Bug Summary

File:include/llvm/Analysis/TargetTransformInfoImpl.h
Warning:line 77, column 25
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name HexagonTargetTransformInfo.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-8/lib/clang/8.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/lib/Target/Hexagon -I /build/llvm-toolchain-snapshot-8~svn350071/lib/Target/Hexagon -I /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/include -I /build/llvm-toolchain-snapshot-8~svn350071/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/include/clang/8.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-8/lib/clang/8.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/lib/Target/Hexagon -fdebug-prefix-map=/build/llvm-toolchain-snapshot-8~svn350071=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-12-27-042839-1215-1 -x c++ /build/llvm-toolchain-snapshot-8~svn350071/lib/Target/Hexagon/HexagonTargetTransformInfo.cpp -faddrsig

/build/llvm-toolchain-snapshot-8~svn350071/lib/Target/Hexagon/HexagonTargetTransformInfo.cpp

1//===- HexagonTargetTransformInfo.cpp - Hexagon specific TTI pass ---------===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8/// \file
9/// This file implements a TargetTransformInfo analysis pass specific to the
10/// Hexagon target machine. It uses the target's detailed information to provide
11/// more precise answers to certain TTI queries, while letting the target
12/// independent and default TTI implementations handle the rest.
13///
14//===----------------------------------------------------------------------===//
15
16#include "HexagonTargetTransformInfo.h"
17#include "HexagonSubtarget.h"
18#include "llvm/Analysis/TargetTransformInfo.h"
19#include "llvm/CodeGen/ValueTypes.h"
20#include "llvm/IR/InstrTypes.h"
21#include "llvm/IR/Instructions.h"
22#include "llvm/IR/User.h"
23#include "llvm/Support/Casting.h"
24#include "llvm/Support/CommandLine.h"
25#include "llvm/Transforms/Utils/UnrollLoop.h"
26
27using namespace llvm;
28
29#define DEBUG_TYPE"hexagontti" "hexagontti"
30
31static cl::opt<bool> HexagonAutoHVX("hexagon-autohvx", cl::init(false),
32 cl::Hidden, cl::desc("Enable loop vectorizer for HVX"));
33
34static cl::opt<bool> EmitLookupTables("hexagon-emit-lookup-tables",
35 cl::init(true), cl::Hidden,
36 cl::desc("Control lookup table emission on Hexagon target"));
37
38// Constant "cost factor" to make floating point operations more expensive
39// in terms of vectorization cost. This isn't the best way, but it should
40// do. Ultimately, the cost should use cycles.
41static const unsigned FloatFactor = 4;
42
43bool HexagonTTIImpl::useHVX() const {
44 return ST.useHVXOps() && HexagonAutoHVX;
45}
46
47bool HexagonTTIImpl::isTypeForHVX(Type *VecTy) const {
48 assert(VecTy->isVectorTy())((VecTy->isVectorTy()) ? static_cast<void> (0) : __assert_fail
("VecTy->isVectorTy()", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Target/Hexagon/HexagonTargetTransformInfo.cpp"
, 48, __PRETTY_FUNCTION__))
;
49 // Avoid types like <2 x i32*>.
50 if (!cast<VectorType>(VecTy)->getElementType()->isIntegerTy())
51 return false;
52 EVT VecVT = EVT::getEVT(VecTy);
53 if (!VecVT.isSimple() || VecVT.getSizeInBits() <= 64)
54 return false;
55 if (ST.isHVXVectorType(VecVT.getSimpleVT()))
56 return true;
57 auto Action = TLI.getPreferredVectorAction(VecVT.getSimpleVT());
58 return Action == TargetLoweringBase::TypeWidenVector;
59}
60
61unsigned HexagonTTIImpl::getTypeNumElements(Type *Ty) const {
62 if (Ty->isVectorTy())
63 return Ty->getVectorNumElements();
64 assert((Ty->isIntegerTy() || Ty->isFloatingPointTy()) &&(((Ty->isIntegerTy() || Ty->isFloatingPointTy()) &&
"Expecting scalar type") ? static_cast<void> (0) : __assert_fail
("(Ty->isIntegerTy() || Ty->isFloatingPointTy()) && \"Expecting scalar type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Target/Hexagon/HexagonTargetTransformInfo.cpp"
, 65, __PRETTY_FUNCTION__))
65 "Expecting scalar type")(((Ty->isIntegerTy() || Ty->isFloatingPointTy()) &&
"Expecting scalar type") ? static_cast<void> (0) : __assert_fail
("(Ty->isIntegerTy() || Ty->isFloatingPointTy()) && \"Expecting scalar type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Target/Hexagon/HexagonTargetTransformInfo.cpp"
, 65, __PRETTY_FUNCTION__))
;
66 return 1;
67}
68
69TargetTransformInfo::PopcntSupportKind
70HexagonTTIImpl::getPopcntSupport(unsigned IntTyWidthInBit) const {
71 // Return fast hardware support as every input < 64 bits will be promoted
72 // to 64 bits.
73 return TargetTransformInfo::PSK_FastHardware;
74}
75
76// The Hexagon target can unroll loops with run-time trip counts.
77void HexagonTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
78 TTI::UnrollingPreferences &UP) {
79 UP.Runtime = UP.Partial = true;
80 // Only try to peel innermost loops with small runtime trip counts.
81 if (L && L->empty() && canPeel(L) &&
82 SE.getSmallConstantTripCount(L) == 0 &&
83 SE.getSmallConstantMaxTripCount(L) > 0 &&
84 SE.getSmallConstantMaxTripCount(L) <= 5) {
85 UP.PeelCount = 2;
86 }
87}
88
89bool HexagonTTIImpl::shouldFavorPostInc() const {
90 return true;
91}
92
93/// --- Vector TTI begin ---
94
95unsigned HexagonTTIImpl::getNumberOfRegisters(bool Vector) const {
96 if (Vector)
97 return useHVX() ? 32 : 0;
98 return 32;
99}
100
101unsigned HexagonTTIImpl::getMaxInterleaveFactor(unsigned VF) {
102 return useHVX() ? 2 : 0;
103}
104
105unsigned HexagonTTIImpl::getRegisterBitWidth(bool Vector) const {
106 return Vector ? getMinVectorRegisterBitWidth() : 32;
107}
108
109unsigned HexagonTTIImpl::getMinVectorRegisterBitWidth() const {
110 return useHVX() ? ST.getVectorLength()*8 : 0;
111}
112
113unsigned HexagonTTIImpl::getMinimumVF(unsigned ElemWidth) const {
114 return (8 * ST.getVectorLength()) / ElemWidth;
115}
116
117unsigned HexagonTTIImpl::getScalarizationOverhead(Type *Ty, bool Insert,
118 bool Extract) {
119 return BaseT::getScalarizationOverhead(Ty, Insert, Extract);
120}
121
122unsigned HexagonTTIImpl::getOperandsScalarizationOverhead(
123 ArrayRef<const Value*> Args, unsigned VF) {
124 return BaseT::getOperandsScalarizationOverhead(Args, VF);
125}
126
127unsigned HexagonTTIImpl::getCallInstrCost(Function *F, Type *RetTy,
128 ArrayRef<Type*> Tys) {
129 return BaseT::getCallInstrCost(F, RetTy, Tys);
130}
131
132unsigned HexagonTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
133 ArrayRef<Value*> Args, FastMathFlags FMF, unsigned VF) {
134 return BaseT::getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF);
135}
136
137unsigned HexagonTTIImpl::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
138 ArrayRef<Type*> Tys, FastMathFlags FMF,
139 unsigned ScalarizationCostPassed) {
140 if (ID == Intrinsic::bswap) {
141 std::pair<int, MVT> LT = TLI.getTypeLegalizationCost(DL, RetTy);
142 return LT.first + 2;
143 }
144 return BaseT::getIntrinsicInstrCost(ID, RetTy, Tys, FMF,
145 ScalarizationCostPassed);
146}
147
148unsigned HexagonTTIImpl::getAddressComputationCost(Type *Tp,
149 ScalarEvolution *SE, const SCEV *S) {
150 return 0;
151}
152
153unsigned HexagonTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
154 unsigned Alignment, unsigned AddressSpace, const Instruction *I) {
155 assert(Opcode == Instruction::Load || Opcode == Instruction::Store)((Opcode == Instruction::Load || Opcode == Instruction::Store
) ? static_cast<void> (0) : __assert_fail ("Opcode == Instruction::Load || Opcode == Instruction::Store"
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Target/Hexagon/HexagonTargetTransformInfo.cpp"
, 155, __PRETTY_FUNCTION__))
;
156 if (Opcode == Instruction::Store)
157 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I);
158
159 if (Src->isVectorTy()) {
160 VectorType *VecTy = cast<VectorType>(Src);
161 unsigned VecWidth = VecTy->getBitWidth();
162 if (useHVX() && isTypeForHVX(VecTy)) {
163 unsigned RegWidth = getRegisterBitWidth(true);
164 Alignment = std::min(Alignment, RegWidth/8);
165 // Cost of HVX loads.
166 if (VecWidth % RegWidth == 0)
167 return VecWidth / RegWidth;
168 // Cost of constructing HVX vector from scalar loads.
169 unsigned AlignWidth = 8 * std::max(1u, Alignment);
170 unsigned NumLoads = alignTo(VecWidth, AlignWidth) / AlignWidth;
171 return 3*NumLoads;
172 }
173
174 // Non-HVX vectors.
175 // Add extra cost for floating point types.
176 unsigned Cost = VecTy->getElementType()->isFloatingPointTy() ? FloatFactor
177 : 1;
178 Alignment = std::min(Alignment, 8u);
179 unsigned AlignWidth = 8 * std::max(1u, Alignment);
180 unsigned NumLoads = alignTo(VecWidth, AlignWidth) / AlignWidth;
181 if (Alignment == 4 || Alignment == 8)
182 return Cost * NumLoads;
183 // Loads of less than 32 bits will need extra inserts to compose a vector.
184 unsigned LogA = Log2_32(Alignment);
185 return (3 - LogA) * Cost * NumLoads;
186 }
187
188 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I);
189}
190
191unsigned HexagonTTIImpl::getMaskedMemoryOpCost(unsigned Opcode,
192 Type *Src, unsigned Alignment, unsigned AddressSpace) {
193 return BaseT::getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
194}
195
196unsigned HexagonTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp,
197 int Index, Type *SubTp) {
198 return 1;
199}
200
201unsigned HexagonTTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
202 Value *Ptr, bool VariableMask, unsigned Alignment) {
203 return BaseT::getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
204 Alignment);
205}
206
207unsigned HexagonTTIImpl::getInterleavedMemoryOpCost(unsigned Opcode,
208 Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
209 unsigned Alignment, unsigned AddressSpace, bool UseMaskForCond,
210 bool UseMaskForGaps) {
211 if (Indices.size() != Factor || UseMaskForCond || UseMaskForGaps)
212 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
213 Alignment, AddressSpace,
214 UseMaskForCond, UseMaskForGaps);
215 return getMemoryOpCost(Opcode, VecTy, Alignment, AddressSpace, nullptr);
216}
217
218unsigned HexagonTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
219 Type *CondTy, const Instruction *I) {
220 if (ValTy->isVectorTy()) {
221 std::pair<int, MVT> LT = TLI.getTypeLegalizationCost(DL, ValTy);
222 if (Opcode == Instruction::FCmp)
223 return LT.first + FloatFactor * getTypeNumElements(ValTy);
224 }
225 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, I);
226}
227
228unsigned HexagonTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
229 TTI::OperandValueKind Opd1Info, TTI::OperandValueKind Opd2Info,
230 TTI::OperandValueProperties Opd1PropInfo,
231 TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value*> Args) {
232 if (Ty->isVectorTy()) {
233 std::pair<int, MVT> LT = TLI.getTypeLegalizationCost(DL, Ty);
234 if (LT.second.isFloatingPoint())
235 return LT.first + FloatFactor * getTypeNumElements(Ty);
236 }
237 return BaseT::getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
238 Opd1PropInfo, Opd2PropInfo, Args);
239}
240
241unsigned HexagonTTIImpl::getCastInstrCost(unsigned Opcode, Type *DstTy,
242 Type *SrcTy, const Instruction *I) {
243 if (SrcTy->isFPOrFPVectorTy() || DstTy->isFPOrFPVectorTy()) {
244 unsigned SrcN = SrcTy->isFPOrFPVectorTy() ? getTypeNumElements(SrcTy) : 0;
245 unsigned DstN = DstTy->isFPOrFPVectorTy() ? getTypeNumElements(DstTy) : 0;
246
247 std::pair<int, MVT> SrcLT = TLI.getTypeLegalizationCost(DL, SrcTy);
248 std::pair<int, MVT> DstLT = TLI.getTypeLegalizationCost(DL, DstTy);
249 return std::max(SrcLT.first, DstLT.first) + FloatFactor * (SrcN + DstN);
250 }
251 return 1;
252}
253
254unsigned HexagonTTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
255 unsigned Index) {
256 Type *ElemTy = Val->isVectorTy() ? cast<VectorType>(Val)->getElementType()
257 : Val;
258 if (Opcode == Instruction::InsertElement) {
259 // Need two rotations for non-zero index.
260 unsigned Cost = (Index != 0) ? 2 : 0;
261 if (ElemTy->isIntegerTy(32))
262 return Cost;
263 // If it's not a 32-bit value, there will need to be an extract.
264 return Cost + getVectorInstrCost(Instruction::ExtractElement, Val, Index);
265 }
266
267 if (Opcode == Instruction::ExtractElement)
268 return 2;
269
270 return 1;
271}
272
273/// --- Vector TTI end ---
274
275unsigned HexagonTTIImpl::getPrefetchDistance() const {
276 return ST.getL1PrefetchDistance();
277}
278
279unsigned HexagonTTIImpl::getCacheLineSize() const {
280 return ST.getL1CacheLineSize();
281}
282
283int HexagonTTIImpl::getUserCost(const User *U,
284 ArrayRef<const Value *> Operands) {
285 auto isCastFoldedIntoLoad = [this](const CastInst *CI) -> bool {
286 if (!CI->isIntegerCast())
287 return false;
288 // Only extensions from an integer type shorter than 32-bit to i32
289 // can be folded into the load.
290 const DataLayout &DL = getDataLayout();
291 unsigned SBW = DL.getTypeSizeInBits(CI->getSrcTy());
292 unsigned DBW = DL.getTypeSizeInBits(CI->getDestTy());
293 if (DBW != 32 || SBW >= DBW)
294 return false;
295
296 const LoadInst *LI = dyn_cast<const LoadInst>(CI->getOperand(0));
297 // Technically, this code could allow multiple uses of the load, and
298 // check if all the uses are the same extension operation, but this
299 // should be sufficient for most cases.
300 return LI && LI->hasOneUse();
301 };
302
303 if (const CastInst *CI = dyn_cast<const CastInst>(U))
1
Taking false branch
304 if (isCastFoldedIntoLoad(CI))
305 return TargetTransformInfo::TCC_Free;
306 return BaseT::getUserCost(U, Operands);
2
Calling 'TargetTransformInfoImplCRTPBase::getUserCost'
307}
308
309bool HexagonTTIImpl::shouldBuildLookupTables() const {
310 return EmitLookupTables;
311}

/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/Analysis/TargetTransformInfoImpl.h

1//===- TargetTransformInfoImpl.h --------------------------------*- C++ -*-===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9/// \file
10/// This file provides helpers for the implementation of
11/// a TargetTransformInfo-conforming class.
12///
13//===----------------------------------------------------------------------===//
14
15#ifndef LLVM_ANALYSIS_TARGETTRANSFORMINFOIMPL_H
16#define LLVM_ANALYSIS_TARGETTRANSFORMINFOIMPL_H
17
18#include "llvm/Analysis/ScalarEvolutionExpressions.h"
19#include "llvm/Analysis/TargetTransformInfo.h"
20#include "llvm/Analysis/VectorUtils.h"
21#include "llvm/IR/CallSite.h"
22#include "llvm/IR/DataLayout.h"
23#include "llvm/IR/Function.h"
24#include "llvm/IR/GetElementPtrTypeIterator.h"
25#include "llvm/IR/Operator.h"
26#include "llvm/IR/Type.h"
27
28namespace llvm {
29
30/// Base class for use as a mix-in that aids implementing
31/// a TargetTransformInfo-compatible class.
32class TargetTransformInfoImplBase {
33protected:
34 typedef TargetTransformInfo TTI;
35
36 const DataLayout &DL;
37
38 explicit TargetTransformInfoImplBase(const DataLayout &DL) : DL(DL) {}
39
40public:
41 // Provide value semantics. MSVC requires that we spell all of these out.
42 TargetTransformInfoImplBase(const TargetTransformInfoImplBase &Arg)
43 : DL(Arg.DL) {}
44 TargetTransformInfoImplBase(TargetTransformInfoImplBase &&Arg) : DL(Arg.DL) {}
45
46 const DataLayout &getDataLayout() const { return DL; }
47
48 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
49 switch (Opcode) {
16
Control jumps to 'case IntToPtr:' at line 74
50 default:
51 // By default, just classify everything as 'basic'.
52 return TTI::TCC_Basic;
53
54 case Instruction::GetElementPtr:
55 llvm_unreachable("Use getGEPCost for GEP operations!")::llvm::llvm_unreachable_internal("Use getGEPCost for GEP operations!"
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 55)
;
56
57 case Instruction::BitCast:
58 assert(OpTy && "Cast instructions must provide the operand type")((OpTy && "Cast instructions must provide the operand type"
) ? static_cast<void> (0) : __assert_fail ("OpTy && \"Cast instructions must provide the operand type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 58, __PRETTY_FUNCTION__))
;
59 if (Ty == OpTy || (Ty->isPointerTy() && OpTy->isPointerTy()))
60 // Identity and pointer-to-pointer casts are free.
61 return TTI::TCC_Free;
62
63 // Otherwise, the default basic cost is used.
64 return TTI::TCC_Basic;
65
66 case Instruction::FDiv:
67 case Instruction::FRem:
68 case Instruction::SDiv:
69 case Instruction::SRem:
70 case Instruction::UDiv:
71 case Instruction::URem:
72 return TTI::TCC_Expensive;
73
74 case Instruction::IntToPtr: {
75 // An inttoptr cast is free so long as the input is a legal integer type
76 // which doesn't contain values outside the range of a pointer.
77 unsigned OpSize = OpTy->getScalarSizeInBits();
17
Called C++ object pointer is null
78 if (DL.isLegalInteger(OpSize) &&
79 OpSize <= DL.getPointerTypeSizeInBits(Ty))
80 return TTI::TCC_Free;
81
82 // Otherwise it's not a no-op.
83 return TTI::TCC_Basic;
84 }
85 case Instruction::PtrToInt: {
86 // A ptrtoint cast is free so long as the result is large enough to store
87 // the pointer, and a legal integer type.
88 unsigned DestSize = Ty->getScalarSizeInBits();
89 if (DL.isLegalInteger(DestSize) &&
90 DestSize >= DL.getPointerTypeSizeInBits(OpTy))
91 return TTI::TCC_Free;
92
93 // Otherwise it's not a no-op.
94 return TTI::TCC_Basic;
95 }
96 case Instruction::Trunc:
97 // trunc to a native type is free (assuming the target has compare and
98 // shift-right of the same width).
99 if (DL.isLegalInteger(DL.getTypeSizeInBits(Ty)))
100 return TTI::TCC_Free;
101
102 return TTI::TCC_Basic;
103 }
104 }
105
106 int getGEPCost(Type *PointeeType, const Value *Ptr,
107 ArrayRef<const Value *> Operands) {
108 // In the basic model, we just assume that all-constant GEPs will be folded
109 // into their uses via addressing modes.
110 for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx)
111 if (!isa<Constant>(Operands[Idx]))
112 return TTI::TCC_Basic;
113
114 return TTI::TCC_Free;
115 }
116
117 unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
118 unsigned &JTSize) {
119 JTSize = 0;
120 return SI.getNumCases();
121 }
122
123 int getExtCost(const Instruction *I, const Value *Src) {
124 return TTI::TCC_Basic;
125 }
126
127 unsigned getCallCost(FunctionType *FTy, int NumArgs) {
128 assert(FTy && "FunctionType must be provided to this routine.")((FTy && "FunctionType must be provided to this routine."
) ? static_cast<void> (0) : __assert_fail ("FTy && \"FunctionType must be provided to this routine.\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 128, __PRETTY_FUNCTION__))
;
129
130 // The target-independent implementation just measures the size of the
131 // function by approximating that each argument will take on average one
132 // instruction to prepare.
133
134 if (NumArgs < 0)
135 // Set the argument number to the number of explicit arguments in the
136 // function.
137 NumArgs = FTy->getNumParams();
138
139 return TTI::TCC_Basic * (NumArgs + 1);
140 }
141
142 unsigned getInliningThresholdMultiplier() { return 1; }
143
144 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
145 ArrayRef<Type *> ParamTys) {
146 switch (IID) {
147 default:
148 // Intrinsics rarely (if ever) have normal argument setup constraints.
149 // Model them as having a basic instruction cost.
150 // FIXME: This is wrong for libc intrinsics.
151 return TTI::TCC_Basic;
152
153 case Intrinsic::annotation:
154 case Intrinsic::assume:
155 case Intrinsic::sideeffect:
156 case Intrinsic::dbg_declare:
157 case Intrinsic::dbg_value:
158 case Intrinsic::dbg_label:
159 case Intrinsic::invariant_start:
160 case Intrinsic::invariant_end:
161 case Intrinsic::launder_invariant_group:
162 case Intrinsic::strip_invariant_group:
163 case Intrinsic::is_constant:
164 case Intrinsic::lifetime_start:
165 case Intrinsic::lifetime_end:
166 case Intrinsic::objectsize:
167 case Intrinsic::ptr_annotation:
168 case Intrinsic::var_annotation:
169 case Intrinsic::experimental_gc_result:
170 case Intrinsic::experimental_gc_relocate:
171 case Intrinsic::coro_alloc:
172 case Intrinsic::coro_begin:
173 case Intrinsic::coro_free:
174 case Intrinsic::coro_end:
175 case Intrinsic::coro_frame:
176 case Intrinsic::coro_size:
177 case Intrinsic::coro_suspend:
178 case Intrinsic::coro_param:
179 case Intrinsic::coro_subfn_addr:
180 // These intrinsics don't actually represent code after lowering.
181 return TTI::TCC_Free;
182 }
183 }
184
185 bool hasBranchDivergence() { return false; }
186
187 bool isSourceOfDivergence(const Value *V) { return false; }
188
189 bool isAlwaysUniform(const Value *V) { return false; }
190
191 unsigned getFlatAddressSpace () {
192 return -1;
193 }
194
195 bool isLoweredToCall(const Function *F) {
196 assert(F && "A concrete function must be provided to this routine.")((F && "A concrete function must be provided to this routine."
) ? static_cast<void> (0) : __assert_fail ("F && \"A concrete function must be provided to this routine.\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 196, __PRETTY_FUNCTION__))
;
197
198 // FIXME: These should almost certainly not be handled here, and instead
199 // handled with the help of TLI or the target itself. This was largely
200 // ported from existing analysis heuristics here so that such refactorings
201 // can take place in the future.
202
203 if (F->isIntrinsic())
204 return false;
205
206 if (F->hasLocalLinkage() || !F->hasName())
207 return true;
208
209 StringRef Name = F->getName();
210
211 // These will all likely lower to a single selection DAG node.
212 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
213 Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" ||
214 Name == "fmin" || Name == "fminf" || Name == "fminl" ||
215 Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" ||
216 Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" ||
217 Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl")
218 return false;
219
220 // These are all likely to be optimized into something smaller.
221 if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" ||
222 Name == "exp2l" || Name == "exp2f" || Name == "floor" ||
223 Name == "floorf" || Name == "ceil" || Name == "round" ||
224 Name == "ffs" || Name == "ffsl" || Name == "abs" || Name == "labs" ||
225 Name == "llabs")
226 return false;
227
228 return true;
229 }
230
231 void getUnrollingPreferences(Loop *, ScalarEvolution &,
232 TTI::UnrollingPreferences &) {}
233
234 bool isLegalAddImmediate(int64_t Imm) { return false; }
235
236 bool isLegalICmpImmediate(int64_t Imm) { return false; }
237
238 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
239 bool HasBaseReg, int64_t Scale,
240 unsigned AddrSpace, Instruction *I = nullptr) {
241 // Guess that only reg and reg+reg addressing is allowed. This heuristic is
242 // taken from the implementation of LSR.
243 return !BaseGV && BaseOffset == 0 && (Scale == 0 || Scale == 1);
244 }
245
246 bool isLSRCostLess(TTI::LSRCost &C1, TTI::LSRCost &C2) {
247 return std::tie(C1.NumRegs, C1.AddRecCost, C1.NumIVMuls, C1.NumBaseAdds,
248 C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
249 std::tie(C2.NumRegs, C2.AddRecCost, C2.NumIVMuls, C2.NumBaseAdds,
250 C2.ScaleCost, C2.ImmCost, C2.SetupCost);
251 }
252
253 bool canMacroFuseCmp() { return false; }
254
255 bool shouldFavorPostInc() const { return false; }
256
257 bool isLegalMaskedStore(Type *DataType) { return false; }
258
259 bool isLegalMaskedLoad(Type *DataType) { return false; }
260
261 bool isLegalMaskedScatter(Type *DataType) { return false; }
262
263 bool isLegalMaskedGather(Type *DataType) { return false; }
264
265 bool hasDivRemOp(Type *DataType, bool IsSigned) { return false; }
266
267 bool hasVolatileVariant(Instruction *I, unsigned AddrSpace) { return false; }
268
269 bool prefersVectorizedAddressing() { return true; }
270
271 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
272 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
273 // Guess that all legal addressing mode are free.
274 if (isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
275 Scale, AddrSpace))
276 return 0;
277 return -1;
278 }
279
280 bool LSRWithInstrQueries() { return false; }
281
282 bool isTruncateFree(Type *Ty1, Type *Ty2) { return false; }
283
284 bool isProfitableToHoist(Instruction *I) { return true; }
285
286 bool useAA() { return false; }
287
288 bool isTypeLegal(Type *Ty) { return false; }
289
290 unsigned getJumpBufAlignment() { return 0; }
291
292 unsigned getJumpBufSize() { return 0; }
293
294 bool shouldBuildLookupTables() { return true; }
295 bool shouldBuildLookupTablesForConstant(Constant *C) { return true; }
296
297 bool useColdCCForColdCall(Function &F) { return false; }
298
299 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
300 return 0;
301 }
302
303 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
304 unsigned VF) { return 0; }
305
306 bool supportsEfficientVectorElementLoadStore() { return false; }
307
308 bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; }
309
310 const TTI::MemCmpExpansionOptions *enableMemCmpExpansion(
311 bool IsZeroCmp) const {
312 return nullptr;
313 }
314
315 bool enableInterleavedAccessVectorization() { return false; }
316
317 bool enableMaskedInterleavedAccessVectorization() { return false; }
318
319 bool isFPVectorizationPotentiallyUnsafe() { return false; }
320
321 bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
322 unsigned BitWidth,
323 unsigned AddressSpace,
324 unsigned Alignment,
325 bool *Fast) { return false; }
326
327 TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) {
328 return TTI::PSK_Software;
329 }
330
331 bool haveFastSqrt(Type *Ty) { return false; }
332
333 bool isFCmpOrdCheaperThanFCmpZero(Type *Ty) { return true; }
334
335 unsigned getFPOpCost(Type *Ty) { return TargetTransformInfo::TCC_Basic; }
336
337 int getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
338 Type *Ty) {
339 return 0;
340 }
341
342 unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; }
343
344 unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
345 Type *Ty) {
346 return TTI::TCC_Free;
347 }
348
349 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
350 Type *Ty) {
351 return TTI::TCC_Free;
352 }
353
354 unsigned getNumberOfRegisters(bool Vector) { return 8; }
355
356 unsigned getRegisterBitWidth(bool Vector) const { return 32; }
357
358 unsigned getMinVectorRegisterBitWidth() { return 128; }
359
360 bool shouldMaximizeVectorBandwidth(bool OptSize) const { return false; }
361
362 unsigned getMinimumVF(unsigned ElemWidth) const { return 0; }
363
364 bool
365 shouldConsiderAddressTypePromotion(const Instruction &I,
366 bool &AllowPromotionWithoutCommonHeader) {
367 AllowPromotionWithoutCommonHeader = false;
368 return false;
369 }
370
371 unsigned getCacheLineSize() { return 0; }
372
373 llvm::Optional<unsigned> getCacheSize(TargetTransformInfo::CacheLevel Level) {
374 switch (Level) {
375 case TargetTransformInfo::CacheLevel::L1D:
376 LLVM_FALLTHROUGH[[clang::fallthrough]];
377 case TargetTransformInfo::CacheLevel::L2D:
378 return llvm::Optional<unsigned>();
379 }
380
381 llvm_unreachable("Unknown TargetTransformInfo::CacheLevel")::llvm::llvm_unreachable_internal("Unknown TargetTransformInfo::CacheLevel"
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 381)
;
382 }
383
384 llvm::Optional<unsigned> getCacheAssociativity(
385 TargetTransformInfo::CacheLevel Level) {
386 switch (Level) {
387 case TargetTransformInfo::CacheLevel::L1D:
388 LLVM_FALLTHROUGH[[clang::fallthrough]];
389 case TargetTransformInfo::CacheLevel::L2D:
390 return llvm::Optional<unsigned>();
391 }
392
393 llvm_unreachable("Unknown TargetTransformInfo::CacheLevel")::llvm::llvm_unreachable_internal("Unknown TargetTransformInfo::CacheLevel"
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 393)
;
394 }
395
396 unsigned getPrefetchDistance() { return 0; }
397
398 unsigned getMinPrefetchStride() { return 1; }
399
400 unsigned getMaxPrefetchIterationsAhead() { return UINT_MAX(2147483647 *2U +1U); }
401
402 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
403
404 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
405 TTI::OperandValueKind Opd1Info,
406 TTI::OperandValueKind Opd2Info,
407 TTI::OperandValueProperties Opd1PropInfo,
408 TTI::OperandValueProperties Opd2PropInfo,
409 ArrayRef<const Value *> Args) {
410 return 1;
411 }
412
413 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Ty, int Index,
414 Type *SubTp) {
415 return 1;
416 }
417
418 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
419 const Instruction *I) { return 1; }
420
421 unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
422 VectorType *VecTy, unsigned Index) {
423 return 1;
424 }
425
426 unsigned getCFInstrCost(unsigned Opcode) { return 1; }
427
428 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
429 const Instruction *I) {
430 return 1;
431 }
432
433 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
434 return 1;
435 }
436
437 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
438 unsigned AddressSpace, const Instruction *I) {
439 return 1;
440 }
441
442 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
443 unsigned AddressSpace) {
444 return 1;
445 }
446
447 unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
448 bool VariableMask,
449 unsigned Alignment) {
450 return 1;
451 }
452
453 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
454 unsigned Factor,
455 ArrayRef<unsigned> Indices,
456 unsigned Alignment, unsigned AddressSpace,
457 bool UseMaskForCond = false,
458 bool UseMaskForGaps = false) {
459 return 1;
460 }
461
462 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
463 ArrayRef<Type *> Tys, FastMathFlags FMF,
464 unsigned ScalarizationCostPassed) {
465 return 1;
466 }
467 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
468 ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) {
469 return 1;
470 }
471
472 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
473 return 1;
474 }
475
476 unsigned getNumberOfParts(Type *Tp) { return 0; }
477
478 unsigned getAddressComputationCost(Type *Tp, ScalarEvolution *,
479 const SCEV *) {
480 return 0;
481 }
482
483 unsigned getArithmeticReductionCost(unsigned, Type *, bool) { return 1; }
484
485 unsigned getMinMaxReductionCost(Type *, Type *, bool, bool) { return 1; }
486
487 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; }
488
489 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) {
490 return false;
491 }
492
493 unsigned getAtomicMemIntrinsicMaxElementSize() const {
494 // Note for overrides: You must ensure for all element unordered-atomic
495 // memory intrinsics that all power-of-2 element sizes up to, and
496 // including, the return value of this method have a corresponding
497 // runtime lib call. These runtime lib call definitions can be found
498 // in RuntimeLibcalls.h
499 return 0;
500 }
501
502 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
503 Type *ExpectedType) {
504 return nullptr;
505 }
506
507 Type *getMemcpyLoopLoweringType(LLVMContext &Context, Value *Length,
508 unsigned SrcAlign, unsigned DestAlign) const {
509 return Type::getInt8Ty(Context);
510 }
511
512 void getMemcpyLoopResidualLoweringType(SmallVectorImpl<Type *> &OpsOut,
513 LLVMContext &Context,
514 unsigned RemainingBytes,
515 unsigned SrcAlign,
516 unsigned DestAlign) const {
517 for (unsigned i = 0; i != RemainingBytes; ++i)
518 OpsOut.push_back(Type::getInt8Ty(Context));
519 }
520
521 bool areInlineCompatible(const Function *Caller,
522 const Function *Callee) const {
523 return (Caller->getFnAttribute("target-cpu") ==
524 Callee->getFnAttribute("target-cpu")) &&
525 (Caller->getFnAttribute("target-features") ==
526 Callee->getFnAttribute("target-features"));
527 }
528
529 bool isIndexedLoadLegal(TTI::MemIndexedMode Mode, Type *Ty,
530 const DataLayout &DL) const {
531 return false;
532 }
533
534 bool isIndexedStoreLegal(TTI::MemIndexedMode Mode, Type *Ty,
535 const DataLayout &DL) const {
536 return false;
537 }
538
539 unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const { return 128; }
540
541 bool isLegalToVectorizeLoad(LoadInst *LI) const { return true; }
542
543 bool isLegalToVectorizeStore(StoreInst *SI) const { return true; }
544
545 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
546 unsigned Alignment,
547 unsigned AddrSpace) const {
548 return true;
549 }
550
551 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
552 unsigned Alignment,
553 unsigned AddrSpace) const {
554 return true;
555 }
556
557 unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
558 unsigned ChainSizeInBytes,
559 VectorType *VecTy) const {
560 return VF;
561 }
562
563 unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
564 unsigned ChainSizeInBytes,
565 VectorType *VecTy) const {
566 return VF;
567 }
568
569 bool useReductionIntrinsic(unsigned Opcode, Type *Ty,
570 TTI::ReductionFlags Flags) const {
571 return false;
572 }
573
574 bool shouldExpandReduction(const IntrinsicInst *II) const {
575 return true;
576 }
577
578protected:
579 // Obtain the minimum required size to hold the value (without the sign)
580 // In case of a vector it returns the min required size for one element.
581 unsigned minRequiredElementSize(const Value* Val, bool &isSigned) {
582 if (isa<ConstantDataVector>(Val) || isa<ConstantVector>(Val)) {
583 const auto* VectorValue = cast<Constant>(Val);
584
585 // In case of a vector need to pick the max between the min
586 // required size for each element
587 auto *VT = cast<VectorType>(Val->getType());
588
589 // Assume unsigned elements
590 isSigned = false;
591
592 // The max required size is the total vector width divided by num
593 // of elements in the vector
594 unsigned MaxRequiredSize = VT->getBitWidth() / VT->getNumElements();
595
596 unsigned MinRequiredSize = 0;
597 for(unsigned i = 0, e = VT->getNumElements(); i < e; ++i) {
598 if (auto* IntElement =
599 dyn_cast<ConstantInt>(VectorValue->getAggregateElement(i))) {
600 bool signedElement = IntElement->getValue().isNegative();
601 // Get the element min required size.
602 unsigned ElementMinRequiredSize =
603 IntElement->getValue().getMinSignedBits() - 1;
604 // In case one element is signed then all the vector is signed.
605 isSigned |= signedElement;
606 // Save the max required bit size between all the elements.
607 MinRequiredSize = std::max(MinRequiredSize, ElementMinRequiredSize);
608 }
609 else {
610 // not an int constant element
611 return MaxRequiredSize;
612 }
613 }
614 return MinRequiredSize;
615 }
616
617 if (const auto* CI = dyn_cast<ConstantInt>(Val)) {
618 isSigned = CI->getValue().isNegative();
619 return CI->getValue().getMinSignedBits() - 1;
620 }
621
622 if (const auto* Cast = dyn_cast<SExtInst>(Val)) {
623 isSigned = true;
624 return Cast->getSrcTy()->getScalarSizeInBits() - 1;
625 }
626
627 if (const auto* Cast = dyn_cast<ZExtInst>(Val)) {
628 isSigned = false;
629 return Cast->getSrcTy()->getScalarSizeInBits();
630 }
631
632 isSigned = false;
633 return Val->getType()->getScalarSizeInBits();
634 }
635
636 bool isStridedAccess(const SCEV *Ptr) {
637 return Ptr && isa<SCEVAddRecExpr>(Ptr);
638 }
639
640 const SCEVConstant *getConstantStrideStep(ScalarEvolution *SE,
641 const SCEV *Ptr) {
642 if (!isStridedAccess(Ptr))
643 return nullptr;
644 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ptr);
645 return dyn_cast<SCEVConstant>(AddRec->getStepRecurrence(*SE));
646 }
647
648 bool isConstantStridedAccessLessThan(ScalarEvolution *SE, const SCEV *Ptr,
649 int64_t MergeDistance) {
650 const SCEVConstant *Step = getConstantStrideStep(SE, Ptr);
651 if (!Step)
652 return false;
653 APInt StrideVal = Step->getAPInt();
654 if (StrideVal.getBitWidth() > 64)
655 return false;
656 // FIXME: Need to take absolute value for negative stride case.
657 return StrideVal.getSExtValue() < MergeDistance;
658 }
659};
660
661/// CRTP base class for use as a mix-in that aids implementing
662/// a TargetTransformInfo-compatible class.
663template <typename T>
664class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase {
665private:
666 typedef TargetTransformInfoImplBase BaseT;
667
668protected:
669 explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {}
670
671public:
672 using BaseT::getCallCost;
673
674 unsigned getCallCost(const Function *F, int NumArgs) {
675 assert(F && "A concrete function must be provided to this routine.")((F && "A concrete function must be provided to this routine."
) ? static_cast<void> (0) : __assert_fail ("F && \"A concrete function must be provided to this routine.\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 675, __PRETTY_FUNCTION__))
;
676
677 if (NumArgs < 0)
678 // Set the argument number to the number of explicit arguments in the
679 // function.
680 NumArgs = F->arg_size();
681
682 if (Intrinsic::ID IID = F->getIntrinsicID()) {
683 FunctionType *FTy = F->getFunctionType();
684 SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
685 return static_cast<T *>(this)
686 ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys);
687 }
688
689 if (!static_cast<T *>(this)->isLoweredToCall(F))
690 return TTI::TCC_Basic; // Give a basic cost if it will be lowered
691 // directly.
692
693 return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs);
694 }
695
696 unsigned getCallCost(const Function *F, ArrayRef<const Value *> Arguments) {
697 // Simply delegate to generic handling of the call.
698 // FIXME: We should use instsimplify or something else to catch calls which
699 // will constant fold with these arguments.
700 return static_cast<T *>(this)->getCallCost(F, Arguments.size());
701 }
702
703 using BaseT::getGEPCost;
704
705 int getGEPCost(Type *PointeeType, const Value *Ptr,
706 ArrayRef<const Value *> Operands) {
707 const GlobalValue *BaseGV = nullptr;
708 if (Ptr != nullptr) {
709 // TODO: will remove this when pointers have an opaque type.
710 assert(Ptr->getType()->getScalarType()->getPointerElementType() ==((Ptr->getType()->getScalarType()->getPointerElementType
() == PointeeType && "explicit pointee type doesn't match operand's pointee type"
) ? static_cast<void> (0) : __assert_fail ("Ptr->getType()->getScalarType()->getPointerElementType() == PointeeType && \"explicit pointee type doesn't match operand's pointee type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 712, __PRETTY_FUNCTION__))
711 PointeeType &&((Ptr->getType()->getScalarType()->getPointerElementType
() == PointeeType && "explicit pointee type doesn't match operand's pointee type"
) ? static_cast<void> (0) : __assert_fail ("Ptr->getType()->getScalarType()->getPointerElementType() == PointeeType && \"explicit pointee type doesn't match operand's pointee type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 712, __PRETTY_FUNCTION__))
712 "explicit pointee type doesn't match operand's pointee type")((Ptr->getType()->getScalarType()->getPointerElementType
() == PointeeType && "explicit pointee type doesn't match operand's pointee type"
) ? static_cast<void> (0) : __assert_fail ("Ptr->getType()->getScalarType()->getPointerElementType() == PointeeType && \"explicit pointee type doesn't match operand's pointee type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 712, __PRETTY_FUNCTION__))
;
713 BaseGV = dyn_cast<GlobalValue>(Ptr->stripPointerCasts());
714 }
715 bool HasBaseReg = (BaseGV == nullptr);
716
717 auto PtrSizeBits = DL.getPointerTypeSizeInBits(Ptr->getType());
718 APInt BaseOffset(PtrSizeBits, 0);
719 int64_t Scale = 0;
720
721 auto GTI = gep_type_begin(PointeeType, Operands);
722 Type *TargetType = nullptr;
723
724 // Handle the case where the GEP instruction has a single operand,
725 // the basis, therefore TargetType is a nullptr.
726 if (Operands.empty())
727 return !BaseGV ? TTI::TCC_Free : TTI::TCC_Basic;
728
729 for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) {
730 TargetType = GTI.getIndexedType();
731 // We assume that the cost of Scalar GEP with constant index and the
732 // cost of Vector GEP with splat constant index are the same.
733 const ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I);
734 if (!ConstIdx)
735 if (auto Splat = getSplatValue(*I))
736 ConstIdx = dyn_cast<ConstantInt>(Splat);
737 if (StructType *STy = GTI.getStructTypeOrNull()) {
738 // For structures the index is always splat or scalar constant
739 assert(ConstIdx && "Unexpected GEP index")((ConstIdx && "Unexpected GEP index") ? static_cast<
void> (0) : __assert_fail ("ConstIdx && \"Unexpected GEP index\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/Analysis/TargetTransformInfoImpl.h"
, 739, __PRETTY_FUNCTION__))
;
740 uint64_t Field = ConstIdx->getZExtValue();
741 BaseOffset += DL.getStructLayout(STy)->getElementOffset(Field);
742 } else {
743 int64_t ElementSize = DL.getTypeAllocSize(GTI.getIndexedType());
744 if (ConstIdx) {
745 BaseOffset +=
746 ConstIdx->getValue().sextOrTrunc(PtrSizeBits) * ElementSize;
747 } else {
748 // Needs scale register.
749 if (Scale != 0)
750 // No addressing mode takes two scale registers.
751 return TTI::TCC_Basic;
752 Scale = ElementSize;
753 }
754 }
755 }
756
757 // Assumes the address space is 0 when Ptr is nullptr.
758 unsigned AS =
759 (Ptr == nullptr ? 0 : Ptr->getType()->getPointerAddressSpace());
760
761 if (static_cast<T *>(this)->isLegalAddressingMode(
762 TargetType, const_cast<GlobalValue *>(BaseGV),
763 BaseOffset.sextOrTrunc(64).getSExtValue(), HasBaseReg, Scale, AS))
764 return TTI::TCC_Free;
765 return TTI::TCC_Basic;
766 }
767
768 using BaseT::getIntrinsicCost;
769
770 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
771 ArrayRef<const Value *> Arguments) {
772 // Delegate to the generic intrinsic handling code. This mostly provides an
773 // opportunity for targets to (for example) special case the cost of
774 // certain intrinsics based on constants used as arguments.
775 SmallVector<Type *, 8> ParamTys;
776 ParamTys.reserve(Arguments.size());
777 for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
778 ParamTys.push_back(Arguments[Idx]->getType());
779 return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys);
780 }
781
782 unsigned getUserCost(const User *U, ArrayRef<const Value *> Operands) {
783 if (isa<PHINode>(U))
3
Taking false branch
784 return TTI::TCC_Free; // Model all PHI nodes as free.
785
786 // Static alloca doesn't generate target instructions.
787 if (auto *A = dyn_cast<AllocaInst>(U))
4
Taking false branch
788 if (A->isStaticAlloca())
789 return TTI::TCC_Free;
790
791 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
5
Taking false branch
792 return static_cast<T *>(this)->getGEPCost(GEP->getSourceElementType(),
793 GEP->getPointerOperand(),
794 Operands.drop_front());
795 }
796
797 if (auto CS = ImmutableCallSite(U)) {
6
Taking false branch
798 const Function *F = CS.getCalledFunction();
799 if (!F) {
800 // Just use the called value type.
801 Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
802 return static_cast<T *>(this)
803 ->getCallCost(cast<FunctionType>(FTy), CS.arg_size());
804 }
805
806 SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end());
807 return static_cast<T *>(this)->getCallCost(F, Arguments);
808 }
809
810 if (const CastInst *CI = dyn_cast<CastInst>(U)) {
7
Taking false branch
811 // Result of a cmp instruction is often extended (to be used by other
812 // cmp instructions, logical or return instructions). These are usually
813 // nop on most sane targets.
814 if (isa<CmpInst>(CI->getOperand(0)))
815 return TTI::TCC_Free;
816 if (isa<SExtInst>(CI) || isa<ZExtInst>(CI) || isa<FPExtInst>(CI))
817 return static_cast<T *>(this)->getExtCost(CI, Operands.back());
818 }
819
820 return static_cast<T *>(this)->getOperationCost(
11
Calling 'BasicTTIImplBase::getOperationCost'
821 Operator::getOpcode(U), U->getType(),
822 U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr);
8
Assuming the condition is false
9
'?' condition is false
10
Passing null pointer value via 3rd parameter 'OpTy'
823 }
824
825 int getInstructionLatency(const Instruction *I) {
826 SmallVector<const Value *, 4> Operands(I->value_op_begin(),
827 I->value_op_end());
828 if (getUserCost(I, Operands) == TTI::TCC_Free)
829 return 0;
830
831 if (isa<LoadInst>(I))
832 return 4;
833
834 Type *DstTy = I->getType();
835
836 // Usually an intrinsic is a simple instruction.
837 // A real function call is much slower.
838 if (auto *CI = dyn_cast<CallInst>(I)) {
839 const Function *F = CI->getCalledFunction();
840 if (!F || static_cast<T *>(this)->isLoweredToCall(F))
841 return 40;
842 // Some intrinsics return a value and a flag, we use the value type
843 // to decide its latency.
844 if (StructType* StructTy = dyn_cast<StructType>(DstTy))
845 DstTy = StructTy->getElementType(0);
846 // Fall through to simple instructions.
847 }
848
849 if (VectorType *VectorTy = dyn_cast<VectorType>(DstTy))
850 DstTy = VectorTy->getElementType();
851 if (DstTy->isFloatingPointTy())
852 return 3;
853
854 return 1;
855 }
856};
857}
858
859#endif

/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h

1//===- BasicTTIImpl.h -------------------------------------------*- C++ -*-===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10/// \file
11/// This file provides a helper that implements much of the TTI interface in
12/// terms of the target-independent code generator and TargetLowering
13/// interfaces.
14//
15//===----------------------------------------------------------------------===//
16
17#ifndef LLVM_CODEGEN_BASICTTIIMPL_H
18#define LLVM_CODEGEN_BASICTTIIMPL_H
19
20#include "llvm/ADT/APInt.h"
21#include "llvm/ADT/ArrayRef.h"
22#include "llvm/ADT/BitVector.h"
23#include "llvm/ADT/SmallPtrSet.h"
24#include "llvm/ADT/SmallVector.h"
25#include "llvm/Analysis/LoopInfo.h"
26#include "llvm/Analysis/TargetTransformInfo.h"
27#include "llvm/Analysis/TargetTransformInfoImpl.h"
28#include "llvm/CodeGen/ISDOpcodes.h"
29#include "llvm/CodeGen/TargetLowering.h"
30#include "llvm/CodeGen/TargetSubtargetInfo.h"
31#include "llvm/CodeGen/ValueTypes.h"
32#include "llvm/IR/BasicBlock.h"
33#include "llvm/IR/CallSite.h"
34#include "llvm/IR/Constant.h"
35#include "llvm/IR/Constants.h"
36#include "llvm/IR/DataLayout.h"
37#include "llvm/IR/DerivedTypes.h"
38#include "llvm/IR/InstrTypes.h"
39#include "llvm/IR/Instruction.h"
40#include "llvm/IR/Instructions.h"
41#include "llvm/IR/Intrinsics.h"
42#include "llvm/IR/Operator.h"
43#include "llvm/IR/Type.h"
44#include "llvm/IR/Value.h"
45#include "llvm/MC/MCSchedule.h"
46#include "llvm/Support/Casting.h"
47#include "llvm/Support/CommandLine.h"
48#include "llvm/Support/ErrorHandling.h"
49#include "llvm/Support/MachineValueType.h"
50#include "llvm/Support/MathExtras.h"
51#include <algorithm>
52#include <cassert>
53#include <cstdint>
54#include <limits>
55#include <utility>
56
57namespace llvm {
58
59class Function;
60class GlobalValue;
61class LLVMContext;
62class ScalarEvolution;
63class SCEV;
64class TargetMachine;
65
66extern cl::opt<unsigned> PartialUnrollingThreshold;
67
68/// Base class which can be used to help build a TTI implementation.
69///
70/// This class provides as much implementation of the TTI interface as is
71/// possible using the target independent parts of the code generator.
72///
73/// In order to subclass it, your class must implement a getST() method to
74/// return the subtarget, and a getTLI() method to return the target lowering.
75/// We need these methods implemented in the derived class so that this class
76/// doesn't have to duplicate storage for them.
77template <typename T>
78class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
79private:
80 using BaseT = TargetTransformInfoImplCRTPBase<T>;
81 using TTI = TargetTransformInfo;
82
83 /// Estimate a cost of Broadcast as an extract and sequence of insert
84 /// operations.
85 unsigned getBroadcastShuffleOverhead(Type *Ty) {
86 assert(Ty->isVectorTy() && "Can only shuffle vectors")((Ty->isVectorTy() && "Can only shuffle vectors") ?
static_cast<void> (0) : __assert_fail ("Ty->isVectorTy() && \"Can only shuffle vectors\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 86, __PRETTY_FUNCTION__))
;
87 unsigned Cost = 0;
88 // Broadcast cost is equal to the cost of extracting the zero'th element
89 // plus the cost of inserting it into every element of the result vector.
90 Cost += static_cast<T *>(this)->getVectorInstrCost(
91 Instruction::ExtractElement, Ty, 0);
92
93 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
94 Cost += static_cast<T *>(this)->getVectorInstrCost(
95 Instruction::InsertElement, Ty, i);
96 }
97 return Cost;
98 }
99
100 /// Estimate a cost of shuffle as a sequence of extract and insert
101 /// operations.
102 unsigned getPermuteShuffleOverhead(Type *Ty) {
103 assert(Ty->isVectorTy() && "Can only shuffle vectors")((Ty->isVectorTy() && "Can only shuffle vectors") ?
static_cast<void> (0) : __assert_fail ("Ty->isVectorTy() && \"Can only shuffle vectors\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 103, __PRETTY_FUNCTION__))
;
104 unsigned Cost = 0;
105 // Shuffle cost is equal to the cost of extracting element from its argument
106 // plus the cost of inserting them onto the result vector.
107
108 // e.g. <4 x float> has a mask of <0,5,2,7> i.e we need to extract from
109 // index 0 of first vector, index 1 of second vector,index 2 of first
110 // vector and finally index 3 of second vector and insert them at index
111 // <0,1,2,3> of result vector.
112 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
113 Cost += static_cast<T *>(this)
114 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
115 Cost += static_cast<T *>(this)
116 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
117 }
118 return Cost;
119 }
120
121 /// Estimate a cost of subvector extraction as a sequence of extract and
122 /// insert operations.
123 unsigned getExtractSubvectorOverhead(Type *Ty, int Index, Type *SubTy) {
124 assert(Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() &&((Ty && Ty->isVectorTy() && SubTy &&
SubTy->isVectorTy() && "Can only extract subvectors from vectors"
) ? static_cast<void> (0) : __assert_fail ("Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() && \"Can only extract subvectors from vectors\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 125, __PRETTY_FUNCTION__))
125 "Can only extract subvectors from vectors")((Ty && Ty->isVectorTy() && SubTy &&
SubTy->isVectorTy() && "Can only extract subvectors from vectors"
) ? static_cast<void> (0) : __assert_fail ("Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() && \"Can only extract subvectors from vectors\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 125, __PRETTY_FUNCTION__))
;
126 int NumSubElts = SubTy->getVectorNumElements();
127 assert((Index + NumSubElts) <= (int)Ty->getVectorNumElements() &&(((Index + NumSubElts) <= (int)Ty->getVectorNumElements
() && "SK_ExtractSubvector index out of range") ? static_cast
<void> (0) : __assert_fail ("(Index + NumSubElts) <= (int)Ty->getVectorNumElements() && \"SK_ExtractSubvector index out of range\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 128, __PRETTY_FUNCTION__))
128 "SK_ExtractSubvector index out of range")(((Index + NumSubElts) <= (int)Ty->getVectorNumElements
() && "SK_ExtractSubvector index out of range") ? static_cast
<void> (0) : __assert_fail ("(Index + NumSubElts) <= (int)Ty->getVectorNumElements() && \"SK_ExtractSubvector index out of range\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 128, __PRETTY_FUNCTION__))
;
129
130 unsigned Cost = 0;
131 // Subvector extraction cost is equal to the cost of extracting element from
132 // the source type plus the cost of inserting them into the result vector
133 // type.
134 for (int i = 0; i != NumSubElts; ++i) {
135 Cost += static_cast<T *>(this)->getVectorInstrCost(
136 Instruction::ExtractElement, Ty, i + Index);
137 Cost += static_cast<T *>(this)->getVectorInstrCost(
138 Instruction::InsertElement, SubTy, i);
139 }
140 return Cost;
141 }
142
143 /// Estimate a cost of subvector insertion as a sequence of extract and
144 /// insert operations.
145 unsigned getInsertSubvectorOverhead(Type *Ty, int Index, Type *SubTy) {
146 assert(Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() &&((Ty && Ty->isVectorTy() && SubTy &&
SubTy->isVectorTy() && "Can only insert subvectors into vectors"
) ? static_cast<void> (0) : __assert_fail ("Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() && \"Can only insert subvectors into vectors\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 147, __PRETTY_FUNCTION__))
147 "Can only insert subvectors into vectors")((Ty && Ty->isVectorTy() && SubTy &&
SubTy->isVectorTy() && "Can only insert subvectors into vectors"
) ? static_cast<void> (0) : __assert_fail ("Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() && \"Can only insert subvectors into vectors\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 147, __PRETTY_FUNCTION__))
;
148 int NumSubElts = SubTy->getVectorNumElements();
149 assert((Index + NumSubElts) <= (int)Ty->getVectorNumElements() &&(((Index + NumSubElts) <= (int)Ty->getVectorNumElements
() && "SK_InsertSubvector index out of range") ? static_cast
<void> (0) : __assert_fail ("(Index + NumSubElts) <= (int)Ty->getVectorNumElements() && \"SK_InsertSubvector index out of range\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 150, __PRETTY_FUNCTION__))
150 "SK_InsertSubvector index out of range")(((Index + NumSubElts) <= (int)Ty->getVectorNumElements
() && "SK_InsertSubvector index out of range") ? static_cast
<void> (0) : __assert_fail ("(Index + NumSubElts) <= (int)Ty->getVectorNumElements() && \"SK_InsertSubvector index out of range\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 150, __PRETTY_FUNCTION__))
;
151
152 unsigned Cost = 0;
153 // Subvector insertion cost is equal to the cost of extracting element from
154 // the source type plus the cost of inserting them into the result vector
155 // type.
156 for (int i = 0; i != NumSubElts; ++i) {
157 Cost += static_cast<T *>(this)->getVectorInstrCost(
158 Instruction::ExtractElement, SubTy, i);
159 Cost += static_cast<T *>(this)->getVectorInstrCost(
160 Instruction::InsertElement, Ty, i + Index);
161 }
162 return Cost;
163 }
164
165 /// Local query method delegates up to T which *must* implement this!
166 const TargetSubtargetInfo *getST() const {
167 return static_cast<const T *>(this)->getST();
168 }
169
170 /// Local query method delegates up to T which *must* implement this!
171 const TargetLoweringBase *getTLI() const {
172 return static_cast<const T *>(this)->getTLI();
173 }
174
175 static ISD::MemIndexedMode getISDIndexedMode(TTI::MemIndexedMode M) {
176 switch (M) {
177 case TTI::MIM_Unindexed:
178 return ISD::UNINDEXED;
179 case TTI::MIM_PreInc:
180 return ISD::PRE_INC;
181 case TTI::MIM_PreDec:
182 return ISD::PRE_DEC;
183 case TTI::MIM_PostInc:
184 return ISD::POST_INC;
185 case TTI::MIM_PostDec:
186 return ISD::POST_DEC;
187 }
188 llvm_unreachable("Unexpected MemIndexedMode")::llvm::llvm_unreachable_internal("Unexpected MemIndexedMode"
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 188)
;
189 }
190
191protected:
192 explicit BasicTTIImplBase(const TargetMachine *TM, const DataLayout &DL)
193 : BaseT(DL) {}
194
195 using TargetTransformInfoImplBase::DL;
196
197public:
198 /// \name Scalar TTI Implementations
199 /// @{
200 bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
201 unsigned BitWidth, unsigned AddressSpace,
202 unsigned Alignment, bool *Fast) const {
203 EVT E = EVT::getIntegerVT(Context, BitWidth);
204 return getTLI()->allowsMisalignedMemoryAccesses(E, AddressSpace, Alignment, Fast);
205 }
206
207 bool hasBranchDivergence() { return false; }
208
209 bool isSourceOfDivergence(const Value *V) { return false; }
210
211 bool isAlwaysUniform(const Value *V) { return false; }
212
213 unsigned getFlatAddressSpace() {
214 // Return an invalid address space.
215 return -1;
216 }
217
218 bool isLegalAddImmediate(int64_t imm) {
219 return getTLI()->isLegalAddImmediate(imm);
220 }
221
222 bool isLegalICmpImmediate(int64_t imm) {
223 return getTLI()->isLegalICmpImmediate(imm);
224 }
225
226 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
227 bool HasBaseReg, int64_t Scale,
228 unsigned AddrSpace, Instruction *I = nullptr) {
229 TargetLoweringBase::AddrMode AM;
230 AM.BaseGV = BaseGV;
231 AM.BaseOffs = BaseOffset;
232 AM.HasBaseReg = HasBaseReg;
233 AM.Scale = Scale;
234 return getTLI()->isLegalAddressingMode(DL, AM, Ty, AddrSpace, I);
235 }
236
237 bool isIndexedLoadLegal(TTI::MemIndexedMode M, Type *Ty,
238 const DataLayout &DL) const {
239 EVT VT = getTLI()->getValueType(DL, Ty);
240 return getTLI()->isIndexedLoadLegal(getISDIndexedMode(M), VT);
241 }
242
243 bool isIndexedStoreLegal(TTI::MemIndexedMode M, Type *Ty,
244 const DataLayout &DL) const {
245 EVT VT = getTLI()->getValueType(DL, Ty);
246 return getTLI()->isIndexedStoreLegal(getISDIndexedMode(M), VT);
247 }
248
249 bool isLSRCostLess(TTI::LSRCost C1, TTI::LSRCost C2) {
250 return TargetTransformInfoImplBase::isLSRCostLess(C1, C2);
251 }
252
253 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
254 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
255 TargetLoweringBase::AddrMode AM;
256 AM.BaseGV = BaseGV;
257 AM.BaseOffs = BaseOffset;
258 AM.HasBaseReg = HasBaseReg;
259 AM.Scale = Scale;
260 return getTLI()->getScalingFactorCost(DL, AM, Ty, AddrSpace);
261 }
262
263 bool isTruncateFree(Type *Ty1, Type *Ty2) {
264 return getTLI()->isTruncateFree(Ty1, Ty2);
265 }
266
267 bool isProfitableToHoist(Instruction *I) {
268 return getTLI()->isProfitableToHoist(I);
269 }
270
271 bool useAA() const { return getST()->useAA(); }
272
273 bool isTypeLegal(Type *Ty) {
274 EVT VT = getTLI()->getValueType(DL, Ty);
275 return getTLI()->isTypeLegal(VT);
276 }
277
278 int getGEPCost(Type *PointeeType, const Value *Ptr,
279 ArrayRef<const Value *> Operands) {
280 return BaseT::getGEPCost(PointeeType, Ptr, Operands);
281 }
282
283 int getExtCost(const Instruction *I, const Value *Src) {
284 if (getTLI()->isExtFree(I))
285 return TargetTransformInfo::TCC_Free;
286
287 if (isa<ZExtInst>(I) || isa<SExtInst>(I))
288 if (const LoadInst *LI = dyn_cast<LoadInst>(Src))
289 if (getTLI()->isExtLoad(LI, I, DL))
290 return TargetTransformInfo::TCC_Free;
291
292 return TargetTransformInfo::TCC_Basic;
293 }
294
295 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
296 ArrayRef<const Value *> Arguments) {
297 return BaseT::getIntrinsicCost(IID, RetTy, Arguments);
298 }
299
300 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
301 ArrayRef<Type *> ParamTys) {
302 if (IID == Intrinsic::cttz) {
303 if (getTLI()->isCheapToSpeculateCttz())
304 return TargetTransformInfo::TCC_Basic;
305 return TargetTransformInfo::TCC_Expensive;
306 }
307
308 if (IID == Intrinsic::ctlz) {
309 if (getTLI()->isCheapToSpeculateCtlz())
310 return TargetTransformInfo::TCC_Basic;
311 return TargetTransformInfo::TCC_Expensive;
312 }
313
314 return BaseT::getIntrinsicCost(IID, RetTy, ParamTys);
315 }
316
317 unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
318 unsigned &JumpTableSize) {
319 /// Try to find the estimated number of clusters. Note that the number of
320 /// clusters identified in this function could be different from the actural
321 /// numbers found in lowering. This function ignore switches that are
322 /// lowered with a mix of jump table / bit test / BTree. This function was
323 /// initially intended to be used when estimating the cost of switch in
324 /// inline cost heuristic, but it's a generic cost model to be used in other
325 /// places (e.g., in loop unrolling).
326 unsigned N = SI.getNumCases();
327 const TargetLoweringBase *TLI = getTLI();
328 const DataLayout &DL = this->getDataLayout();
329
330 JumpTableSize = 0;
331 bool IsJTAllowed = TLI->areJTsAllowed(SI.getParent()->getParent());
332
333 // Early exit if both a jump table and bit test are not allowed.
334 if (N < 1 || (!IsJTAllowed && DL.getIndexSizeInBits(0u) < N))
335 return N;
336
337 APInt MaxCaseVal = SI.case_begin()->getCaseValue()->getValue();
338 APInt MinCaseVal = MaxCaseVal;
339 for (auto CI : SI.cases()) {
340 const APInt &CaseVal = CI.getCaseValue()->getValue();
341 if (CaseVal.sgt(MaxCaseVal))
342 MaxCaseVal = CaseVal;
343 if (CaseVal.slt(MinCaseVal))
344 MinCaseVal = CaseVal;
345 }
346
347 // Check if suitable for a bit test
348 if (N <= DL.getIndexSizeInBits(0u)) {
349 SmallPtrSet<const BasicBlock *, 4> Dests;
350 for (auto I : SI.cases())
351 Dests.insert(I.getCaseSuccessor());
352
353 if (TLI->isSuitableForBitTests(Dests.size(), N, MinCaseVal, MaxCaseVal,
354 DL))
355 return 1;
356 }
357
358 // Check if suitable for a jump table.
359 if (IsJTAllowed) {
360 if (N < 2 || N < TLI->getMinimumJumpTableEntries())
361 return N;
362 uint64_t Range =
363 (MaxCaseVal - MinCaseVal)
364 .getLimitedValue(std::numeric_limits<uint64_t>::max() - 1) + 1;
365 // Check whether a range of clusters is dense enough for a jump table
366 if (TLI->isSuitableForJumpTable(&SI, N, Range)) {
367 JumpTableSize = Range;
368 return 1;
369 }
370 }
371 return N;
372 }
373
374 unsigned getJumpBufAlignment() { return getTLI()->getJumpBufAlignment(); }
375
376 unsigned getJumpBufSize() { return getTLI()->getJumpBufSize(); }
377
378 bool shouldBuildLookupTables() {
379 const TargetLoweringBase *TLI = getTLI();
380 return TLI->isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
381 TLI->isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
382 }
383
384 bool haveFastSqrt(Type *Ty) {
385 const TargetLoweringBase *TLI = getTLI();
386 EVT VT = TLI->getValueType(DL, Ty);
387 return TLI->isTypeLegal(VT) &&
388 TLI->isOperationLegalOrCustom(ISD::FSQRT, VT);
389 }
390
391 bool isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
392 return true;
393 }
394
395 unsigned getFPOpCost(Type *Ty) {
396 // Check whether FADD is available, as a proxy for floating-point in
397 // general.
398 const TargetLoweringBase *TLI = getTLI();
399 EVT VT = TLI->getValueType(DL, Ty);
400 if (TLI->isOperationLegalOrCustomOrPromote(ISD::FADD, VT))
401 return TargetTransformInfo::TCC_Basic;
402 return TargetTransformInfo::TCC_Expensive;
403 }
404
405 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
406 const TargetLoweringBase *TLI = getTLI();
407 switch (Opcode) {
12
Control jumps to the 'default' case at line 408
408 default: break;
13
Execution continues on line 419
409 case Instruction::Trunc:
410 if (TLI->isTruncateFree(OpTy, Ty))
411 return TargetTransformInfo::TCC_Free;
412 return TargetTransformInfo::TCC_Basic;
413 case Instruction::ZExt:
414 if (TLI->isZExtFree(OpTy, Ty))
415 return TargetTransformInfo::TCC_Free;
416 return TargetTransformInfo::TCC_Basic;
417 }
418
419 return BaseT::getOperationCost(Opcode, Ty, OpTy);
14
Passing null pointer value via 3rd parameter 'OpTy'
15
Calling 'TargetTransformInfoImplBase::getOperationCost'
420 }
421
422 unsigned getInliningThresholdMultiplier() { return 1; }
423
424 void getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
425 TTI::UnrollingPreferences &UP) {
426 // This unrolling functionality is target independent, but to provide some
427 // motivation for its intended use, for x86:
428
429 // According to the Intel 64 and IA-32 Architectures Optimization Reference
430 // Manual, Intel Core models and later have a loop stream detector (and
431 // associated uop queue) that can benefit from partial unrolling.
432 // The relevant requirements are:
433 // - The loop must have no more than 4 (8 for Nehalem and later) branches
434 // taken, and none of them may be calls.
435 // - The loop can have no more than 18 (28 for Nehalem and later) uops.
436
437 // According to the Software Optimization Guide for AMD Family 15h
438 // Processors, models 30h-4fh (Steamroller and later) have a loop predictor
439 // and loop buffer which can benefit from partial unrolling.
440 // The relevant requirements are:
441 // - The loop must have fewer than 16 branches
442 // - The loop must have less than 40 uops in all executed loop branches
443
444 // The number of taken branches in a loop is hard to estimate here, and
445 // benchmarking has revealed that it is better not to be conservative when
446 // estimating the branch count. As a result, we'll ignore the branch limits
447 // until someone finds a case where it matters in practice.
448
449 unsigned MaxOps;
450 const TargetSubtargetInfo *ST = getST();
451 if (PartialUnrollingThreshold.getNumOccurrences() > 0)
452 MaxOps = PartialUnrollingThreshold;
453 else if (ST->getSchedModel().LoopMicroOpBufferSize > 0)
454 MaxOps = ST->getSchedModel().LoopMicroOpBufferSize;
455 else
456 return;
457
458 // Scan the loop: don't unroll loops with calls.
459 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E;
460 ++I) {
461 BasicBlock *BB = *I;
462
463 for (BasicBlock::iterator J = BB->begin(), JE = BB->end(); J != JE; ++J)
464 if (isa<CallInst>(J) || isa<InvokeInst>(J)) {
465 ImmutableCallSite CS(&*J);
466 if (const Function *F = CS.getCalledFunction()) {
467 if (!static_cast<T *>(this)->isLoweredToCall(F))
468 continue;
469 }
470
471 return;
472 }
473 }
474
475 // Enable runtime and partial unrolling up to the specified size.
476 // Enable using trip count upper bound to unroll loops.
477 UP.Partial = UP.Runtime = UP.UpperBound = true;
478 UP.PartialThreshold = MaxOps;
479
480 // Avoid unrolling when optimizing for size.
481 UP.OptSizeThreshold = 0;
482 UP.PartialOptSizeThreshold = 0;
483
484 // Set number of instructions optimized when "back edge"
485 // becomes "fall through" to default value of 2.
486 UP.BEInsns = 2;
487 }
488
489 int getInstructionLatency(const Instruction *I) {
490 if (isa<LoadInst>(I))
491 return getST()->getSchedModel().DefaultLoadLatency;
492
493 return BaseT::getInstructionLatency(I);
494 }
495
496 /// @}
497
498 /// \name Vector TTI Implementations
499 /// @{
500
501 unsigned getNumberOfRegisters(bool Vector) { return Vector ? 0 : 1; }
502
503 unsigned getRegisterBitWidth(bool Vector) const { return 32; }
504
505 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
506 /// are set if the result needs to be inserted and/or extracted from vectors.
507 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
508 assert(Ty->isVectorTy() && "Can only scalarize vectors")((Ty->isVectorTy() && "Can only scalarize vectors"
) ? static_cast<void> (0) : __assert_fail ("Ty->isVectorTy() && \"Can only scalarize vectors\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 508, __PRETTY_FUNCTION__))
;
509 unsigned Cost = 0;
510
511 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
512 if (Insert)
513 Cost += static_cast<T *>(this)
514 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
515 if (Extract)
516 Cost += static_cast<T *>(this)
517 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
518 }
519
520 return Cost;
521 }
522
523 /// Estimate the overhead of scalarizing an instructions unique
524 /// non-constant operands. The types of the arguments are ordinarily
525 /// scalar, in which case the costs are multiplied with VF.
526 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
527 unsigned VF) {
528 unsigned Cost = 0;
529 SmallPtrSet<const Value*, 4> UniqueOperands;
530 for (const Value *A : Args) {
531 if (!isa<Constant>(A) && UniqueOperands.insert(A).second) {
532 Type *VecTy = nullptr;
533 if (A->getType()->isVectorTy()) {
534 VecTy = A->getType();
535 // If A is a vector operand, VF should be 1 or correspond to A.
536 assert((VF == 1 || VF == VecTy->getVectorNumElements()) &&(((VF == 1 || VF == VecTy->getVectorNumElements()) &&
"Vector argument does not match VF") ? static_cast<void>
(0) : __assert_fail ("(VF == 1 || VF == VecTy->getVectorNumElements()) && \"Vector argument does not match VF\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 537, __PRETTY_FUNCTION__))
537 "Vector argument does not match VF")(((VF == 1 || VF == VecTy->getVectorNumElements()) &&
"Vector argument does not match VF") ? static_cast<void>
(0) : __assert_fail ("(VF == 1 || VF == VecTy->getVectorNumElements()) && \"Vector argument does not match VF\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 537, __PRETTY_FUNCTION__))
;
538 }
539 else
540 VecTy = VectorType::get(A->getType(), VF);
541
542 Cost += getScalarizationOverhead(VecTy, false, true);
543 }
544 }
545
546 return Cost;
547 }
548
549 unsigned getScalarizationOverhead(Type *VecTy, ArrayRef<const Value *> Args) {
550 assert(VecTy->isVectorTy())((VecTy->isVectorTy()) ? static_cast<void> (0) : __assert_fail
("VecTy->isVectorTy()", "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 550, __PRETTY_FUNCTION__))
;
551
552 unsigned Cost = 0;
553
554 Cost += getScalarizationOverhead(VecTy, true, false);
555 if (!Args.empty())
556 Cost += getOperandsScalarizationOverhead(Args,
557 VecTy->getVectorNumElements());
558 else
559 // When no information on arguments is provided, we add the cost
560 // associated with one argument as a heuristic.
561 Cost += getScalarizationOverhead(VecTy, false, true);
562
563 return Cost;
564 }
565
566 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
567
568 unsigned getArithmeticInstrCost(
569 unsigned Opcode, Type *Ty,
570 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue,
571 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue,
572 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
573 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
574 ArrayRef<const Value *> Args = ArrayRef<const Value *>()) {
575 // Check if any of the operands are vector operands.
576 const TargetLoweringBase *TLI = getTLI();
577 int ISD = TLI->InstructionOpcodeToISD(Opcode);
578 assert(ISD && "Invalid opcode")((ISD && "Invalid opcode") ? static_cast<void> (
0) : __assert_fail ("ISD && \"Invalid opcode\"", "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 578, __PRETTY_FUNCTION__))
;
579
580 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
581
582 bool IsFloat = Ty->isFPOrFPVectorTy();
583 // Assume that floating point arithmetic operations cost twice as much as
584 // integer operations.
585 unsigned OpCost = (IsFloat ? 2 : 1);
586
587 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
588 // The operation is legal. Assume it costs 1.
589 // TODO: Once we have extract/insert subvector cost we need to use them.
590 return LT.first * OpCost;
591 }
592
593 if (!TLI->isOperationExpand(ISD, LT.second)) {
594 // If the operation is custom lowered, then assume that the code is twice
595 // as expensive.
596 return LT.first * 2 * OpCost;
597 }
598
599 // Else, assume that we need to scalarize this op.
600 // TODO: If one of the types get legalized by splitting, handle this
601 // similarly to what getCastInstrCost() does.
602 if (Ty->isVectorTy()) {
603 unsigned Num = Ty->getVectorNumElements();
604 unsigned Cost = static_cast<T *>(this)
605 ->getArithmeticInstrCost(Opcode, Ty->getScalarType());
606 // Return the cost of multiple scalar invocation plus the cost of
607 // inserting and extracting the values.
608 return getScalarizationOverhead(Ty, Args) + Num * Cost;
609 }
610
611 // We don't know anything about this scalar instruction.
612 return OpCost;
613 }
614
615 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
616 Type *SubTp) {
617 switch (Kind) {
618 case TTI::SK_Broadcast:
619 return getBroadcastShuffleOverhead(Tp);
620 case TTI::SK_Select:
621 case TTI::SK_Reverse:
622 case TTI::SK_Transpose:
623 case TTI::SK_PermuteSingleSrc:
624 case TTI::SK_PermuteTwoSrc:
625 return getPermuteShuffleOverhead(Tp);
626 case TTI::SK_ExtractSubvector:
627 return getExtractSubvectorOverhead(Tp, Index, SubTp);
628 case TTI::SK_InsertSubvector:
629 return getInsertSubvectorOverhead(Tp, Index, SubTp);
630 }
631 llvm_unreachable("Unknown TTI::ShuffleKind")::llvm::llvm_unreachable_internal("Unknown TTI::ShuffleKind",
"/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 631)
;
632 }
633
634 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
635 const Instruction *I = nullptr) {
636 const TargetLoweringBase *TLI = getTLI();
637 int ISD = TLI->InstructionOpcodeToISD(Opcode);
638 assert(ISD && "Invalid opcode")((ISD && "Invalid opcode") ? static_cast<void> (
0) : __assert_fail ("ISD && \"Invalid opcode\"", "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 638, __PRETTY_FUNCTION__))
;
639 std::pair<unsigned, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, Src);
640 std::pair<unsigned, MVT> DstLT = TLI->getTypeLegalizationCost(DL, Dst);
641
642 // Check for NOOP conversions.
643 if (SrcLT.first == DstLT.first &&
644 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
645
646 // Bitcast between types that are legalized to the same type are free.
647 if (Opcode == Instruction::BitCast || Opcode == Instruction::Trunc)
648 return 0;
649 }
650
651 if (Opcode == Instruction::Trunc &&
652 TLI->isTruncateFree(SrcLT.second, DstLT.second))
653 return 0;
654
655 if (Opcode == Instruction::ZExt &&
656 TLI->isZExtFree(SrcLT.second, DstLT.second))
657 return 0;
658
659 if (Opcode == Instruction::AddrSpaceCast &&
660 TLI->isNoopAddrSpaceCast(Src->getPointerAddressSpace(),
661 Dst->getPointerAddressSpace()))
662 return 0;
663
664 // If this is a zext/sext of a load, return 0 if the corresponding
665 // extending load exists on target.
666 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
667 I && isa<LoadInst>(I->getOperand(0))) {
668 EVT ExtVT = EVT::getEVT(Dst);
669 EVT LoadVT = EVT::getEVT(Src);
670 unsigned LType =
671 ((Opcode == Instruction::ZExt) ? ISD::ZEXTLOAD : ISD::SEXTLOAD);
672 if (TLI->isLoadExtLegal(LType, ExtVT, LoadVT))
673 return 0;
674 }
675
676 // If the cast is marked as legal (or promote) then assume low cost.
677 if (SrcLT.first == DstLT.first &&
678 TLI->isOperationLegalOrPromote(ISD, DstLT.second))
679 return 1;
680
681 // Handle scalar conversions.
682 if (!Src->isVectorTy() && !Dst->isVectorTy()) {
683 // Scalar bitcasts are usually free.
684 if (Opcode == Instruction::BitCast)
685 return 0;
686
687 // Just check the op cost. If the operation is legal then assume it costs
688 // 1.
689 if (!TLI->isOperationExpand(ISD, DstLT.second))
690 return 1;
691
692 // Assume that illegal scalar instruction are expensive.
693 return 4;
694 }
695
696 // Check vector-to-vector casts.
697 if (Dst->isVectorTy() && Src->isVectorTy()) {
698 // If the cast is between same-sized registers, then the check is simple.
699 if (SrcLT.first == DstLT.first &&
700 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
701
702 // Assume that Zext is done using AND.
703 if (Opcode == Instruction::ZExt)
704 return 1;
705
706 // Assume that sext is done using SHL and SRA.
707 if (Opcode == Instruction::SExt)
708 return 2;
709
710 // Just check the op cost. If the operation is legal then assume it
711 // costs
712 // 1 and multiply by the type-legalization overhead.
713 if (!TLI->isOperationExpand(ISD, DstLT.second))
714 return SrcLT.first * 1;
715 }
716
717 // If we are legalizing by splitting, query the concrete TTI for the cost
718 // of casting the original vector twice. We also need to factor in the
719 // cost of the split itself. Count that as 1, to be consistent with
720 // TLI->getTypeLegalizationCost().
721 if ((TLI->getTypeAction(Src->getContext(), TLI->getValueType(DL, Src)) ==
722 TargetLowering::TypeSplitVector) ||
723 (TLI->getTypeAction(Dst->getContext(), TLI->getValueType(DL, Dst)) ==
724 TargetLowering::TypeSplitVector)) {
725 Type *SplitDst = VectorType::get(Dst->getVectorElementType(),
726 Dst->getVectorNumElements() / 2);
727 Type *SplitSrc = VectorType::get(Src->getVectorElementType(),
728 Src->getVectorNumElements() / 2);
729 T *TTI = static_cast<T *>(this);
730 return TTI->getVectorSplitCost() +
731 (2 * TTI->getCastInstrCost(Opcode, SplitDst, SplitSrc, I));
732 }
733
734 // In other cases where the source or destination are illegal, assume
735 // the operation will get scalarized.
736 unsigned Num = Dst->getVectorNumElements();
737 unsigned Cost = static_cast<T *>(this)->getCastInstrCost(
738 Opcode, Dst->getScalarType(), Src->getScalarType(), I);
739
740 // Return the cost of multiple scalar invocation plus the cost of
741 // inserting and extracting the values.
742 return getScalarizationOverhead(Dst, true, true) + Num * Cost;
743 }
744
745 // We already handled vector-to-vector and scalar-to-scalar conversions.
746 // This
747 // is where we handle bitcast between vectors and scalars. We need to assume
748 // that the conversion is scalarized in one way or another.
749 if (Opcode == Instruction::BitCast)
750 // Illegal bitcasts are done by storing and loading from a stack slot.
751 return (Src->isVectorTy() ? getScalarizationOverhead(Src, false, true)
752 : 0) +
753 (Dst->isVectorTy() ? getScalarizationOverhead(Dst, true, false)
754 : 0);
755
756 llvm_unreachable("Unhandled cast")::llvm::llvm_unreachable_internal("Unhandled cast", "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 756)
;
757 }
758
759 unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
760 VectorType *VecTy, unsigned Index) {
761 return static_cast<T *>(this)->getVectorInstrCost(
762 Instruction::ExtractElement, VecTy, Index) +
763 static_cast<T *>(this)->getCastInstrCost(Opcode, Dst,
764 VecTy->getElementType());
765 }
766
767 unsigned getCFInstrCost(unsigned Opcode) {
768 // Branches are assumed to be predicted.
769 return 0;
770 }
771
772 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
773 const Instruction *I) {
774 const TargetLoweringBase *TLI = getTLI();
775 int ISD = TLI->InstructionOpcodeToISD(Opcode);
776 assert(ISD && "Invalid opcode")((ISD && "Invalid opcode") ? static_cast<void> (
0) : __assert_fail ("ISD && \"Invalid opcode\"", "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 776, __PRETTY_FUNCTION__))
;
777
778 // Selects on vectors are actually vector selects.
779 if (ISD == ISD::SELECT) {
780 assert(CondTy && "CondTy must exist")((CondTy && "CondTy must exist") ? static_cast<void
> (0) : __assert_fail ("CondTy && \"CondTy must exist\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 780, __PRETTY_FUNCTION__))
;
781 if (CondTy->isVectorTy())
782 ISD = ISD::VSELECT;
783 }
784 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
785
786 if (!(ValTy->isVectorTy() && !LT.second.isVector()) &&
787 !TLI->isOperationExpand(ISD, LT.second)) {
788 // The operation is legal. Assume it costs 1. Multiply
789 // by the type-legalization overhead.
790 return LT.first * 1;
791 }
792
793 // Otherwise, assume that the cast is scalarized.
794 // TODO: If one of the types get legalized by splitting, handle this
795 // similarly to what getCastInstrCost() does.
796 if (ValTy->isVectorTy()) {
797 unsigned Num = ValTy->getVectorNumElements();
798 if (CondTy)
799 CondTy = CondTy->getScalarType();
800 unsigned Cost = static_cast<T *>(this)->getCmpSelInstrCost(
801 Opcode, ValTy->getScalarType(), CondTy, I);
802
803 // Return the cost of multiple scalar invocation plus the cost of
804 // inserting and extracting the values.
805 return getScalarizationOverhead(ValTy, true, false) + Num * Cost;
806 }
807
808 // Unknown scalar opcode.
809 return 1;
810 }
811
812 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
813 std::pair<unsigned, MVT> LT =
814 getTLI()->getTypeLegalizationCost(DL, Val->getScalarType());
815
816 return LT.first;
817 }
818
819 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
820 unsigned AddressSpace, const Instruction *I = nullptr) {
821 assert(!Src->isVoidTy() && "Invalid type")((!Src->isVoidTy() && "Invalid type") ? static_cast
<void> (0) : __assert_fail ("!Src->isVoidTy() && \"Invalid type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 821, __PRETTY_FUNCTION__))
;
822 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Src);
823
824 // Assuming that all loads of legal types cost 1.
825 unsigned Cost = LT.first;
826
827 if (Src->isVectorTy() &&
828 Src->getPrimitiveSizeInBits() < LT.second.getSizeInBits()) {
829 // This is a vector load that legalizes to a larger type than the vector
830 // itself. Unless the corresponding extending load or truncating store is
831 // legal, then this will scalarize.
832 TargetLowering::LegalizeAction LA = TargetLowering::Expand;
833 EVT MemVT = getTLI()->getValueType(DL, Src);
834 if (Opcode == Instruction::Store)
835 LA = getTLI()->getTruncStoreAction(LT.second, MemVT);
836 else
837 LA = getTLI()->getLoadExtAction(ISD::EXTLOAD, LT.second, MemVT);
838
839 if (LA != TargetLowering::Legal && LA != TargetLowering::Custom) {
840 // This is a vector load/store for some illegal type that is scalarized.
841 // We must account for the cost of building or decomposing the vector.
842 Cost += getScalarizationOverhead(Src, Opcode != Instruction::Store,
843 Opcode == Instruction::Store);
844 }
845 }
846
847 return Cost;
848 }
849
850 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
851 unsigned Factor,
852 ArrayRef<unsigned> Indices,
853 unsigned Alignment, unsigned AddressSpace,
854 bool UseMaskForCond = false,
855 bool UseMaskForGaps = false) {
856 VectorType *VT = dyn_cast<VectorType>(VecTy);
857 assert(VT && "Expect a vector type for interleaved memory op")((VT && "Expect a vector type for interleaved memory op"
) ? static_cast<void> (0) : __assert_fail ("VT && \"Expect a vector type for interleaved memory op\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 857, __PRETTY_FUNCTION__))
;
858
859 unsigned NumElts = VT->getNumElements();
860 assert(Factor > 1 && NumElts % Factor == 0 && "Invalid interleave factor")((Factor > 1 && NumElts % Factor == 0 && "Invalid interleave factor"
) ? static_cast<void> (0) : __assert_fail ("Factor > 1 && NumElts % Factor == 0 && \"Invalid interleave factor\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 860, __PRETTY_FUNCTION__))
;
861
862 unsigned NumSubElts = NumElts / Factor;
863 VectorType *SubVT = VectorType::get(VT->getElementType(), NumSubElts);
864
865 // Firstly, the cost of load/store operation.
866 unsigned Cost;
867 if (UseMaskForCond || UseMaskForGaps)
868 Cost = static_cast<T *>(this)->getMaskedMemoryOpCost(
869 Opcode, VecTy, Alignment, AddressSpace);
870 else
871 Cost = static_cast<T *>(this)->getMemoryOpCost(Opcode, VecTy, Alignment,
872 AddressSpace);
873
874 // Legalize the vector type, and get the legalized and unlegalized type
875 // sizes.
876 MVT VecTyLT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
877 unsigned VecTySize =
878 static_cast<T *>(this)->getDataLayout().getTypeStoreSize(VecTy);
879 unsigned VecTyLTSize = VecTyLT.getStoreSize();
880
881 // Return the ceiling of dividing A by B.
882 auto ceil = [](unsigned A, unsigned B) { return (A + B - 1) / B; };
883
884 // Scale the cost of the memory operation by the fraction of legalized
885 // instructions that will actually be used. We shouldn't account for the
886 // cost of dead instructions since they will be removed.
887 //
888 // E.g., An interleaved load of factor 8:
889 // %vec = load <16 x i64>, <16 x i64>* %ptr
890 // %v0 = shufflevector %vec, undef, <0, 8>
891 //
892 // If <16 x i64> is legalized to 8 v2i64 loads, only 2 of the loads will be
893 // used (those corresponding to elements [0:1] and [8:9] of the unlegalized
894 // type). The other loads are unused.
895 //
896 // We only scale the cost of loads since interleaved store groups aren't
897 // allowed to have gaps.
898 if (Opcode == Instruction::Load && VecTySize > VecTyLTSize) {
899 // The number of loads of a legal type it will take to represent a load
900 // of the unlegalized vector type.
901 unsigned NumLegalInsts = ceil(VecTySize, VecTyLTSize);
902
903 // The number of elements of the unlegalized type that correspond to a
904 // single legal instruction.
905 unsigned NumEltsPerLegalInst = ceil(NumElts, NumLegalInsts);
906
907 // Determine which legal instructions will be used.
908 BitVector UsedInsts(NumLegalInsts, false);
909 for (unsigned Index : Indices)
910 for (unsigned Elt = 0; Elt < NumSubElts; ++Elt)
911 UsedInsts.set((Index + Elt * Factor) / NumEltsPerLegalInst);
912
913 // Scale the cost of the load by the fraction of legal instructions that
914 // will be used.
915 Cost *= UsedInsts.count() / NumLegalInsts;
916 }
917
918 // Then plus the cost of interleave operation.
919 if (Opcode == Instruction::Load) {
920 // The interleave cost is similar to extract sub vectors' elements
921 // from the wide vector, and insert them into sub vectors.
922 //
923 // E.g. An interleaved load of factor 2 (with one member of index 0):
924 // %vec = load <8 x i32>, <8 x i32>* %ptr
925 // %v0 = shuffle %vec, undef, <0, 2, 4, 6> ; Index 0
926 // The cost is estimated as extract elements at 0, 2, 4, 6 from the
927 // <8 x i32> vector and insert them into a <4 x i32> vector.
928
929 assert(Indices.size() <= Factor &&((Indices.size() <= Factor && "Interleaved memory op has too many members"
) ? static_cast<void> (0) : __assert_fail ("Indices.size() <= Factor && \"Interleaved memory op has too many members\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 930, __PRETTY_FUNCTION__))
930 "Interleaved memory op has too many members")((Indices.size() <= Factor && "Interleaved memory op has too many members"
) ? static_cast<void> (0) : __assert_fail ("Indices.size() <= Factor && \"Interleaved memory op has too many members\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 930, __PRETTY_FUNCTION__))
;
931
932 for (unsigned Index : Indices) {
933 assert(Index < Factor && "Invalid index for interleaved memory op")((Index < Factor && "Invalid index for interleaved memory op"
) ? static_cast<void> (0) : __assert_fail ("Index < Factor && \"Invalid index for interleaved memory op\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 933, __PRETTY_FUNCTION__))
;
934
935 // Extract elements from loaded vector for each sub vector.
936 for (unsigned i = 0; i < NumSubElts; i++)
937 Cost += static_cast<T *>(this)->getVectorInstrCost(
938 Instruction::ExtractElement, VT, Index + i * Factor);
939 }
940
941 unsigned InsSubCost = 0;
942 for (unsigned i = 0; i < NumSubElts; i++)
943 InsSubCost += static_cast<T *>(this)->getVectorInstrCost(
944 Instruction::InsertElement, SubVT, i);
945
946 Cost += Indices.size() * InsSubCost;
947 } else {
948 // The interleave cost is extract all elements from sub vectors, and
949 // insert them into the wide vector.
950 //
951 // E.g. An interleaved store of factor 2:
952 // %v0_v1 = shuffle %v0, %v1, <0, 4, 1, 5, 2, 6, 3, 7>
953 // store <8 x i32> %interleaved.vec, <8 x i32>* %ptr
954 // The cost is estimated as extract all elements from both <4 x i32>
955 // vectors and insert into the <8 x i32> vector.
956
957 unsigned ExtSubCost = 0;
958 for (unsigned i = 0; i < NumSubElts; i++)
959 ExtSubCost += static_cast<T *>(this)->getVectorInstrCost(
960 Instruction::ExtractElement, SubVT, i);
961 Cost += ExtSubCost * Factor;
962
963 for (unsigned i = 0; i < NumElts; i++)
964 Cost += static_cast<T *>(this)
965 ->getVectorInstrCost(Instruction::InsertElement, VT, i);
966 }
967
968 if (!UseMaskForCond)
969 return Cost;
970
971 Type *I8Type = Type::getInt8Ty(VT->getContext());
972 VectorType *MaskVT = VectorType::get(I8Type, NumElts);
973 SubVT = VectorType::get(I8Type, NumSubElts);
974
975 // The Mask shuffling cost is extract all the elements of the Mask
976 // and insert each of them Factor times into the wide vector:
977 //
978 // E.g. an interleaved group with factor 3:
979 // %mask = icmp ult <8 x i32> %vec1, %vec2
980 // %interleaved.mask = shufflevector <8 x i1> %mask, <8 x i1> undef,
981 // <24 x i32> <0,0,0,1,1,1,2,2,2,3,3,3,4,4,4,5,5,5,6,6,6,7,7,7>
982 // The cost is estimated as extract all mask elements from the <8xi1> mask
983 // vector and insert them factor times into the <24xi1> shuffled mask
984 // vector.
985 for (unsigned i = 0; i < NumSubElts; i++)
986 Cost += static_cast<T *>(this)->getVectorInstrCost(
987 Instruction::ExtractElement, SubVT, i);
988
989 for (unsigned i = 0; i < NumElts; i++)
990 Cost += static_cast<T *>(this)->getVectorInstrCost(
991 Instruction::InsertElement, MaskVT, i);
992
993 // The Gaps mask is invariant and created outside the loop, therefore the
994 // cost of creating it is not accounted for here. However if we have both
995 // a MaskForGaps and some other mask that guards the execution of the
996 // memory access, we need to account for the cost of And-ing the two masks
997 // inside the loop.
998 if (UseMaskForGaps)
999 Cost += static_cast<T *>(this)->getArithmeticInstrCost(
1000 BinaryOperator::And, MaskVT);
1001
1002 return Cost;
1003 }
1004
1005 /// Get intrinsic cost based on arguments.
1006 unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1007 ArrayRef<Value *> Args, FastMathFlags FMF,
1008 unsigned VF = 1) {
1009 unsigned RetVF = (RetTy->isVectorTy() ? RetTy->getVectorNumElements() : 1);
1010 assert((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type")(((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type"
) ? static_cast<void> (0) : __assert_fail ("(RetVF == 1 || VF == 1) && \"VF > 1 and RetVF is a vector type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 1010, __PRETTY_FUNCTION__))
;
1011 auto *ConcreteTTI = static_cast<T *>(this);
1012
1013 switch (IID) {
1014 default: {
1015 // Assume that we need to scalarize this intrinsic.
1016 SmallVector<Type *, 4> Types;
1017 for (Value *Op : Args) {
1018 Type *OpTy = Op->getType();
1019 assert(VF == 1 || !OpTy->isVectorTy())((VF == 1 || !OpTy->isVectorTy()) ? static_cast<void>
(0) : __assert_fail ("VF == 1 || !OpTy->isVectorTy()", "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 1019, __PRETTY_FUNCTION__))
;
1020 Types.push_back(VF == 1 ? OpTy : VectorType::get(OpTy, VF));
1021 }
1022
1023 if (VF > 1 && !RetTy->isVoidTy())
1024 RetTy = VectorType::get(RetTy, VF);
1025
1026 // Compute the scalarization overhead based on Args for a vector
1027 // intrinsic. A vectorizer will pass a scalar RetTy and VF > 1, while
1028 // CostModel will pass a vector RetTy and VF is 1.
1029 unsigned ScalarizationCost = std::numeric_limits<unsigned>::max();
1030 if (RetVF > 1 || VF > 1) {
1031 ScalarizationCost = 0;
1032 if (!RetTy->isVoidTy())
1033 ScalarizationCost += getScalarizationOverhead(RetTy, true, false);
1034 ScalarizationCost += getOperandsScalarizationOverhead(Args, VF);
1035 }
1036
1037 return ConcreteTTI->getIntrinsicInstrCost(IID, RetTy, Types, FMF,
1038 ScalarizationCost);
1039 }
1040 case Intrinsic::masked_scatter: {
1041 assert(VF == 1 && "Can't vectorize types here.")((VF == 1 && "Can't vectorize types here.") ? static_cast
<void> (0) : __assert_fail ("VF == 1 && \"Can't vectorize types here.\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 1041, __PRETTY_FUNCTION__))
;
1042 Value *Mask = Args[3];
1043 bool VarMask = !isa<Constant>(Mask);
1044 unsigned Alignment = cast<ConstantInt>(Args[2])->getZExtValue();
1045 return ConcreteTTI->getGatherScatterOpCost(
1046 Instruction::Store, Args[0]->getType(), Args[1], VarMask, Alignment);
1047 }
1048 case Intrinsic::masked_gather: {
1049 assert(VF == 1 && "Can't vectorize types here.")((VF == 1 && "Can't vectorize types here.") ? static_cast
<void> (0) : __assert_fail ("VF == 1 && \"Can't vectorize types here.\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 1049, __PRETTY_FUNCTION__))
;
1050 Value *Mask = Args[2];
1051 bool VarMask = !isa<Constant>(Mask);
1052 unsigned Alignment = cast<ConstantInt>(Args[1])->getZExtValue();
1053 return ConcreteTTI->getGatherScatterOpCost(Instruction::Load, RetTy,
1054 Args[0], VarMask, Alignment);
1055 }
1056 case Intrinsic::experimental_vector_reduce_add:
1057 case Intrinsic::experimental_vector_reduce_mul:
1058 case Intrinsic::experimental_vector_reduce_and:
1059 case Intrinsic::experimental_vector_reduce_or:
1060 case Intrinsic::experimental_vector_reduce_xor:
1061 case Intrinsic::experimental_vector_reduce_fadd:
1062 case Intrinsic::experimental_vector_reduce_fmul:
1063 case Intrinsic::experimental_vector_reduce_smax:
1064 case Intrinsic::experimental_vector_reduce_smin:
1065 case Intrinsic::experimental_vector_reduce_fmax:
1066 case Intrinsic::experimental_vector_reduce_fmin:
1067 case Intrinsic::experimental_vector_reduce_umax:
1068 case Intrinsic::experimental_vector_reduce_umin:
1069 return getIntrinsicInstrCost(IID, RetTy, Args[0]->getType(), FMF);
1070 case Intrinsic::fshl:
1071 case Intrinsic::fshr: {
1072 Value *X = Args[0];
1073 Value *Y = Args[1];
1074 Value *Z = Args[2];
1075 TTI::OperandValueProperties OpPropsX, OpPropsY, OpPropsZ, OpPropsBW;
1076 TTI::OperandValueKind OpKindX = TTI::getOperandInfo(X, OpPropsX);
1077 TTI::OperandValueKind OpKindY = TTI::getOperandInfo(Y, OpPropsY);
1078 TTI::OperandValueKind OpKindZ = TTI::getOperandInfo(Z, OpPropsZ);
1079 TTI::OperandValueKind OpKindBW = TTI::OK_UniformConstantValue;
1080 OpPropsBW = isPowerOf2_32(RetTy->getScalarSizeInBits()) ? TTI::OP_PowerOf2
1081 : TTI::OP_None;
1082 // fshl: (X << (Z % BW)) | (Y >> (BW - (Z % BW)))
1083 // fshr: (X << (BW - (Z % BW))) | (Y >> (Z % BW))
1084 unsigned Cost = 0;
1085 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::Or, RetTy);
1086 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::Sub, RetTy);
1087 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::Shl, RetTy,
1088 OpKindX, OpKindZ, OpPropsX);
1089 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::LShr, RetTy,
1090 OpKindY, OpKindZ, OpPropsY);
1091 // Non-constant shift amounts requires a modulo.
1092 if (OpKindZ != TTI::OK_UniformConstantValue &&
1093 OpKindZ != TTI::OK_NonUniformConstantValue)
1094 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::URem, RetTy,
1095 OpKindZ, OpKindBW, OpPropsZ,
1096 OpPropsBW);
1097 // For non-rotates (X != Y) we must add shift-by-zero handling costs.
1098 if (X != Y) {
1099 Type *CondTy = Type::getInt1Ty(RetTy->getContext());
1100 if (RetVF > 1)
1101 CondTy = VectorType::get(CondTy, RetVF);
1102 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::ICmp, RetTy,
1103 CondTy, nullptr);
1104 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::Select, RetTy,
1105 CondTy, nullptr);
1106 }
1107 return Cost;
1108 }
1109 }
1110 }
1111
1112 /// Get intrinsic cost based on argument types.
1113 /// If ScalarizationCostPassed is std::numeric_limits<unsigned>::max(), the
1114 /// cost of scalarizing the arguments and the return value will be computed
1115 /// based on types.
1116 unsigned getIntrinsicInstrCost(
1117 Intrinsic::ID IID, Type *RetTy, ArrayRef<Type *> Tys, FastMathFlags FMF,
1118 unsigned ScalarizationCostPassed = std::numeric_limits<unsigned>::max()) {
1119 SmallVector<unsigned, 2> ISDs;
1120 unsigned SingleCallCost = 10; // Library call cost. Make it expensive.
1121 switch (IID) {
1122 default: {
1123 // Assume that we need to scalarize this intrinsic.
1124 unsigned ScalarizationCost = ScalarizationCostPassed;
1125 unsigned ScalarCalls = 1;
1126 Type *ScalarRetTy = RetTy;
1127 if (RetTy->isVectorTy()) {
1128 if (ScalarizationCostPassed == std::numeric_limits<unsigned>::max())
1129 ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
1130 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
1131 ScalarRetTy = RetTy->getScalarType();
1132 }
1133 SmallVector<Type *, 4> ScalarTys;
1134 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1135 Type *Ty = Tys[i];
1136 if (Ty->isVectorTy()) {
1137 if (ScalarizationCostPassed == std::numeric_limits<unsigned>::max())
1138 ScalarizationCost += getScalarizationOverhead(Ty, false, true);
1139 ScalarCalls = std::max(ScalarCalls, Ty->getVectorNumElements());
1140 Ty = Ty->getScalarType();
1141 }
1142 ScalarTys.push_back(Ty);
1143 }
1144 if (ScalarCalls == 1)
1145 return 1; // Return cost of a scalar intrinsic. Assume it to be cheap.
1146
1147 unsigned ScalarCost = static_cast<T *>(this)->getIntrinsicInstrCost(
1148 IID, ScalarRetTy, ScalarTys, FMF);
1149
1150 return ScalarCalls * ScalarCost + ScalarizationCost;
1151 }
1152 // Look for intrinsics that can be lowered directly or turned into a scalar
1153 // intrinsic call.
1154 case Intrinsic::sqrt:
1155 ISDs.push_back(ISD::FSQRT);
1156 break;
1157 case Intrinsic::sin:
1158 ISDs.push_back(ISD::FSIN);
1159 break;
1160 case Intrinsic::cos:
1161 ISDs.push_back(ISD::FCOS);
1162 break;
1163 case Intrinsic::exp:
1164 ISDs.push_back(ISD::FEXP);
1165 break;
1166 case Intrinsic::exp2:
1167 ISDs.push_back(ISD::FEXP2);
1168 break;
1169 case Intrinsic::log:
1170 ISDs.push_back(ISD::FLOG);
1171 break;
1172 case Intrinsic::log10:
1173 ISDs.push_back(ISD::FLOG10);
1174 break;
1175 case Intrinsic::log2:
1176 ISDs.push_back(ISD::FLOG2);
1177 break;
1178 case Intrinsic::fabs:
1179 ISDs.push_back(ISD::FABS);
1180 break;
1181 case Intrinsic::canonicalize:
1182 ISDs.push_back(ISD::FCANONICALIZE);
1183 break;
1184 case Intrinsic::minnum:
1185 ISDs.push_back(ISD::FMINNUM);
1186 if (FMF.noNaNs())
1187 ISDs.push_back(ISD::FMINIMUM);
1188 break;
1189 case Intrinsic::maxnum:
1190 ISDs.push_back(ISD::FMAXNUM);
1191 if (FMF.noNaNs())
1192 ISDs.push_back(ISD::FMAXIMUM);
1193 break;
1194 case Intrinsic::copysign:
1195 ISDs.push_back(ISD::FCOPYSIGN);
1196 break;
1197 case Intrinsic::floor:
1198 ISDs.push_back(ISD::FFLOOR);
1199 break;
1200 case Intrinsic::ceil:
1201 ISDs.push_back(ISD::FCEIL);
1202 break;
1203 case Intrinsic::trunc:
1204 ISDs.push_back(ISD::FTRUNC);
1205 break;
1206 case Intrinsic::nearbyint:
1207 ISDs.push_back(ISD::FNEARBYINT);
1208 break;
1209 case Intrinsic::rint:
1210 ISDs.push_back(ISD::FRINT);
1211 break;
1212 case Intrinsic::round:
1213 ISDs.push_back(ISD::FROUND);
1214 break;
1215 case Intrinsic::pow:
1216 ISDs.push_back(ISD::FPOW);
1217 break;
1218 case Intrinsic::fma:
1219 ISDs.push_back(ISD::FMA);
1220 break;
1221 case Intrinsic::fmuladd:
1222 ISDs.push_back(ISD::FMA);
1223 break;
1224 // FIXME: We should return 0 whenever getIntrinsicCost == TCC_Free.
1225 case Intrinsic::lifetime_start:
1226 case Intrinsic::lifetime_end:
1227 case Intrinsic::sideeffect:
1228 return 0;
1229 case Intrinsic::masked_store:
1230 return static_cast<T *>(this)
1231 ->getMaskedMemoryOpCost(Instruction::Store, Tys[0], 0, 0);
1232 case Intrinsic::masked_load:
1233 return static_cast<T *>(this)
1234 ->getMaskedMemoryOpCost(Instruction::Load, RetTy, 0, 0);
1235 case Intrinsic::experimental_vector_reduce_add:
1236 return static_cast<T *>(this)->getArithmeticReductionCost(
1237 Instruction::Add, Tys[0], /*IsPairwiseForm=*/false);
1238 case Intrinsic::experimental_vector_reduce_mul:
1239 return static_cast<T *>(this)->getArithmeticReductionCost(
1240 Instruction::Mul, Tys[0], /*IsPairwiseForm=*/false);
1241 case Intrinsic::experimental_vector_reduce_and:
1242 return static_cast<T *>(this)->getArithmeticReductionCost(
1243 Instruction::And, Tys[0], /*IsPairwiseForm=*/false);
1244 case Intrinsic::experimental_vector_reduce_or:
1245 return static_cast<T *>(this)->getArithmeticReductionCost(
1246 Instruction::Or, Tys[0], /*IsPairwiseForm=*/false);
1247 case Intrinsic::experimental_vector_reduce_xor:
1248 return static_cast<T *>(this)->getArithmeticReductionCost(
1249 Instruction::Xor, Tys[0], /*IsPairwiseForm=*/false);
1250 case Intrinsic::experimental_vector_reduce_fadd:
1251 return static_cast<T *>(this)->getArithmeticReductionCost(
1252 Instruction::FAdd, Tys[0], /*IsPairwiseForm=*/false);
1253 case Intrinsic::experimental_vector_reduce_fmul:
1254 return static_cast<T *>(this)->getArithmeticReductionCost(
1255 Instruction::FMul, Tys[0], /*IsPairwiseForm=*/false);
1256 case Intrinsic::experimental_vector_reduce_smax:
1257 case Intrinsic::experimental_vector_reduce_smin:
1258 case Intrinsic::experimental_vector_reduce_fmax:
1259 case Intrinsic::experimental_vector_reduce_fmin:
1260 return static_cast<T *>(this)->getMinMaxReductionCost(
1261 Tys[0], CmpInst::makeCmpResultType(Tys[0]), /*IsPairwiseForm=*/false,
1262 /*IsSigned=*/true);
1263 case Intrinsic::experimental_vector_reduce_umax:
1264 case Intrinsic::experimental_vector_reduce_umin:
1265 return static_cast<T *>(this)->getMinMaxReductionCost(
1266 Tys[0], CmpInst::makeCmpResultType(Tys[0]), /*IsPairwiseForm=*/false,
1267 /*IsSigned=*/false);
1268 case Intrinsic::ctpop:
1269 ISDs.push_back(ISD::CTPOP);
1270 // In case of legalization use TCC_Expensive. This is cheaper than a
1271 // library call but still not a cheap instruction.
1272 SingleCallCost = TargetTransformInfo::TCC_Expensive;
1273 break;
1274 // FIXME: ctlz, cttz, ...
1275 }
1276
1277 const TargetLoweringBase *TLI = getTLI();
1278 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
1279
1280 SmallVector<unsigned, 2> LegalCost;
1281 SmallVector<unsigned, 2> CustomCost;
1282 for (unsigned ISD : ISDs) {
1283 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
1284 if (IID == Intrinsic::fabs && LT.second.isFloatingPoint() &&
1285 TLI->isFAbsFree(LT.second)) {
1286 return 0;
1287 }
1288
1289 // The operation is legal. Assume it costs 1.
1290 // If the type is split to multiple registers, assume that there is some
1291 // overhead to this.
1292 // TODO: Once we have extract/insert subvector cost we need to use them.
1293 if (LT.first > 1)
1294 LegalCost.push_back(LT.first * 2);
1295 else
1296 LegalCost.push_back(LT.first * 1);
1297 } else if (!TLI->isOperationExpand(ISD, LT.second)) {
1298 // If the operation is custom lowered then assume
1299 // that the code is twice as expensive.
1300 CustomCost.push_back(LT.first * 2);
1301 }
1302 }
1303
1304 auto MinLegalCostI = std::min_element(LegalCost.begin(), LegalCost.end());
1305 if (MinLegalCostI != LegalCost.end())
1306 return *MinLegalCostI;
1307
1308 auto MinCustomCostI = std::min_element(CustomCost.begin(), CustomCost.end());
1309 if (MinCustomCostI != CustomCost.end())
1310 return *MinCustomCostI;
1311
1312 // If we can't lower fmuladd into an FMA estimate the cost as a floating
1313 // point mul followed by an add.
1314 if (IID == Intrinsic::fmuladd)
1315 return static_cast<T *>(this)
1316 ->getArithmeticInstrCost(BinaryOperator::FMul, RetTy) +
1317 static_cast<T *>(this)
1318 ->getArithmeticInstrCost(BinaryOperator::FAdd, RetTy);
1319
1320 // Else, assume that we need to scalarize this intrinsic. For math builtins
1321 // this will emit a costly libcall, adding call overhead and spills. Make it
1322 // very expensive.
1323 if (RetTy->isVectorTy()) {
1324 unsigned ScalarizationCost =
1325 ((ScalarizationCostPassed != std::numeric_limits<unsigned>::max())
1326 ? ScalarizationCostPassed
1327 : getScalarizationOverhead(RetTy, true, false));
1328 unsigned ScalarCalls = RetTy->getVectorNumElements();
1329 SmallVector<Type *, 4> ScalarTys;
1330 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1331 Type *Ty = Tys[i];
1332 if (Ty->isVectorTy())
1333 Ty = Ty->getScalarType();
1334 ScalarTys.push_back(Ty);
1335 }
1336 unsigned ScalarCost = static_cast<T *>(this)->getIntrinsicInstrCost(
1337 IID, RetTy->getScalarType(), ScalarTys, FMF);
1338 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1339 if (Tys[i]->isVectorTy()) {
1340 if (ScalarizationCostPassed == std::numeric_limits<unsigned>::max())
1341 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
1342 ScalarCalls = std::max(ScalarCalls, Tys[i]->getVectorNumElements());
1343 }
1344 }
1345
1346 return ScalarCalls * ScalarCost + ScalarizationCost;
1347 }
1348
1349 // This is going to be turned into a library call, make it expensive.
1350 return SingleCallCost;
1351 }
1352
1353 /// Compute a cost of the given call instruction.
1354 ///
1355 /// Compute the cost of calling function F with return type RetTy and
1356 /// argument types Tys. F might be nullptr, in this case the cost of an
1357 /// arbitrary call with the specified signature will be returned.
1358 /// This is used, for instance, when we estimate call of a vector
1359 /// counterpart of the given function.
1360 /// \param F Called function, might be nullptr.
1361 /// \param RetTy Return value types.
1362 /// \param Tys Argument types.
1363 /// \returns The cost of Call instruction.
1364 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
1365 return 10;
1366 }
1367
1368 unsigned getNumberOfParts(Type *Tp) {
1369 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Tp);
1370 return LT.first;
1371 }
1372
1373 unsigned getAddressComputationCost(Type *Ty, ScalarEvolution *,
1374 const SCEV *) {
1375 return 0;
1376 }
1377
1378 /// Try to calculate arithmetic and shuffle op costs for reduction operations.
1379 /// We're assuming that reduction operation are performing the following way:
1380 /// 1. Non-pairwise reduction
1381 /// %val1 = shufflevector<n x t> %val, <n x t> %undef,
1382 /// <n x i32> <i32 n/2, i32 n/2 + 1, ..., i32 n, i32 undef, ..., i32 undef>
1383 /// \----------------v-------------/ \----------v------------/
1384 /// n/2 elements n/2 elements
1385 /// %red1 = op <n x t> %val, <n x t> val1
1386 /// After this operation we have a vector %red1 where only the first n/2
1387 /// elements are meaningful, the second n/2 elements are undefined and can be
1388 /// dropped. All other operations are actually working with the vector of
1389 /// length n/2, not n, though the real vector length is still n.
1390 /// %val2 = shufflevector<n x t> %red1, <n x t> %undef,
1391 /// <n x i32> <i32 n/4, i32 n/4 + 1, ..., i32 n/2, i32 undef, ..., i32 undef>
1392 /// \----------------v-------------/ \----------v------------/
1393 /// n/4 elements 3*n/4 elements
1394 /// %red2 = op <n x t> %red1, <n x t> val2 - working with the vector of
1395 /// length n/2, the resulting vector has length n/4 etc.
1396 /// 2. Pairwise reduction:
1397 /// Everything is the same except for an additional shuffle operation which
1398 /// is used to produce operands for pairwise kind of reductions.
1399 /// %val1 = shufflevector<n x t> %val, <n x t> %undef,
1400 /// <n x i32> <i32 0, i32 2, ..., i32 n-2, i32 undef, ..., i32 undef>
1401 /// \-------------v----------/ \----------v------------/
1402 /// n/2 elements n/2 elements
1403 /// %val2 = shufflevector<n x t> %val, <n x t> %undef,
1404 /// <n x i32> <i32 1, i32 3, ..., i32 n-1, i32 undef, ..., i32 undef>
1405 /// \-------------v----------/ \----------v------------/
1406 /// n/2 elements n/2 elements
1407 /// %red1 = op <n x t> %val1, <n x t> val2
1408 /// Again, the operation is performed on <n x t> vector, but the resulting
1409 /// vector %red1 is <n/2 x t> vector.
1410 ///
1411 /// The cost model should take into account that the actual length of the
1412 /// vector is reduced on each iteration.
1413 unsigned getArithmeticReductionCost(unsigned Opcode, Type *Ty,
1414 bool IsPairwise) {
1415 assert(Ty->isVectorTy() && "Expect a vector type")((Ty->isVectorTy() && "Expect a vector type") ? static_cast
<void> (0) : __assert_fail ("Ty->isVectorTy() && \"Expect a vector type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 1415, __PRETTY_FUNCTION__))
;
1416 Type *ScalarTy = Ty->getVectorElementType();
1417 unsigned NumVecElts = Ty->getVectorNumElements();
1418 unsigned NumReduxLevels = Log2_32(NumVecElts);
1419 unsigned ArithCost = 0;
1420 unsigned ShuffleCost = 0;
1421 auto *ConcreteTTI = static_cast<T *>(this);
1422 std::pair<unsigned, MVT> LT =
1423 ConcreteTTI->getTLI()->getTypeLegalizationCost(DL, Ty);
1424 unsigned LongVectorCount = 0;
1425 unsigned MVTLen =
1426 LT.second.isVector() ? LT.second.getVectorNumElements() : 1;
1427 while (NumVecElts > MVTLen) {
1428 NumVecElts /= 2;
1429 Type *SubTy = VectorType::get(ScalarTy, NumVecElts);
1430 // Assume the pairwise shuffles add a cost.
1431 ShuffleCost += (IsPairwise + 1) *
1432 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1433 NumVecElts, SubTy);
1434 ArithCost += ConcreteTTI->getArithmeticInstrCost(Opcode, SubTy);
1435 Ty = SubTy;
1436 ++LongVectorCount;
1437 }
1438
1439 NumReduxLevels -= LongVectorCount;
1440
1441 // The minimal length of the vector is limited by the real length of vector
1442 // operations performed on the current platform. That's why several final
1443 // reduction operations are performed on the vectors with the same
1444 // architecture-dependent length.
1445
1446 // Non pairwise reductions need one shuffle per reduction level. Pairwise
1447 // reductions need two shuffles on every level, but the last one. On that
1448 // level one of the shuffles is <0, u, u, ...> which is identity.
1449 unsigned NumShuffles = NumReduxLevels;
1450 if (IsPairwise && NumReduxLevels >= 1)
1451 NumShuffles += NumReduxLevels - 1;
1452 ShuffleCost += NumShuffles *
1453 ConcreteTTI->getShuffleCost(TTI::SK_PermuteSingleSrc, Ty,
1454 0, Ty);
1455 ArithCost += NumReduxLevels *
1456 ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1457 return ShuffleCost + ArithCost +
1458 ConcreteTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
1459 }
1460
1461 /// Try to calculate op costs for min/max reduction operations.
1462 /// \param CondTy Conditional type for the Select instruction.
1463 unsigned getMinMaxReductionCost(Type *Ty, Type *CondTy, bool IsPairwise,
1464 bool) {
1465 assert(Ty->isVectorTy() && "Expect a vector type")((Ty->isVectorTy() && "Expect a vector type") ? static_cast
<void> (0) : __assert_fail ("Ty->isVectorTy() && \"Expect a vector type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 1465, __PRETTY_FUNCTION__))
;
1466 Type *ScalarTy = Ty->getVectorElementType();
1467 Type *ScalarCondTy = CondTy->getVectorElementType();
1468 unsigned NumVecElts = Ty->getVectorNumElements();
1469 unsigned NumReduxLevels = Log2_32(NumVecElts);
1470 unsigned CmpOpcode;
1471 if (Ty->isFPOrFPVectorTy()) {
1472 CmpOpcode = Instruction::FCmp;
1473 } else {
1474 assert(Ty->isIntOrIntVectorTy() &&((Ty->isIntOrIntVectorTy() && "expecting floating point or integer type for min/max reduction"
) ? static_cast<void> (0) : __assert_fail ("Ty->isIntOrIntVectorTy() && \"expecting floating point or integer type for min/max reduction\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 1475, __PRETTY_FUNCTION__))
1475 "expecting floating point or integer type for min/max reduction")((Ty->isIntOrIntVectorTy() && "expecting floating point or integer type for min/max reduction"
) ? static_cast<void> (0) : __assert_fail ("Ty->isIntOrIntVectorTy() && \"expecting floating point or integer type for min/max reduction\""
, "/build/llvm-toolchain-snapshot-8~svn350071/include/llvm/CodeGen/BasicTTIImpl.h"
, 1475, __PRETTY_FUNCTION__))
;
1476 CmpOpcode = Instruction::ICmp;
1477 }
1478 unsigned MinMaxCost = 0;
1479 unsigned ShuffleCost = 0;
1480 auto *ConcreteTTI = static_cast<T *>(this);
1481 std::pair<unsigned, MVT> LT =
1482 ConcreteTTI->getTLI()->getTypeLegalizationCost(DL, Ty);
1483 unsigned LongVectorCount = 0;
1484 unsigned MVTLen =
1485 LT.second.isVector() ? LT.second.getVectorNumElements() : 1;
1486 while (NumVecElts > MVTLen) {
1487 NumVecElts /= 2;
1488 Type *SubTy = VectorType::get(ScalarTy, NumVecElts);
1489 CondTy = VectorType::get(ScalarCondTy, NumVecElts);
1490
1491 // Assume the pairwise shuffles add a cost.
1492 ShuffleCost += (IsPairwise + 1) *
1493 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1494 NumVecElts, SubTy);
1495 MinMaxCost +=
1496 ConcreteTTI->getCmpSelInstrCost(CmpOpcode, SubTy, CondTy, nullptr) +
1497 ConcreteTTI->getCmpSelInstrCost(Instruction::Select, SubTy, CondTy,
1498 nullptr);
1499 Ty = SubTy;
1500 ++LongVectorCount;
1501 }
1502
1503 NumReduxLevels -= LongVectorCount;
1504
1505 // The minimal length of the vector is limited by the real length of vector
1506 // operations performed on the current platform. That's why several final
1507 // reduction opertions are perfomed on the vectors with the same
1508 // architecture-dependent length.
1509
1510 // Non pairwise reductions need one shuffle per reduction level. Pairwise
1511 // reductions need two shuffles on every level, but the last one. On that
1512 // level one of the shuffles is <0, u, u, ...> which is identity.
1513 unsigned NumShuffles = NumReduxLevels;
1514 if (IsPairwise && NumReduxLevels >= 1)
1515 NumShuffles += NumReduxLevels - 1;
1516 ShuffleCost += NumShuffles *
1517 ConcreteTTI->getShuffleCost(TTI::SK_PermuteSingleSrc, Ty,
1518 0, Ty);
1519 MinMaxCost +=
1520 NumReduxLevels *
1521 (ConcreteTTI->getCmpSelInstrCost(CmpOpcode, Ty, CondTy, nullptr) +
1522 ConcreteTTI->getCmpSelInstrCost(Instruction::Select, Ty, CondTy,
1523 nullptr));
1524 // The last min/max should be in vector registers and we counted it above.
1525 // So just need a single extractelement.
1526 return ShuffleCost + MinMaxCost +
1527 ConcreteTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
1528 }
1529
1530 unsigned getVectorSplitCost() { return 1; }
1531
1532 /// @}
1533};
1534
1535/// Concrete BasicTTIImpl that can be used if no further customization
1536/// is needed.
1537class BasicTTIImpl : public BasicTTIImplBase<BasicTTIImpl> {
1538 using BaseT = BasicTTIImplBase<BasicTTIImpl>;
1539
1540 friend class BasicTTIImplBase<BasicTTIImpl>;
1541
1542 const TargetSubtargetInfo *ST;
1543 const TargetLoweringBase *TLI;
1544
1545 const TargetSubtargetInfo *getST() const { return ST; }
1546 const TargetLoweringBase *getTLI() const { return TLI; }
1547
1548public:
1549 explicit BasicTTIImpl(const TargetMachine *TM, const Function &F);
1550};
1551
1552} // end namespace llvm
1553
1554#endif // LLVM_CODEGEN_BASICTTIIMPL_H