35   cl::desc("Control lookup table emission on Hexagon target"));
36 
37 // Constant "cost factor" to make floating point operations more expensive
38 // in terms of vectorization cost. This isn't the best way, but it should
39 // do. Ultimately, the cost should use cycles.
40 static const unsigned FloatFactor = 4;
41 
42 bool HexagonTTIImpl::useHVX() const {
43   return ST.useHVXOps() && HexagonAutoHVX;
44 }
45 
46 bool HexagonTTIImpl::isTypeForHVX(Type *VecTy) const {
47   assert(VecTy->isVectorTy());
48   if (isa<ScalableVectorType>(VecTy))
49     return false;
50   // Avoid types like <2 x i32*>.
51   if (!cast<VectorType>(VecTy)->getElementType()->isIntegerTy())
52     return false;
53   EVT VecVT = EVT::getEVT(VecTy);
54   if (!VecVT.isSimple() || VecVT.getSizeInBits() <= 64)
55     return false;
56   if (ST.isHVXVectorType(VecVT.getSimpleVT()))
57     return true;
58   auto Action = TLI.getPreferredVectorAction(VecVT.getSimpleVT());
59   return Action == TargetLoweringBase::TypeWidenVector;
60 }
61 
62 unsigned HexagonTTIImpl::getTypeNumElements(Type *Ty) const {
63   if (auto *VTy = dyn_cast<VectorType>(Ty))
64     return VTy->getNumElements();
65   assert((Ty->isIntegerTy() || Ty->isFloatingPointTy()) &&
66          "Expecting scalar type");
67   return 1;
68 }
69 
70 TargetTransformInfo::PopcntSupportKind
71 HexagonTTIImpl::getPopcntSupport(unsigned IntTyWidthInBit) const {
72   // Return fast hardware support as every input < 64 bits will be promoted
73   // to 64 bits.
74   return TargetTransformInfo::PSK_FastHardware;
75 }
76 
77 // The Hexagon target can unroll loops with run-time trip counts.
78 void HexagonTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
79                                              TTI::UnrollingPreferences &UP) {
80   UP.Runtime = UP.Partial = true;
81   // Only try to peel innermost loops with small runtime trip counts.
82   if (L && L->empty() && canPeel(L) &&
83       SE.getSmallConstantTripCount(L) == 0 &&
84       SE.getSmallConstantMaxTripCount(L) > 0 &&
85       SE.getSmallConstantMaxTripCount(L) <= 5) {
86     UP.PeelCount = 2;
87   }
88 }
89 
90 bool HexagonTTIImpl::shouldFavorPostInc() const {
91   return true;
92 }
93 
94 /// --- Vector TTI begin ---
95 
96 unsigned HexagonTTIImpl::getNumberOfRegisters(bool Vector) const {
97   if (Vector)
98     return useHVX() ? 32 : 0;
99   return 32;
100 }
101 
102 unsigned HexagonTTIImpl::getMaxInterleaveFactor(unsigned VF) {
103   return useHVX() ? 2 : 0;
104 }
105 
106 unsigned HexagonTTIImpl::getRegisterBitWidth(bool Vector) const {
107   return Vector ? getMinVectorRegisterBitWidth() : 32;
108 }
109 
110 unsigned HexagonTTIImpl::getMinVectorRegisterBitWidth() const {
111   return useHVX() ? ST.getVectorLength()*8 : 0;
112 }
113 
114 unsigned HexagonTTIImpl::getMinimumVF(unsigned ElemWidth) const {
115   return (8 * ST.getVectorLength()) / ElemWidth;
116 }
117 
118 unsigned HexagonTTIImpl::getScalarizationOverhead(VectorType *Ty,
119                                                   const APInt &DemandedElts,
120                                                   bool Insert, bool Extract) {
121   return BaseT::getScalarizationOverhead(Ty, DemandedElts, Insert, Extract);
122 }
123 
124 unsigned HexagonTTIImpl::getOperandsScalarizationOverhead(
125       ArrayRef<const Value*> Args, unsigned VF) {
126   return BaseT::getOperandsScalarizationOverhead(Args, VF);
127 }
128 
129 unsigned HexagonTTIImpl::getCallInstrCost(Function *F, Type *RetTy,
130       ArrayRef<Type*> Tys, TTI::TargetCostKind CostKind) {
131   return BaseT::getCallInstrCost(F, RetTy, Tys, CostKind);
132 }
133 
134 unsigned HexagonTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
135                                                ArrayRef<Value *> Args,
136                                                FastMathFlags FMF, unsigned VF,
137                                                TTI::TargetCostKind CostKind,
138                                                const Instruction *I) {
139   return BaseT::getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF, CostKind, I);
140 }
141 
142 unsigned HexagonTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
143                                                ArrayRef<Type *> Tys,
144                                                FastMathFlags FMF,
145                                                unsigned ScalarizationCostPassed,
146                                                TTI::TargetCostKind CostKind,
147                                                const Instruction *I) {
148   if (ID == Intrinsic::bswap) {
149     std::pair<int, MVT> LT = TLI.getTypeLegalizationCost(DL, RetTy);
150     return LT.first + 2;
151   }
152   return BaseT::getIntrinsicInstrCost(ID, RetTy, Tys, FMF,
153                                       ScalarizationCostPassed, CostKind, I);
154 }
155 
156 unsigned HexagonTTIImpl::getAddressComputationCost(Type *Tp,
157       ScalarEvolution *SE, const SCEV *S) {
158   return 0;
159 }
160 
161 unsigned HexagonTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
162                                          MaybeAlign Alignment,
163                                          unsigned AddressSpace,
164                                          TTI::TargetCostKind CostKind,
165                                          const Instruction *I) {
166   assert(Opcode == Instruction::Load || Opcode == Instruction::Store);
167   if (Opcode == Instruction::Store)
168     return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
169                                   CostKind, I);
170 
171   if (Src->isVectorTy()) {
172     VectorType *VecTy = cast<VectorType>(Src);
173     unsigned VecWidth = VecTy->getPrimitiveSizeInBits().getFixedSize();
174     if (useHVX() && isTypeForHVX(VecTy)) {
175       unsigned RegWidth = getRegisterBitWidth(true);
176       assert(RegWidth && "Non-zero vector register width expected");
177       // Cost of HVX loads.
178       if (VecWidth % RegWidth == 0)
179         return VecWidth / RegWidth;
180       // Cost of constructing HVX vector from scalar loads
181       const Align RegAlign(RegWidth / 8);
182       if (!Alignment || *Alignment > RegAlign)
183         Alignment = RegAlign;
184       assert(Alignment);
185       unsigned AlignWidth = 8 * Alignment->value();
186       unsigned NumLoads = alignTo(VecWidth, AlignWidth) / AlignWidth;
187       return 3 * NumLoads;
188     }
189 
190     // Non-HVX vectors.
191     // Add extra cost for floating point types.
192     unsigned Cost =
193         VecTy->getElementType()->isFloatingPointTy() ? FloatFactor : 1;
194 
195     // At this point unspecified alignment is considered as Align(1).
196     const Align BoundAlignment = std::min(Alignment.valueOrOne(), Align(8));
197     unsigned AlignWidth = 8 * BoundAlignment.value();
198     unsigned NumLoads = alignTo(VecWidth, AlignWidth) / AlignWidth;
199     if (Alignment == Align(4) || Alignment == Align(8))
200       return Cost * NumLoads;
201     // Loads of less than 32 bits will need extra inserts to compose a vector.
202     assert(BoundAlignment <= Align(8));
203     unsigned LogA = Log2(BoundAlignment);
204     return (3 - LogA) * Cost * NumLoads;
205   }
206 
207   return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
208                                 CostKind, I);
209 }
210 
211 unsigned HexagonTTIImpl::getMaskedMemoryOpCost(unsigned Opcode,
212       Type *Src, unsigned Alignment, unsigned AddressSpace,
213       TTI::TargetCostKind CostKind) {
214   return BaseT::getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
215                                       CostKind);
216 }
217 
218 unsigned HexagonTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp,
219       int Index, Type *SubTp) {
220   return 1;
221 }
222 
223 unsigned HexagonTTIImpl::getGatherScatterOpCost(
224     unsigned Opcode, Type *DataTy, Value *Ptr, bool VariableMask,
225     unsigned Alignment, TTI::TargetCostKind CostKind,
226     const Instruction *I) {
227   return BaseT::getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
228                                        Alignment, CostKind, I);
229 }
230 
231 unsigned HexagonTTIImpl::getInterleavedMemoryOpCost(unsigned Opcode,
232       Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
233       unsigned Alignment, unsigned AddressSpace,
234       TTI::TargetCostKind CostKind, bool UseMaskForCond,
235       bool UseMaskForGaps) {
236   if (Indices.size() != Factor || UseMaskForCond || UseMaskForGaps)
237     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
238                                              Alignment, AddressSpace,
239                                              CostKind,
240                                              UseMaskForCond, UseMaskForGaps);
241   return getMemoryOpCost(Opcode, VecTy, MaybeAlign(Alignment), AddressSpace,
242                          CostKind);
243 }
244 
245 unsigned HexagonTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
246       Type *CondTy, TTI::TargetCostKind CostKind, const Instruction *I) {
247   if (ValTy->isVectorTy()) {
248     std::pair<int, MVT> LT = TLI.getTypeLegalizationCost(DL, ValTy);
249     if (Opcode == Instruction::FCmp)
250       return LT.first + FloatFactor * getTypeNumElements(ValTy);
251   }
252   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, CostKind, I);
253 }
254 
255 unsigned HexagonTTIImpl::getArithmeticInstrCost(
256     unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
257     TTI::OperandValueKind Opd1Info,
258     TTI::OperandValueKind Opd2Info, TTI::OperandValueProperties Opd1PropInfo,
259     TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args,
260     const Instruction *CxtI) {
261   if (Ty->isVectorTy()) {
262     std::pair<int, MVT> LT = TLI.getTypeLegalizationCost(DL, Ty);
263     if (LT.second.isFloatingPoint())
264       return LT.first + FloatFactor * getTypeNumElements(Ty);
265   }
266   return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Opd1Info, Opd2Info,
267                                        Opd1PropInfo, Opd2PropInfo, Args, CxtI);
268 }
269 
270 unsigned HexagonTTIImpl::getCastInstrCost(unsigned Opcode, Type *DstTy,
271       Type *SrcTy, TTI::TargetCostKind CostKind, const Instruction *I) {
272   if (SrcTy->isFPOrFPVectorTy() || DstTy->isFPOrFPVectorTy()) {
273     unsigned SrcN = SrcTy->isFPOrFPVectorTy() ? getTypeNumElements(SrcTy) : 0;
274     unsigned DstN = DstTy->isFPOrFPVectorTy() ? getTypeNumElements(DstTy) : 0;
275 
276     std::pair<int, MVT> SrcLT = TLI.getTypeLegalizationCost(DL, SrcTy);
277     std::pair<int, MVT> DstLT = TLI.getTypeLegalizationCost(DL, DstTy);
278     return std::max(SrcLT.first, DstLT.first) + FloatFactor * (SrcN + DstN);
279   }
280   return 1;
281 }
282 
283 unsigned HexagonTTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
284       unsigned Index) {
285   Type *ElemTy = Val->isVectorTy() ? cast<VectorType>(Val)->getElementType()
286                                    : Val;
287   if (Opcode == Instruction::InsertElement) {
288     // Need two rotations for non-zero index.
289     unsigned Cost = (Index != 0) ? 2 : 0;
290     if (ElemTy->isIntegerTy(32))
291       return Cost;
292     // If it's not a 32-bit value, there will need to be an extract.
293     return Cost + getVectorInstrCost(Instruction::ExtractElement, Val, Index);
294   }
295 
296   if (Opcode == Instruction::ExtractElement)
297     return 2;
298 
299   return 1;
300 }
301 
302 /// --- Vector TTI end ---
303 
304 unsigned HexagonTTIImpl::getPrefetchDistance() const {
305   return ST.getL1PrefetchDistance();
306 }
307 
308 unsigned HexagonTTIImpl::getCacheLineSize() const {
309   return ST.getL1CacheLineSize();
310 }
311 
312 int
313 HexagonTTIImpl::getUserCost(const User *U,
314                             ArrayRef<const Value *> Operands,
315                             TTI::TargetCostKind CostKind) {
316   auto isCastFoldedIntoLoad = [this](const CastInst *CI) -> bool {
317     if (!CI->isIntegerCast())
318       return false;
319     // Only extensions from an integer type shorter than 32-bit to i32
320     // can be folded into the load.
321     const DataLayout &DL = getDataLayout();
322     unsigned SBW = DL.getTypeSizeInBits(CI->getSrcTy());
323     unsigned DBW = DL.getTypeSizeInBits(CI->getDestTy());
324     if (DBW != 32 || SBW >= DBW)
325       return false;
326 
327     const LoadInst *LI = dyn_cast<const LoadInst>(CI->getOperand(0));
328     // Technically, this code could allow multiple uses of the load, and
329     // check if all the uses are the same extension operation, but this
330     // should be sufficient for most cases.
331     return LI && LI->hasOneUse();
332   };
333 
334   if (const CastInst *CI = dyn_cast<const CastInst>(U))
335     if (isCastFoldedIntoLoad(CI))
336       return TargetTransformInfo::TCC_Free;
337   return BaseT::getUserCost(U, Operands, CostKind);
338 }
339 
340 bool HexagonTTIImpl::shouldBuildLookupTables() const {
341   return EmitLookupTables;
342 }
343