1 //===-- ARMTargetTransformInfo.cpp - ARM specific TTI ---------------------===//
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 #include "ARMTargetTransformInfo.h"
11 #include "llvm/Support/Debug.h"
12 #include "llvm/Target/CostTable.h"
13 #include "llvm/Target/TargetLowering.h"
14 using namespace llvm;
15
16 #define DEBUG_TYPE "armtti"
17
getIntImmCost(const APInt & Imm,Type * Ty)18 unsigned ARMTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) {
19 assert(Ty->isIntegerTy());
20
21 unsigned Bits = Ty->getPrimitiveSizeInBits();
22 if (Bits == 0 || Bits > 32)
23 return 4;
24
25 int32_t SImmVal = Imm.getSExtValue();
26 uint32_t ZImmVal = Imm.getZExtValue();
27 if (!ST->isThumb()) {
28 if ((SImmVal >= 0 && SImmVal < 65536) ||
29 (ARM_AM::getSOImmVal(ZImmVal) != -1) ||
30 (ARM_AM::getSOImmVal(~ZImmVal) != -1))
31 return 1;
32 return ST->hasV6T2Ops() ? 2 : 3;
33 }
34 if (ST->isThumb2()) {
35 if ((SImmVal >= 0 && SImmVal < 65536) ||
36 (ARM_AM::getT2SOImmVal(ZImmVal) != -1) ||
37 (ARM_AM::getT2SOImmVal(~ZImmVal) != -1))
38 return 1;
39 return ST->hasV6T2Ops() ? 2 : 3;
40 }
41 // Thumb1.
42 if (SImmVal >= 0 && SImmVal < 256)
43 return 1;
44 if ((~ZImmVal < 256) || ARM_AM::isThumbImmShiftedVal(ZImmVal))
45 return 2;
46 // Load from constantpool.
47 return 3;
48 }
49
getCastInstrCost(unsigned Opcode,Type * Dst,Type * Src)50 unsigned ARMTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) {
51 int ISD = TLI->InstructionOpcodeToISD(Opcode);
52 assert(ISD && "Invalid opcode");
53
54 // Single to/from double precision conversions.
55 static const CostTblEntry<MVT::SimpleValueType> NEONFltDblTbl[] = {
56 // Vector fptrunc/fpext conversions.
57 { ISD::FP_ROUND, MVT::v2f64, 2 },
58 { ISD::FP_EXTEND, MVT::v2f32, 2 },
59 { ISD::FP_EXTEND, MVT::v4f32, 4 }
60 };
61
62 if (Src->isVectorTy() && ST->hasNEON() && (ISD == ISD::FP_ROUND ||
63 ISD == ISD::FP_EXTEND)) {
64 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
65 int Idx = CostTableLookup(NEONFltDblTbl, ISD, LT.second);
66 if (Idx != -1)
67 return LT.first * NEONFltDblTbl[Idx].Cost;
68 }
69
70 EVT SrcTy = TLI->getValueType(DL, Src);
71 EVT DstTy = TLI->getValueType(DL, Dst);
72
73 if (!SrcTy.isSimple() || !DstTy.isSimple())
74 return BaseT::getCastInstrCost(Opcode, Dst, Src);
75
76 // Some arithmetic, load and store operations have specific instructions
77 // to cast up/down their types automatically at no extra cost.
78 // TODO: Get these tables to know at least what the related operations are.
79 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
80 NEONVectorConversionTbl[] = {
81 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 0 },
82 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 0 },
83 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i32, 1 },
84 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i32, 1 },
85 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 0 },
86 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 1 },
87
88 // The number of vmovl instructions for the extension.
89 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
90 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
91 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
92 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
93 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i8, 7 },
94 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i8, 7 },
95 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i16, 6 },
96 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i16, 6 },
97 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 6 },
98 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 6 },
99
100 // Operations that we legalize using splitting.
101 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 6 },
102 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 3 },
103
104 // Vector float <-> i32 conversions.
105 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
106 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
107
108 { ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i8, 3 },
109 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i8, 3 },
110 { ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i16, 2 },
111 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i16, 2 },
112 { ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i32, 1 },
113 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 1 },
114 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 },
115 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 },
116 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 },
117 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 },
118 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 },
119 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 },
120 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 4 },
121 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 4 },
122 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 2 },
123 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 2 },
124 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i16, 8 },
125 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i16, 8 },
126 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i32, 4 },
127 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i32, 4 },
128
129 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 1 },
130 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1 },
131 { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 3 },
132 { ISD::FP_TO_UINT, MVT::v4i8, MVT::v4f32, 3 },
133 { ISD::FP_TO_SINT, MVT::v4i16, MVT::v4f32, 2 },
134 { ISD::FP_TO_UINT, MVT::v4i16, MVT::v4f32, 2 },
135
136 // Vector double <-> i32 conversions.
137 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 },
138 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 },
139
140 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i8, 4 },
141 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i8, 4 },
142 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i16, 3 },
143 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i16, 3 },
144 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 },
145 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 },
146
147 { ISD::FP_TO_SINT, MVT::v2i32, MVT::v2f64, 2 },
148 { ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f64, 2 },
149 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f32, 4 },
150 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f32, 4 },
151 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v16f32, 8 },
152 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v16f32, 8 }
153 };
154
155 if (SrcTy.isVector() && ST->hasNEON()) {
156 int Idx = ConvertCostTableLookup(NEONVectorConversionTbl, ISD,
157 DstTy.getSimpleVT(), SrcTy.getSimpleVT());
158 if (Idx != -1)
159 return NEONVectorConversionTbl[Idx].Cost;
160 }
161
162 // Scalar float to integer conversions.
163 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
164 NEONFloatConversionTbl[] = {
165 { ISD::FP_TO_SINT, MVT::i1, MVT::f32, 2 },
166 { ISD::FP_TO_UINT, MVT::i1, MVT::f32, 2 },
167 { ISD::FP_TO_SINT, MVT::i1, MVT::f64, 2 },
168 { ISD::FP_TO_UINT, MVT::i1, MVT::f64, 2 },
169 { ISD::FP_TO_SINT, MVT::i8, MVT::f32, 2 },
170 { ISD::FP_TO_UINT, MVT::i8, MVT::f32, 2 },
171 { ISD::FP_TO_SINT, MVT::i8, MVT::f64, 2 },
172 { ISD::FP_TO_UINT, MVT::i8, MVT::f64, 2 },
173 { ISD::FP_TO_SINT, MVT::i16, MVT::f32, 2 },
174 { ISD::FP_TO_UINT, MVT::i16, MVT::f32, 2 },
175 { ISD::FP_TO_SINT, MVT::i16, MVT::f64, 2 },
176 { ISD::FP_TO_UINT, MVT::i16, MVT::f64, 2 },
177 { ISD::FP_TO_SINT, MVT::i32, MVT::f32, 2 },
178 { ISD::FP_TO_UINT, MVT::i32, MVT::f32, 2 },
179 { ISD::FP_TO_SINT, MVT::i32, MVT::f64, 2 },
180 { ISD::FP_TO_UINT, MVT::i32, MVT::f64, 2 },
181 { ISD::FP_TO_SINT, MVT::i64, MVT::f32, 10 },
182 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 10 },
183 { ISD::FP_TO_SINT, MVT::i64, MVT::f64, 10 },
184 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 10 }
185 };
186 if (SrcTy.isFloatingPoint() && ST->hasNEON()) {
187 int Idx = ConvertCostTableLookup(NEONFloatConversionTbl, ISD,
188 DstTy.getSimpleVT(), SrcTy.getSimpleVT());
189 if (Idx != -1)
190 return NEONFloatConversionTbl[Idx].Cost;
191 }
192
193 // Scalar integer to float conversions.
194 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
195 NEONIntegerConversionTbl[] = {
196 { ISD::SINT_TO_FP, MVT::f32, MVT::i1, 2 },
197 { ISD::UINT_TO_FP, MVT::f32, MVT::i1, 2 },
198 { ISD::SINT_TO_FP, MVT::f64, MVT::i1, 2 },
199 { ISD::UINT_TO_FP, MVT::f64, MVT::i1, 2 },
200 { ISD::SINT_TO_FP, MVT::f32, MVT::i8, 2 },
201 { ISD::UINT_TO_FP, MVT::f32, MVT::i8, 2 },
202 { ISD::SINT_TO_FP, MVT::f64, MVT::i8, 2 },
203 { ISD::UINT_TO_FP, MVT::f64, MVT::i8, 2 },
204 { ISD::SINT_TO_FP, MVT::f32, MVT::i16, 2 },
205 { ISD::UINT_TO_FP, MVT::f32, MVT::i16, 2 },
206 { ISD::SINT_TO_FP, MVT::f64, MVT::i16, 2 },
207 { ISD::UINT_TO_FP, MVT::f64, MVT::i16, 2 },
208 { ISD::SINT_TO_FP, MVT::f32, MVT::i32, 2 },
209 { ISD::UINT_TO_FP, MVT::f32, MVT::i32, 2 },
210 { ISD::SINT_TO_FP, MVT::f64, MVT::i32, 2 },
211 { ISD::UINT_TO_FP, MVT::f64, MVT::i32, 2 },
212 { ISD::SINT_TO_FP, MVT::f32, MVT::i64, 10 },
213 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 10 },
214 { ISD::SINT_TO_FP, MVT::f64, MVT::i64, 10 },
215 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 10 }
216 };
217
218 if (SrcTy.isInteger() && ST->hasNEON()) {
219 int Idx = ConvertCostTableLookup(NEONIntegerConversionTbl, ISD,
220 DstTy.getSimpleVT(), SrcTy.getSimpleVT());
221 if (Idx != -1)
222 return NEONIntegerConversionTbl[Idx].Cost;
223 }
224
225 // Scalar integer conversion costs.
226 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
227 ARMIntegerConversionTbl[] = {
228 // i16 -> i64 requires two dependent operations.
229 { ISD::SIGN_EXTEND, MVT::i64, MVT::i16, 2 },
230
231 // Truncates on i64 are assumed to be free.
232 { ISD::TRUNCATE, MVT::i32, MVT::i64, 0 },
233 { ISD::TRUNCATE, MVT::i16, MVT::i64, 0 },
234 { ISD::TRUNCATE, MVT::i8, MVT::i64, 0 },
235 { ISD::TRUNCATE, MVT::i1, MVT::i64, 0 }
236 };
237
238 if (SrcTy.isInteger()) {
239 int Idx = ConvertCostTableLookup(ARMIntegerConversionTbl, ISD,
240 DstTy.getSimpleVT(), SrcTy.getSimpleVT());
241 if (Idx != -1)
242 return ARMIntegerConversionTbl[Idx].Cost;
243 }
244
245 return BaseT::getCastInstrCost(Opcode, Dst, Src);
246 }
247
getVectorInstrCost(unsigned Opcode,Type * ValTy,unsigned Index)248 unsigned ARMTTIImpl::getVectorInstrCost(unsigned Opcode, Type *ValTy,
249 unsigned Index) {
250 // Penalize inserting into an D-subregister. We end up with a three times
251 // lower estimated throughput on swift.
252 if (ST->isSwift() &&
253 Opcode == Instruction::InsertElement &&
254 ValTy->isVectorTy() &&
255 ValTy->getScalarSizeInBits() <= 32)
256 return 3;
257
258 // Cross-class copies are expensive on many microarchitectures,
259 // so assume they are expensive by default.
260 if ((Opcode == Instruction::InsertElement ||
261 Opcode == Instruction::ExtractElement) &&
262 ValTy->getVectorElementType()->isIntegerTy())
263 return 3;
264
265 return BaseT::getVectorInstrCost(Opcode, ValTy, Index);
266 }
267
getCmpSelInstrCost(unsigned Opcode,Type * ValTy,Type * CondTy)268 unsigned ARMTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
269 Type *CondTy) {
270
271 int ISD = TLI->InstructionOpcodeToISD(Opcode);
272 // On NEON a a vector select gets lowered to vbsl.
273 if (ST->hasNEON() && ValTy->isVectorTy() && ISD == ISD::SELECT) {
274 // Lowering of some vector selects is currently far from perfect.
275 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
276 NEONVectorSelectTbl[] = {
277 { ISD::SELECT, MVT::v16i1, MVT::v16i16, 2*16 + 1 + 3*1 + 4*1 },
278 { ISD::SELECT, MVT::v8i1, MVT::v8i32, 4*8 + 1*3 + 1*4 + 1*2 },
279 { ISD::SELECT, MVT::v16i1, MVT::v16i32, 4*16 + 1*6 + 1*8 + 1*4 },
280 { ISD::SELECT, MVT::v4i1, MVT::v4i64, 4*4 + 1*2 + 1 },
281 { ISD::SELECT, MVT::v8i1, MVT::v8i64, 50 },
282 { ISD::SELECT, MVT::v16i1, MVT::v16i64, 100 }
283 };
284
285 EVT SelCondTy = TLI->getValueType(DL, CondTy);
286 EVT SelValTy = TLI->getValueType(DL, ValTy);
287 if (SelCondTy.isSimple() && SelValTy.isSimple()) {
288 int Idx = ConvertCostTableLookup(NEONVectorSelectTbl, ISD,
289 SelCondTy.getSimpleVT(),
290 SelValTy.getSimpleVT());
291 if (Idx != -1)
292 return NEONVectorSelectTbl[Idx].Cost;
293 }
294
295 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
296 return LT.first;
297 }
298
299 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy);
300 }
301
getAddressComputationCost(Type * Ty,bool IsComplex)302 unsigned ARMTTIImpl::getAddressComputationCost(Type *Ty, bool IsComplex) {
303 // Address computations in vectorized code with non-consecutive addresses will
304 // likely result in more instructions compared to scalar code where the
305 // computation can more often be merged into the index mode. The resulting
306 // extra micro-ops can significantly decrease throughput.
307 unsigned NumVectorInstToHideOverhead = 10;
308
309 if (Ty->isVectorTy() && IsComplex)
310 return NumVectorInstToHideOverhead;
311
312 // In many cases the address computation is not merged into the instruction
313 // addressing mode.
314 return 1;
315 }
316
getFPOpCost(Type * Ty)317 unsigned ARMTTIImpl::getFPOpCost(Type *Ty) {
318 // Use similar logic that's in ARMISelLowering:
319 // Any ARM CPU with VFP2 has floating point, but Thumb1 didn't have access
320 // to VFP.
321
322 if (ST->hasVFP2() && !ST->isThumb1Only()) {
323 if (Ty->isFloatTy()) {
324 return TargetTransformInfo::TCC_Basic;
325 }
326
327 if (Ty->isDoubleTy()) {
328 return ST->isFPOnlySP() ? TargetTransformInfo::TCC_Expensive :
329 TargetTransformInfo::TCC_Basic;
330 }
331 }
332
333 return TargetTransformInfo::TCC_Expensive;
334 }
335
getShuffleCost(TTI::ShuffleKind Kind,Type * Tp,int Index,Type * SubTp)336 unsigned ARMTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
337 Type *SubTp) {
338 // We only handle costs of reverse and alternate shuffles for now.
339 if (Kind != TTI::SK_Reverse && Kind != TTI::SK_Alternate)
340 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
341
342 if (Kind == TTI::SK_Reverse) {
343 static const CostTblEntry<MVT::SimpleValueType> NEONShuffleTbl[] = {
344 // Reverse shuffle cost one instruction if we are shuffling within a
345 // double word (vrev) or two if we shuffle a quad word (vrev, vext).
346 {ISD::VECTOR_SHUFFLE, MVT::v2i32, 1},
347 {ISD::VECTOR_SHUFFLE, MVT::v2f32, 1},
348 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
349 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
350
351 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
352 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
353 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 2},
354 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 2}};
355
356 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
357
358 int Idx = CostTableLookup(NEONShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
359 if (Idx == -1)
360 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
361
362 return LT.first * NEONShuffleTbl[Idx].Cost;
363 }
364 if (Kind == TTI::SK_Alternate) {
365 static const CostTblEntry<MVT::SimpleValueType> NEONAltShuffleTbl[] = {
366 // Alt shuffle cost table for ARM. Cost is the number of instructions
367 // required to create the shuffled vector.
368
369 {ISD::VECTOR_SHUFFLE, MVT::v2f32, 1},
370 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
371 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
372 {ISD::VECTOR_SHUFFLE, MVT::v2i32, 1},
373
374 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
375 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
376 {ISD::VECTOR_SHUFFLE, MVT::v4i16, 2},
377
378 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 16},
379
380 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 32}};
381
382 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
383 int Idx =
384 CostTableLookup(NEONAltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
385 if (Idx == -1)
386 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
387 return LT.first * NEONAltShuffleTbl[Idx].Cost;
388 }
389 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
390 }
391
getArithmeticInstrCost(unsigned Opcode,Type * Ty,TTI::OperandValueKind Op1Info,TTI::OperandValueKind Op2Info,TTI::OperandValueProperties Opd1PropInfo,TTI::OperandValueProperties Opd2PropInfo)392 unsigned ARMTTIImpl::getArithmeticInstrCost(
393 unsigned Opcode, Type *Ty, TTI::OperandValueKind Op1Info,
394 TTI::OperandValueKind Op2Info, TTI::OperandValueProperties Opd1PropInfo,
395 TTI::OperandValueProperties Opd2PropInfo) {
396
397 int ISDOpcode = TLI->InstructionOpcodeToISD(Opcode);
398 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
399
400 const unsigned FunctionCallDivCost = 20;
401 const unsigned ReciprocalDivCost = 10;
402 static const CostTblEntry<MVT::SimpleValueType> CostTbl[] = {
403 // Division.
404 // These costs are somewhat random. Choose a cost of 20 to indicate that
405 // vectorizing devision (added function call) is going to be very expensive.
406 // Double registers types.
407 { ISD::SDIV, MVT::v1i64, 1 * FunctionCallDivCost},
408 { ISD::UDIV, MVT::v1i64, 1 * FunctionCallDivCost},
409 { ISD::SREM, MVT::v1i64, 1 * FunctionCallDivCost},
410 { ISD::UREM, MVT::v1i64, 1 * FunctionCallDivCost},
411 { ISD::SDIV, MVT::v2i32, 2 * FunctionCallDivCost},
412 { ISD::UDIV, MVT::v2i32, 2 * FunctionCallDivCost},
413 { ISD::SREM, MVT::v2i32, 2 * FunctionCallDivCost},
414 { ISD::UREM, MVT::v2i32, 2 * FunctionCallDivCost},
415 { ISD::SDIV, MVT::v4i16, ReciprocalDivCost},
416 { ISD::UDIV, MVT::v4i16, ReciprocalDivCost},
417 { ISD::SREM, MVT::v4i16, 4 * FunctionCallDivCost},
418 { ISD::UREM, MVT::v4i16, 4 * FunctionCallDivCost},
419 { ISD::SDIV, MVT::v8i8, ReciprocalDivCost},
420 { ISD::UDIV, MVT::v8i8, ReciprocalDivCost},
421 { ISD::SREM, MVT::v8i8, 8 * FunctionCallDivCost},
422 { ISD::UREM, MVT::v8i8, 8 * FunctionCallDivCost},
423 // Quad register types.
424 { ISD::SDIV, MVT::v2i64, 2 * FunctionCallDivCost},
425 { ISD::UDIV, MVT::v2i64, 2 * FunctionCallDivCost},
426 { ISD::SREM, MVT::v2i64, 2 * FunctionCallDivCost},
427 { ISD::UREM, MVT::v2i64, 2 * FunctionCallDivCost},
428 { ISD::SDIV, MVT::v4i32, 4 * FunctionCallDivCost},
429 { ISD::UDIV, MVT::v4i32, 4 * FunctionCallDivCost},
430 { ISD::SREM, MVT::v4i32, 4 * FunctionCallDivCost},
431 { ISD::UREM, MVT::v4i32, 4 * FunctionCallDivCost},
432 { ISD::SDIV, MVT::v8i16, 8 * FunctionCallDivCost},
433 { ISD::UDIV, MVT::v8i16, 8 * FunctionCallDivCost},
434 { ISD::SREM, MVT::v8i16, 8 * FunctionCallDivCost},
435 { ISD::UREM, MVT::v8i16, 8 * FunctionCallDivCost},
436 { ISD::SDIV, MVT::v16i8, 16 * FunctionCallDivCost},
437 { ISD::UDIV, MVT::v16i8, 16 * FunctionCallDivCost},
438 { ISD::SREM, MVT::v16i8, 16 * FunctionCallDivCost},
439 { ISD::UREM, MVT::v16i8, 16 * FunctionCallDivCost},
440 // Multiplication.
441 };
442
443 int Idx = -1;
444
445 if (ST->hasNEON())
446 Idx = CostTableLookup(CostTbl, ISDOpcode, LT.second);
447
448 if (Idx != -1)
449 return LT.first * CostTbl[Idx].Cost;
450
451 unsigned Cost = BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info,
452 Opd1PropInfo, Opd2PropInfo);
453
454 // This is somewhat of a hack. The problem that we are facing is that SROA
455 // creates a sequence of shift, and, or instructions to construct values.
456 // These sequences are recognized by the ISel and have zero-cost. Not so for
457 // the vectorized code. Because we have support for v2i64 but not i64 those
458 // sequences look particularly beneficial to vectorize.
459 // To work around this we increase the cost of v2i64 operations to make them
460 // seem less beneficial.
461 if (LT.second == MVT::v2i64 &&
462 Op2Info == TargetTransformInfo::OK_UniformConstantValue)
463 Cost += 4;
464
465 return Cost;
466 }
467
getMemoryOpCost(unsigned Opcode,Type * Src,unsigned Alignment,unsigned AddressSpace)468 unsigned ARMTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
469 unsigned Alignment,
470 unsigned AddressSpace) {
471 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
472
473 if (Src->isVectorTy() && Alignment != 16 &&
474 Src->getVectorElementType()->isDoubleTy()) {
475 // Unaligned loads/stores are extremely inefficient.
476 // We need 4 uops for vst.1/vld.1 vs 1uop for vldr/vstr.
477 return LT.first * 4;
478 }
479 return LT.first;
480 }
481
getInterleavedMemoryOpCost(unsigned Opcode,Type * VecTy,unsigned Factor,ArrayRef<unsigned> Indices,unsigned Alignment,unsigned AddressSpace)482 unsigned ARMTTIImpl::getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
483 unsigned Factor,
484 ArrayRef<unsigned> Indices,
485 unsigned Alignment,
486 unsigned AddressSpace) {
487 assert(Factor >= 2 && "Invalid interleave factor");
488 assert(isa<VectorType>(VecTy) && "Expect a vector type");
489
490 // vldN/vstN doesn't support vector types of i64/f64 element.
491 bool EltIs64Bits = DL.getTypeAllocSizeInBits(VecTy->getScalarType()) == 64;
492
493 if (Factor <= TLI->getMaxSupportedInterleaveFactor() && !EltIs64Bits) {
494 unsigned NumElts = VecTy->getVectorNumElements();
495 Type *SubVecTy = VectorType::get(VecTy->getScalarType(), NumElts / Factor);
496 unsigned SubVecSize = DL.getTypeAllocSize(SubVecTy);
497
498 // vldN/vstN only support legal vector types of size 64 or 128 in bits.
499 if (NumElts % Factor == 0 && (SubVecSize == 64 || SubVecSize == 128))
500 return Factor;
501 }
502
503 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
504 Alignment, AddressSpace);
505 }
506