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