1 //===-- X86TargetTransformInfo.cpp - X86 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 //===----------------------------------------------------------------------===//
9 /// \file
10 /// This file implements a TargetTransformInfo analysis pass specific to the
11 /// X86 target machine. It uses the target's detailed information to provide
12 /// more precise answers to certain TTI queries, while letting the target
13 /// independent and default TTI implementations handle the rest.
14 ///
15 //===----------------------------------------------------------------------===//
16
17 #include "X86.h"
18 #include "X86TargetMachine.h"
19 #include "llvm/Analysis/TargetTransformInfo.h"
20 #include "llvm/IR/IntrinsicInst.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Target/CostTable.h"
23 #include "llvm/Target/TargetLowering.h"
24 using namespace llvm;
25
26 #define DEBUG_TYPE "x86tti"
27
28 // Declare the pass initialization routine locally as target-specific passes
29 // don't have a target-wide initialization entry point, and so we rely on the
30 // pass constructor initialization.
31 namespace llvm {
32 void initializeX86TTIPass(PassRegistry &);
33 }
34
35 namespace {
36
37 class X86TTI final : public ImmutablePass, public TargetTransformInfo {
38 const X86Subtarget *ST;
39 const X86TargetLowering *TLI;
40
41 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
42 /// are set if the result needs to be inserted and/or extracted from vectors.
43 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const;
44
45 public:
X86TTI()46 X86TTI() : ImmutablePass(ID), ST(nullptr), TLI(nullptr) {
47 llvm_unreachable("This pass cannot be directly constructed");
48 }
49
X86TTI(const X86TargetMachine * TM)50 X86TTI(const X86TargetMachine *TM)
51 : ImmutablePass(ID), ST(TM->getSubtargetImpl()),
52 TLI(TM->getSubtargetImpl()->getTargetLowering()) {
53 initializeX86TTIPass(*PassRegistry::getPassRegistry());
54 }
55
initializePass()56 void initializePass() override {
57 pushTTIStack(this);
58 }
59
getAnalysisUsage(AnalysisUsage & AU) const60 void getAnalysisUsage(AnalysisUsage &AU) const override {
61 TargetTransformInfo::getAnalysisUsage(AU);
62 }
63
64 /// Pass identification.
65 static char ID;
66
67 /// Provide necessary pointer adjustments for the two base classes.
getAdjustedAnalysisPointer(const void * ID)68 void *getAdjustedAnalysisPointer(const void *ID) override {
69 if (ID == &TargetTransformInfo::ID)
70 return (TargetTransformInfo*)this;
71 return this;
72 }
73
74 /// \name Scalar TTI Implementations
75 /// @{
76 PopcntSupportKind getPopcntSupport(unsigned TyWidth) const override;
77
78 /// @}
79
80 /// \name Vector TTI Implementations
81 /// @{
82
83 unsigned getNumberOfRegisters(bool Vector) const override;
84 unsigned getRegisterBitWidth(bool Vector) const override;
85 unsigned getMaxInterleaveFactor() const override;
86 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind,
87 OperandValueKind, OperandValueProperties,
88 OperandValueProperties) const override;
89 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp,
90 int Index, Type *SubTp) const override;
91 unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
92 Type *Src) const override;
93 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
94 Type *CondTy) const override;
95 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
96 unsigned Index) const override;
97 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
98 unsigned AddressSpace) const override;
99
100 unsigned getAddressComputationCost(Type *PtrTy,
101 bool IsComplex) const override;
102
103 unsigned getReductionCost(unsigned Opcode, Type *Ty,
104 bool IsPairwiseForm) const override;
105
106 unsigned getIntImmCost(int64_t) const;
107
108 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const override;
109
110 unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
111 Type *Ty) const override;
112 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
113 Type *Ty) const override;
114 bool isLegalMaskedLoad (Type *DataType, int Consecutive) const override;
115 bool isLegalMaskedStore(Type *DataType, int Consecutive) const override;
116
117 /// @}
118 };
119
120 } // end anonymous namespace
121
122 INITIALIZE_AG_PASS(X86TTI, TargetTransformInfo, "x86tti",
123 "X86 Target Transform Info", true, true, false)
124 char X86TTI::ID = 0;
125
126 ImmutablePass *
createX86TargetTransformInfoPass(const X86TargetMachine * TM)127 llvm::createX86TargetTransformInfoPass(const X86TargetMachine *TM) {
128 return new X86TTI(TM);
129 }
130
131
132 //===----------------------------------------------------------------------===//
133 //
134 // X86 cost model.
135 //
136 //===----------------------------------------------------------------------===//
137
getPopcntSupport(unsigned TyWidth) const138 X86TTI::PopcntSupportKind X86TTI::getPopcntSupport(unsigned TyWidth) const {
139 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
140 // TODO: Currently the __builtin_popcount() implementation using SSE3
141 // instructions is inefficient. Once the problem is fixed, we should
142 // call ST->hasSSE3() instead of ST->hasPOPCNT().
143 return ST->hasPOPCNT() ? PSK_FastHardware : PSK_Software;
144 }
145
getNumberOfRegisters(bool Vector) const146 unsigned X86TTI::getNumberOfRegisters(bool Vector) const {
147 if (Vector && !ST->hasSSE1())
148 return 0;
149
150 if (ST->is64Bit()) {
151 if (Vector && ST->hasAVX512())
152 return 32;
153 return 16;
154 }
155 return 8;
156 }
157
getRegisterBitWidth(bool Vector) const158 unsigned X86TTI::getRegisterBitWidth(bool Vector) const {
159 if (Vector) {
160 if (ST->hasAVX512()) return 512;
161 if (ST->hasAVX()) return 256;
162 if (ST->hasSSE1()) return 128;
163 return 0;
164 }
165
166 if (ST->is64Bit())
167 return 64;
168 return 32;
169
170 }
171
getMaxInterleaveFactor() const172 unsigned X86TTI::getMaxInterleaveFactor() const {
173 if (ST->isAtom())
174 return 1;
175
176 // Sandybridge and Haswell have multiple execution ports and pipelined
177 // vector units.
178 if (ST->hasAVX())
179 return 4;
180
181 return 2;
182 }
183
getArithmeticInstrCost(unsigned Opcode,Type * Ty,OperandValueKind Op1Info,OperandValueKind Op2Info,OperandValueProperties Opd1PropInfo,OperandValueProperties Opd2PropInfo) const184 unsigned X86TTI::getArithmeticInstrCost(
185 unsigned Opcode, Type *Ty, OperandValueKind Op1Info,
186 OperandValueKind Op2Info, OperandValueProperties Opd1PropInfo,
187 OperandValueProperties Opd2PropInfo) const {
188 // Legalize the type.
189 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty);
190
191 int ISD = TLI->InstructionOpcodeToISD(Opcode);
192 assert(ISD && "Invalid opcode");
193
194 if (ISD == ISD::SDIV &&
195 Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
196 Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
197 // On X86, vector signed division by constants power-of-two are
198 // normally expanded to the sequence SRA + SRL + ADD + SRA.
199 // The OperandValue properties many not be same as that of previous
200 // operation;conservatively assume OP_None.
201 unsigned Cost =
202 2 * getArithmeticInstrCost(Instruction::AShr, Ty, Op1Info, Op2Info,
203 TargetTransformInfo::OP_None,
204 TargetTransformInfo::OP_None);
205 Cost += getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info,
206 TargetTransformInfo::OP_None,
207 TargetTransformInfo::OP_None);
208 Cost += getArithmeticInstrCost(Instruction::Add, Ty, Op1Info, Op2Info,
209 TargetTransformInfo::OP_None,
210 TargetTransformInfo::OP_None);
211
212 return Cost;
213 }
214
215 static const CostTblEntry<MVT::SimpleValueType>
216 AVX2UniformConstCostTable[] = {
217 { ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence
218 { ISD::UDIV, MVT::v16i16, 6 }, // vpmulhuw sequence
219 { ISD::SDIV, MVT::v8i32, 15 }, // vpmuldq sequence
220 { ISD::UDIV, MVT::v8i32, 15 }, // vpmuludq sequence
221 };
222
223 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
224 ST->hasAVX2()) {
225 int Idx = CostTableLookup(AVX2UniformConstCostTable, ISD, LT.second);
226 if (Idx != -1)
227 return LT.first * AVX2UniformConstCostTable[Idx].Cost;
228 }
229
230 static const CostTblEntry<MVT::SimpleValueType> AVX512CostTable[] = {
231 { ISD::SHL, MVT::v16i32, 1 },
232 { ISD::SRL, MVT::v16i32, 1 },
233 { ISD::SRA, MVT::v16i32, 1 },
234 { ISD::SHL, MVT::v8i64, 1 },
235 { ISD::SRL, MVT::v8i64, 1 },
236 { ISD::SRA, MVT::v8i64, 1 },
237 };
238
239 static const CostTblEntry<MVT::SimpleValueType> AVX2CostTable[] = {
240 // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
241 // customize them to detect the cases where shift amount is a scalar one.
242 { ISD::SHL, MVT::v4i32, 1 },
243 { ISD::SRL, MVT::v4i32, 1 },
244 { ISD::SRA, MVT::v4i32, 1 },
245 { ISD::SHL, MVT::v8i32, 1 },
246 { ISD::SRL, MVT::v8i32, 1 },
247 { ISD::SRA, MVT::v8i32, 1 },
248 { ISD::SHL, MVT::v2i64, 1 },
249 { ISD::SRL, MVT::v2i64, 1 },
250 { ISD::SHL, MVT::v4i64, 1 },
251 { ISD::SRL, MVT::v4i64, 1 },
252
253 { ISD::SHL, MVT::v32i8, 42 }, // cmpeqb sequence.
254 { ISD::SHL, MVT::v16i16, 16*10 }, // Scalarized.
255
256 { ISD::SRL, MVT::v32i8, 32*10 }, // Scalarized.
257 { ISD::SRL, MVT::v16i16, 8*10 }, // Scalarized.
258
259 { ISD::SRA, MVT::v32i8, 32*10 }, // Scalarized.
260 { ISD::SRA, MVT::v16i16, 16*10 }, // Scalarized.
261 { ISD::SRA, MVT::v4i64, 4*10 }, // Scalarized.
262
263 // Vectorizing division is a bad idea. See the SSE2 table for more comments.
264 { ISD::SDIV, MVT::v32i8, 32*20 },
265 { ISD::SDIV, MVT::v16i16, 16*20 },
266 { ISD::SDIV, MVT::v8i32, 8*20 },
267 { ISD::SDIV, MVT::v4i64, 4*20 },
268 { ISD::UDIV, MVT::v32i8, 32*20 },
269 { ISD::UDIV, MVT::v16i16, 16*20 },
270 { ISD::UDIV, MVT::v8i32, 8*20 },
271 { ISD::UDIV, MVT::v4i64, 4*20 },
272 };
273
274 if (ST->hasAVX512()) {
275 int Idx = CostTableLookup(AVX512CostTable, ISD, LT.second);
276 if (Idx != -1)
277 return LT.first * AVX512CostTable[Idx].Cost;
278 }
279 // Look for AVX2 lowering tricks.
280 if (ST->hasAVX2()) {
281 if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
282 (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
283 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
284 // On AVX2, a packed v16i16 shift left by a constant build_vector
285 // is lowered into a vector multiply (vpmullw).
286 return LT.first;
287
288 int Idx = CostTableLookup(AVX2CostTable, ISD, LT.second);
289 if (Idx != -1)
290 return LT.first * AVX2CostTable[Idx].Cost;
291 }
292
293 static const CostTblEntry<MVT::SimpleValueType>
294 SSE2UniformConstCostTable[] = {
295 // We don't correctly identify costs of casts because they are marked as
296 // custom.
297 // Constant splats are cheaper for the following instructions.
298 { ISD::SHL, MVT::v16i8, 1 }, // psllw.
299 { ISD::SHL, MVT::v8i16, 1 }, // psllw.
300 { ISD::SHL, MVT::v4i32, 1 }, // pslld
301 { ISD::SHL, MVT::v2i64, 1 }, // psllq.
302
303 { ISD::SRL, MVT::v16i8, 1 }, // psrlw.
304 { ISD::SRL, MVT::v8i16, 1 }, // psrlw.
305 { ISD::SRL, MVT::v4i32, 1 }, // psrld.
306 { ISD::SRL, MVT::v2i64, 1 }, // psrlq.
307
308 { ISD::SRA, MVT::v16i8, 4 }, // psrlw, pand, pxor, psubb.
309 { ISD::SRA, MVT::v8i16, 1 }, // psraw.
310 { ISD::SRA, MVT::v4i32, 1 }, // psrad.
311
312 { ISD::SDIV, MVT::v8i16, 6 }, // pmulhw sequence
313 { ISD::UDIV, MVT::v8i16, 6 }, // pmulhuw sequence
314 { ISD::SDIV, MVT::v4i32, 19 }, // pmuludq sequence
315 { ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence
316 };
317
318 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
319 ST->hasSSE2()) {
320 // pmuldq sequence.
321 if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
322 return LT.first * 15;
323
324 int Idx = CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second);
325 if (Idx != -1)
326 return LT.first * SSE2UniformConstCostTable[Idx].Cost;
327 }
328
329 if (ISD == ISD::SHL &&
330 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
331 EVT VT = LT.second;
332 if ((VT == MVT::v8i16 && ST->hasSSE2()) ||
333 (VT == MVT::v4i32 && ST->hasSSE41()))
334 // Vector shift left by non uniform constant can be lowered
335 // into vector multiply (pmullw/pmulld).
336 return LT.first;
337 if (VT == MVT::v4i32 && ST->hasSSE2())
338 // A vector shift left by non uniform constant is converted
339 // into a vector multiply; the new multiply is eventually
340 // lowered into a sequence of shuffles and 2 x pmuludq.
341 ISD = ISD::MUL;
342 }
343
344 static const CostTblEntry<MVT::SimpleValueType> SSE2CostTable[] = {
345 // We don't correctly identify costs of casts because they are marked as
346 // custom.
347 // For some cases, where the shift amount is a scalar we would be able
348 // to generate better code. Unfortunately, when this is the case the value
349 // (the splat) will get hoisted out of the loop, thereby making it invisible
350 // to ISel. The cost model must return worst case assumptions because it is
351 // used for vectorization and we don't want to make vectorized code worse
352 // than scalar code.
353 { ISD::SHL, MVT::v16i8, 30 }, // cmpeqb sequence.
354 { ISD::SHL, MVT::v8i16, 8*10 }, // Scalarized.
355 { ISD::SHL, MVT::v4i32, 2*5 }, // We optimized this using mul.
356 { ISD::SHL, MVT::v2i64, 2*10 }, // Scalarized.
357 { ISD::SHL, MVT::v4i64, 4*10 }, // Scalarized.
358
359 { ISD::SRL, MVT::v16i8, 16*10 }, // Scalarized.
360 { ISD::SRL, MVT::v8i16, 8*10 }, // Scalarized.
361 { ISD::SRL, MVT::v4i32, 4*10 }, // Scalarized.
362 { ISD::SRL, MVT::v2i64, 2*10 }, // Scalarized.
363
364 { ISD::SRA, MVT::v16i8, 16*10 }, // Scalarized.
365 { ISD::SRA, MVT::v8i16, 8*10 }, // Scalarized.
366 { ISD::SRA, MVT::v4i32, 4*10 }, // Scalarized.
367 { ISD::SRA, MVT::v2i64, 2*10 }, // Scalarized.
368
369 // It is not a good idea to vectorize division. We have to scalarize it and
370 // in the process we will often end up having to spilling regular
371 // registers. The overhead of division is going to dominate most kernels
372 // anyways so try hard to prevent vectorization of division - it is
373 // generally a bad idea. Assume somewhat arbitrarily that we have to be able
374 // to hide "20 cycles" for each lane.
375 { ISD::SDIV, MVT::v16i8, 16*20 },
376 { ISD::SDIV, MVT::v8i16, 8*20 },
377 { ISD::SDIV, MVT::v4i32, 4*20 },
378 { ISD::SDIV, MVT::v2i64, 2*20 },
379 { ISD::UDIV, MVT::v16i8, 16*20 },
380 { ISD::UDIV, MVT::v8i16, 8*20 },
381 { ISD::UDIV, MVT::v4i32, 4*20 },
382 { ISD::UDIV, MVT::v2i64, 2*20 },
383 };
384
385 if (ST->hasSSE2()) {
386 int Idx = CostTableLookup(SSE2CostTable, ISD, LT.second);
387 if (Idx != -1)
388 return LT.first * SSE2CostTable[Idx].Cost;
389 }
390
391 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTable[] = {
392 // We don't have to scalarize unsupported ops. We can issue two half-sized
393 // operations and we only need to extract the upper YMM half.
394 // Two ops + 1 extract + 1 insert = 4.
395 { ISD::MUL, MVT::v16i16, 4 },
396 { ISD::MUL, MVT::v8i32, 4 },
397 { ISD::SUB, MVT::v8i32, 4 },
398 { ISD::ADD, MVT::v8i32, 4 },
399 { ISD::SUB, MVT::v4i64, 4 },
400 { ISD::ADD, MVT::v4i64, 4 },
401 // A v4i64 multiply is custom lowered as two split v2i64 vectors that then
402 // are lowered as a series of long multiplies(3), shifts(4) and adds(2)
403 // Because we believe v4i64 to be a legal type, we must also include the
404 // split factor of two in the cost table. Therefore, the cost here is 18
405 // instead of 9.
406 { ISD::MUL, MVT::v4i64, 18 },
407 };
408
409 // Look for AVX1 lowering tricks.
410 if (ST->hasAVX() && !ST->hasAVX2()) {
411 EVT VT = LT.second;
412
413 // v16i16 and v8i32 shifts by non-uniform constants are lowered into a
414 // sequence of extract + two vector multiply + insert.
415 if (ISD == ISD::SHL && (VT == MVT::v8i32 || VT == MVT::v16i16) &&
416 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)
417 ISD = ISD::MUL;
418
419 int Idx = CostTableLookup(AVX1CostTable, ISD, VT);
420 if (Idx != -1)
421 return LT.first * AVX1CostTable[Idx].Cost;
422 }
423
424 // Custom lowering of vectors.
425 static const CostTblEntry<MVT::SimpleValueType> CustomLowered[] = {
426 // A v2i64/v4i64 and multiply is custom lowered as a series of long
427 // multiplies(3), shifts(4) and adds(2).
428 { ISD::MUL, MVT::v2i64, 9 },
429 { ISD::MUL, MVT::v4i64, 9 },
430 };
431 int Idx = CostTableLookup(CustomLowered, ISD, LT.second);
432 if (Idx != -1)
433 return LT.first * CustomLowered[Idx].Cost;
434
435 // Special lowering of v4i32 mul on sse2, sse3: Lower v4i32 mul as 2x shuffle,
436 // 2x pmuludq, 2x shuffle.
437 if (ISD == ISD::MUL && LT.second == MVT::v4i32 && ST->hasSSE2() &&
438 !ST->hasSSE41())
439 return LT.first * 6;
440
441 // Fallback to the default implementation.
442 return TargetTransformInfo::getArithmeticInstrCost(Opcode, Ty, Op1Info,
443 Op2Info);
444 }
445
getShuffleCost(ShuffleKind Kind,Type * Tp,int Index,Type * SubTp) const446 unsigned X86TTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
447 Type *SubTp) const {
448 // We only estimate the cost of reverse and alternate shuffles.
449 if (Kind != SK_Reverse && Kind != SK_Alternate)
450 return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
451
452 if (Kind == SK_Reverse) {
453 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
454 unsigned Cost = 1;
455 if (LT.second.getSizeInBits() > 128)
456 Cost = 3; // Extract + insert + copy.
457
458 // Multiple by the number of parts.
459 return Cost * LT.first;
460 }
461
462 if (Kind == SK_Alternate) {
463 // 64-bit packed float vectors (v2f32) are widened to type v4f32.
464 // 64-bit packed integer vectors (v2i32) are promoted to type v2i64.
465 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
466
467 // The backend knows how to generate a single VEX.256 version of
468 // instruction VPBLENDW if the target supports AVX2.
469 if (ST->hasAVX2() && LT.second == MVT::v16i16)
470 return LT.first;
471
472 static const CostTblEntry<MVT::SimpleValueType> AVXAltShuffleTbl[] = {
473 {ISD::VECTOR_SHUFFLE, MVT::v4i64, 1}, // vblendpd
474 {ISD::VECTOR_SHUFFLE, MVT::v4f64, 1}, // vblendpd
475
476 {ISD::VECTOR_SHUFFLE, MVT::v8i32, 1}, // vblendps
477 {ISD::VECTOR_SHUFFLE, MVT::v8f32, 1}, // vblendps
478
479 // This shuffle is custom lowered into a sequence of:
480 // 2x vextractf128 , 2x vpblendw , 1x vinsertf128
481 {ISD::VECTOR_SHUFFLE, MVT::v16i16, 5},
482
483 // This shuffle is custom lowered into a long sequence of:
484 // 2x vextractf128 , 4x vpshufb , 2x vpor , 1x vinsertf128
485 {ISD::VECTOR_SHUFFLE, MVT::v32i8, 9}
486 };
487
488 if (ST->hasAVX()) {
489 int Idx = CostTableLookup(AVXAltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
490 if (Idx != -1)
491 return LT.first * AVXAltShuffleTbl[Idx].Cost;
492 }
493
494 static const CostTblEntry<MVT::SimpleValueType> SSE41AltShuffleTbl[] = {
495 // These are lowered into movsd.
496 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
497 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
498
499 // packed float vectors with four elements are lowered into BLENDI dag
500 // nodes. A v4i32/v4f32 BLENDI generates a single 'blendps'/'blendpd'.
501 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 1},
502 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 1},
503
504 // This shuffle generates a single pshufw.
505 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 1},
506
507 // There is no instruction that matches a v16i8 alternate shuffle.
508 // The backend will expand it into the sequence 'pshufb + pshufb + or'.
509 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 3}
510 };
511
512 if (ST->hasSSE41()) {
513 int Idx = CostTableLookup(SSE41AltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
514 if (Idx != -1)
515 return LT.first * SSE41AltShuffleTbl[Idx].Cost;
516 }
517
518 static const CostTblEntry<MVT::SimpleValueType> SSSE3AltShuffleTbl[] = {
519 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1}, // movsd
520 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1}, // movsd
521
522 // SSE3 doesn't have 'blendps'. The following shuffles are expanded into
523 // the sequence 'shufps + pshufd'
524 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
525 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
526
527 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 3}, // pshufb + pshufb + or
528 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 3} // pshufb + pshufb + or
529 };
530
531 if (ST->hasSSSE3()) {
532 int Idx = CostTableLookup(SSSE3AltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
533 if (Idx != -1)
534 return LT.first * SSSE3AltShuffleTbl[Idx].Cost;
535 }
536
537 static const CostTblEntry<MVT::SimpleValueType> SSEAltShuffleTbl[] = {
538 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1}, // movsd
539 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1}, // movsd
540
541 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2}, // shufps + pshufd
542 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2}, // shufps + pshufd
543
544 // This is expanded into a long sequence of four extract + four insert.
545 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 8}, // 4 x pextrw + 4 pinsrw.
546
547 // 8 x (pinsrw + pextrw + and + movb + movzb + or)
548 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 48}
549 };
550
551 // Fall-back (SSE3 and SSE2).
552 int Idx = CostTableLookup(SSEAltShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second);
553 if (Idx != -1)
554 return LT.first * SSEAltShuffleTbl[Idx].Cost;
555 return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
556 }
557
558 return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
559 }
560
getCastInstrCost(unsigned Opcode,Type * Dst,Type * Src) const561 unsigned X86TTI::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const {
562 int ISD = TLI->InstructionOpcodeToISD(Opcode);
563 assert(ISD && "Invalid opcode");
564
565 std::pair<unsigned, MVT> LTSrc = TLI->getTypeLegalizationCost(Src);
566 std::pair<unsigned, MVT> LTDest = TLI->getTypeLegalizationCost(Dst);
567
568 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
569 SSE2ConvTbl[] = {
570 // These are somewhat magic numbers justified by looking at the output of
571 // Intel's IACA, running some kernels and making sure when we take
572 // legalization into account the throughput will be overestimated.
573 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
574 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
575 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
576 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
577 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
578 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
579 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
580 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
581 // There are faster sequences for float conversions.
582 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
583 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 8 },
584 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
585 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
586 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
587 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
588 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
589 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
590 };
591
592 if (ST->hasSSE2() && !ST->hasAVX()) {
593 int Idx =
594 ConvertCostTableLookup(SSE2ConvTbl, ISD, LTDest.second, LTSrc.second);
595 if (Idx != -1)
596 return LTSrc.first * SSE2ConvTbl[Idx].Cost;
597 }
598
599 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
600 AVX512ConversionTbl[] = {
601 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 1 },
602 { ISD::FP_EXTEND, MVT::v8f64, MVT::v16f32, 3 },
603 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 1 },
604 { ISD::FP_ROUND, MVT::v16f32, MVT::v8f64, 3 },
605
606 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 1 },
607 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 1 },
608 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 1 },
609 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 1 },
610 { ISD::TRUNCATE, MVT::v16i32, MVT::v8i64, 4 },
611
612 // v16i1 -> v16i32 - load + broadcast
613 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1, 2 },
614 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1, 2 },
615
616 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 1 },
617 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 1 },
618 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
619 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
620 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v16i32, 3 },
621 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v16i32, 3 },
622
623 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 },
624 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i8, 2 },
625 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i16, 2 },
626 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 },
627 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 },
628 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 },
629 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 },
630 };
631
632 if (ST->hasAVX512()) {
633 int Idx = ConvertCostTableLookup(AVX512ConversionTbl, ISD, LTDest.second,
634 LTSrc.second);
635 if (Idx != -1)
636 return AVX512ConversionTbl[Idx].Cost;
637 }
638 EVT SrcTy = TLI->getValueType(Src);
639 EVT DstTy = TLI->getValueType(Dst);
640
641 // The function getSimpleVT only handles simple value types.
642 if (!SrcTy.isSimple() || !DstTy.isSimple())
643 return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
644
645 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
646 AVX2ConversionTbl[] = {
647 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1 },
648 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1 },
649 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 3 },
650 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 3 },
651 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
652 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
653 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
654 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
655 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 3 },
656 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 3 },
657 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 3 },
658 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 3 },
659 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
660 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
661 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
662 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
663
664 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 2 },
665 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2 },
666 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 2 },
667 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2 },
668 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 2 },
669 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 4 },
670
671 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 3 },
672 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 3 },
673
674 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 8 },
675 };
676
677 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
678 AVXConversionTbl[] = {
679 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
680 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
681 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 7 },
682 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 4 },
683 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 7 },
684 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 4 },
685 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 4 },
686 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 4 },
687 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 6 },
688 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 4 },
689 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 6 },
690 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 4 },
691 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 6 },
692 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
693 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 4 },
694 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 4 },
695
696 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 4 },
697 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 4 },
698 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 4 },
699 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 4 },
700 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 },
701 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 4 },
702 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 9 },
703
704 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i1, 8 },
705 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 8 },
706 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 },
707 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 },
708 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 },
709 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 },
710 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 3 },
711 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
712 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i1, 3 },
713 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i8, 3 },
714 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i16, 3 },
715 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 },
716
717 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i1, 6 },
718 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 5 },
719 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 },
720 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 9 },
721 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 7 },
722 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 2 },
723 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 },
724 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 6 },
725 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i1, 7 },
726 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i8, 2 },
727 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i16, 2 },
728 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 6 },
729 // The generic code to compute the scalar overhead is currently broken.
730 // Workaround this limitation by estimating the scalarization overhead
731 // here. We have roughly 10 instructions per scalar element.
732 // Multiply that by the vector width.
733 // FIXME: remove that when PR19268 is fixed.
734 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
735 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 4*10 },
736
737 { ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f32, 7 },
738 { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 1 },
739 // This node is expanded into scalarized operations but BasicTTI is overly
740 // optimistic estimating its cost. It computes 3 per element (one
741 // vector-extract, one scalar conversion and one vector-insert). The
742 // problem is that the inserts form a read-modify-write chain so latency
743 // should be factored in too. Inflating the cost per element by 1.
744 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 8*4 },
745 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 4*4 },
746 };
747
748 if (ST->hasAVX2()) {
749 int Idx = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
750 DstTy.getSimpleVT(), SrcTy.getSimpleVT());
751 if (Idx != -1)
752 return AVX2ConversionTbl[Idx].Cost;
753 }
754
755 if (ST->hasAVX()) {
756 int Idx = ConvertCostTableLookup(AVXConversionTbl, ISD, DstTy.getSimpleVT(),
757 SrcTy.getSimpleVT());
758 if (Idx != -1)
759 return AVXConversionTbl[Idx].Cost;
760 }
761
762 return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
763 }
764
getCmpSelInstrCost(unsigned Opcode,Type * ValTy,Type * CondTy) const765 unsigned X86TTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
766 Type *CondTy) const {
767 // Legalize the type.
768 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
769
770 MVT MTy = LT.second;
771
772 int ISD = TLI->InstructionOpcodeToISD(Opcode);
773 assert(ISD && "Invalid opcode");
774
775 static const CostTblEntry<MVT::SimpleValueType> SSE42CostTbl[] = {
776 { ISD::SETCC, MVT::v2f64, 1 },
777 { ISD::SETCC, MVT::v4f32, 1 },
778 { ISD::SETCC, MVT::v2i64, 1 },
779 { ISD::SETCC, MVT::v4i32, 1 },
780 { ISD::SETCC, MVT::v8i16, 1 },
781 { ISD::SETCC, MVT::v16i8, 1 },
782 };
783
784 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTbl[] = {
785 { ISD::SETCC, MVT::v4f64, 1 },
786 { ISD::SETCC, MVT::v8f32, 1 },
787 // AVX1 does not support 8-wide integer compare.
788 { ISD::SETCC, MVT::v4i64, 4 },
789 { ISD::SETCC, MVT::v8i32, 4 },
790 { ISD::SETCC, MVT::v16i16, 4 },
791 { ISD::SETCC, MVT::v32i8, 4 },
792 };
793
794 static const CostTblEntry<MVT::SimpleValueType> AVX2CostTbl[] = {
795 { ISD::SETCC, MVT::v4i64, 1 },
796 { ISD::SETCC, MVT::v8i32, 1 },
797 { ISD::SETCC, MVT::v16i16, 1 },
798 { ISD::SETCC, MVT::v32i8, 1 },
799 };
800
801 static const CostTblEntry<MVT::SimpleValueType> AVX512CostTbl[] = {
802 { ISD::SETCC, MVT::v8i64, 1 },
803 { ISD::SETCC, MVT::v16i32, 1 },
804 { ISD::SETCC, MVT::v8f64, 1 },
805 { ISD::SETCC, MVT::v16f32, 1 },
806 };
807
808 if (ST->hasAVX512()) {
809 int Idx = CostTableLookup(AVX512CostTbl, ISD, MTy);
810 if (Idx != -1)
811 return LT.first * AVX512CostTbl[Idx].Cost;
812 }
813
814 if (ST->hasAVX2()) {
815 int Idx = CostTableLookup(AVX2CostTbl, ISD, MTy);
816 if (Idx != -1)
817 return LT.first * AVX2CostTbl[Idx].Cost;
818 }
819
820 if (ST->hasAVX()) {
821 int Idx = CostTableLookup(AVX1CostTbl, ISD, MTy);
822 if (Idx != -1)
823 return LT.first * AVX1CostTbl[Idx].Cost;
824 }
825
826 if (ST->hasSSE42()) {
827 int Idx = CostTableLookup(SSE42CostTbl, ISD, MTy);
828 if (Idx != -1)
829 return LT.first * SSE42CostTbl[Idx].Cost;
830 }
831
832 return TargetTransformInfo::getCmpSelInstrCost(Opcode, ValTy, CondTy);
833 }
834
getVectorInstrCost(unsigned Opcode,Type * Val,unsigned Index) const835 unsigned X86TTI::getVectorInstrCost(unsigned Opcode, Type *Val,
836 unsigned Index) const {
837 assert(Val->isVectorTy() && "This must be a vector type");
838
839 if (Index != -1U) {
840 // Legalize the type.
841 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Val);
842
843 // This type is legalized to a scalar type.
844 if (!LT.second.isVector())
845 return 0;
846
847 // The type may be split. Normalize the index to the new type.
848 unsigned Width = LT.second.getVectorNumElements();
849 Index = Index % Width;
850
851 // Floating point scalars are already located in index #0.
852 if (Val->getScalarType()->isFloatingPointTy() && Index == 0)
853 return 0;
854 }
855
856 return TargetTransformInfo::getVectorInstrCost(Opcode, Val, Index);
857 }
858
getScalarizationOverhead(Type * Ty,bool Insert,bool Extract) const859 unsigned X86TTI::getScalarizationOverhead(Type *Ty, bool Insert,
860 bool Extract) const {
861 assert (Ty->isVectorTy() && "Can only scalarize vectors");
862 unsigned Cost = 0;
863
864 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
865 if (Insert)
866 Cost += TopTTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
867 if (Extract)
868 Cost += TopTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
869 }
870
871 return Cost;
872 }
873
getMemoryOpCost(unsigned Opcode,Type * Src,unsigned Alignment,unsigned AddressSpace) const874 unsigned X86TTI::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
875 unsigned AddressSpace) const {
876 // Handle non-power-of-two vectors such as <3 x float>
877 if (VectorType *VTy = dyn_cast<VectorType>(Src)) {
878 unsigned NumElem = VTy->getVectorNumElements();
879
880 // Handle a few common cases:
881 // <3 x float>
882 if (NumElem == 3 && VTy->getScalarSizeInBits() == 32)
883 // Cost = 64 bit store + extract + 32 bit store.
884 return 3;
885
886 // <3 x double>
887 if (NumElem == 3 && VTy->getScalarSizeInBits() == 64)
888 // Cost = 128 bit store + unpack + 64 bit store.
889 return 3;
890
891 // Assume that all other non-power-of-two numbers are scalarized.
892 if (!isPowerOf2_32(NumElem)) {
893 unsigned Cost = TargetTransformInfo::getMemoryOpCost(Opcode,
894 VTy->getScalarType(),
895 Alignment,
896 AddressSpace);
897 unsigned SplitCost = getScalarizationOverhead(Src,
898 Opcode == Instruction::Load,
899 Opcode==Instruction::Store);
900 return NumElem * Cost + SplitCost;
901 }
902 }
903
904 // Legalize the type.
905 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Src);
906 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
907 "Invalid Opcode");
908
909 // Each load/store unit costs 1.
910 unsigned Cost = LT.first * 1;
911
912 // On Sandybridge 256bit load/stores are double pumped
913 // (but not on Haswell).
914 if (LT.second.getSizeInBits() > 128 && !ST->hasAVX2())
915 Cost*=2;
916
917 return Cost;
918 }
919
getAddressComputationCost(Type * Ty,bool IsComplex) const920 unsigned X86TTI::getAddressComputationCost(Type *Ty, bool IsComplex) const {
921 // Address computations in vectorized code with non-consecutive addresses will
922 // likely result in more instructions compared to scalar code where the
923 // computation can more often be merged into the index mode. The resulting
924 // extra micro-ops can significantly decrease throughput.
925 unsigned NumVectorInstToHideOverhead = 10;
926
927 if (Ty->isVectorTy() && IsComplex)
928 return NumVectorInstToHideOverhead;
929
930 return TargetTransformInfo::getAddressComputationCost(Ty, IsComplex);
931 }
932
getReductionCost(unsigned Opcode,Type * ValTy,bool IsPairwise) const933 unsigned X86TTI::getReductionCost(unsigned Opcode, Type *ValTy,
934 bool IsPairwise) const {
935
936 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
937
938 MVT MTy = LT.second;
939
940 int ISD = TLI->InstructionOpcodeToISD(Opcode);
941 assert(ISD && "Invalid opcode");
942
943 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
944 // and make it as the cost.
945
946 static const CostTblEntry<MVT::SimpleValueType> SSE42CostTblPairWise[] = {
947 { ISD::FADD, MVT::v2f64, 2 },
948 { ISD::FADD, MVT::v4f32, 4 },
949 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6".
950 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.5".
951 { ISD::ADD, MVT::v8i16, 5 },
952 };
953
954 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTblPairWise[] = {
955 { ISD::FADD, MVT::v4f32, 4 },
956 { ISD::FADD, MVT::v4f64, 5 },
957 { ISD::FADD, MVT::v8f32, 7 },
958 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5".
959 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.5".
960 { ISD::ADD, MVT::v4i64, 5 }, // The data reported by the IACA tool is "4.8".
961 { ISD::ADD, MVT::v8i16, 5 },
962 { ISD::ADD, MVT::v8i32, 5 },
963 };
964
965 static const CostTblEntry<MVT::SimpleValueType> SSE42CostTblNoPairWise[] = {
966 { ISD::FADD, MVT::v2f64, 2 },
967 { ISD::FADD, MVT::v4f32, 4 },
968 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6".
969 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.3".
970 { ISD::ADD, MVT::v8i16, 4 }, // The data reported by the IACA tool is "4.3".
971 };
972
973 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTblNoPairWise[] = {
974 { ISD::FADD, MVT::v4f32, 3 },
975 { ISD::FADD, MVT::v4f64, 3 },
976 { ISD::FADD, MVT::v8f32, 4 },
977 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5".
978 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "2.8".
979 { ISD::ADD, MVT::v4i64, 3 },
980 { ISD::ADD, MVT::v8i16, 4 },
981 { ISD::ADD, MVT::v8i32, 5 },
982 };
983
984 if (IsPairwise) {
985 if (ST->hasAVX()) {
986 int Idx = CostTableLookup(AVX1CostTblPairWise, ISD, MTy);
987 if (Idx != -1)
988 return LT.first * AVX1CostTblPairWise[Idx].Cost;
989 }
990
991 if (ST->hasSSE42()) {
992 int Idx = CostTableLookup(SSE42CostTblPairWise, ISD, MTy);
993 if (Idx != -1)
994 return LT.first * SSE42CostTblPairWise[Idx].Cost;
995 }
996 } else {
997 if (ST->hasAVX()) {
998 int Idx = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy);
999 if (Idx != -1)
1000 return LT.first * AVX1CostTblNoPairWise[Idx].Cost;
1001 }
1002
1003 if (ST->hasSSE42()) {
1004 int Idx = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy);
1005 if (Idx != -1)
1006 return LT.first * SSE42CostTblNoPairWise[Idx].Cost;
1007 }
1008 }
1009
1010 return TargetTransformInfo::getReductionCost(Opcode, ValTy, IsPairwise);
1011 }
1012
1013 /// \brief Calculate the cost of materializing a 64-bit value. This helper
1014 /// method might only calculate a fraction of a larger immediate. Therefore it
1015 /// is valid to return a cost of ZERO.
getIntImmCost(int64_t Val) const1016 unsigned X86TTI::getIntImmCost(int64_t Val) const {
1017 if (Val == 0)
1018 return TCC_Free;
1019
1020 if (isInt<32>(Val))
1021 return TCC_Basic;
1022
1023 return 2 * TCC_Basic;
1024 }
1025
getIntImmCost(const APInt & Imm,Type * Ty) const1026 unsigned X86TTI::getIntImmCost(const APInt &Imm, Type *Ty) const {
1027 assert(Ty->isIntegerTy());
1028
1029 unsigned BitSize = Ty->getPrimitiveSizeInBits();
1030 if (BitSize == 0)
1031 return ~0U;
1032
1033 // Never hoist constants larger than 128bit, because this might lead to
1034 // incorrect code generation or assertions in codegen.
1035 // Fixme: Create a cost model for types larger than i128 once the codegen
1036 // issues have been fixed.
1037 if (BitSize > 128)
1038 return TCC_Free;
1039
1040 if (Imm == 0)
1041 return TCC_Free;
1042
1043 // Sign-extend all constants to a multiple of 64-bit.
1044 APInt ImmVal = Imm;
1045 if (BitSize & 0x3f)
1046 ImmVal = Imm.sext((BitSize + 63) & ~0x3fU);
1047
1048 // Split the constant into 64-bit chunks and calculate the cost for each
1049 // chunk.
1050 unsigned Cost = 0;
1051 for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
1052 APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
1053 int64_t Val = Tmp.getSExtValue();
1054 Cost += getIntImmCost(Val);
1055 }
1056 // We need at least one instruction to materialze the constant.
1057 return std::max(1U, Cost);
1058 }
1059
getIntImmCost(unsigned Opcode,unsigned Idx,const APInt & Imm,Type * Ty) const1060 unsigned X86TTI::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
1061 Type *Ty) const {
1062 assert(Ty->isIntegerTy());
1063
1064 unsigned BitSize = Ty->getPrimitiveSizeInBits();
1065 // There is no cost model for constants with a bit size of 0. Return TCC_Free
1066 // here, so that constant hoisting will ignore this constant.
1067 if (BitSize == 0)
1068 return TCC_Free;
1069
1070 unsigned ImmIdx = ~0U;
1071 switch (Opcode) {
1072 default: return TCC_Free;
1073 case Instruction::GetElementPtr:
1074 // Always hoist the base address of a GetElementPtr. This prevents the
1075 // creation of new constants for every base constant that gets constant
1076 // folded with the offset.
1077 if (Idx == 0)
1078 return 2 * TCC_Basic;
1079 return TCC_Free;
1080 case Instruction::Store:
1081 ImmIdx = 0;
1082 break;
1083 case Instruction::Add:
1084 case Instruction::Sub:
1085 case Instruction::Mul:
1086 case Instruction::UDiv:
1087 case Instruction::SDiv:
1088 case Instruction::URem:
1089 case Instruction::SRem:
1090 case Instruction::And:
1091 case Instruction::Or:
1092 case Instruction::Xor:
1093 case Instruction::ICmp:
1094 ImmIdx = 1;
1095 break;
1096 // Always return TCC_Free for the shift value of a shift instruction.
1097 case Instruction::Shl:
1098 case Instruction::LShr:
1099 case Instruction::AShr:
1100 if (Idx == 1)
1101 return TCC_Free;
1102 break;
1103 case Instruction::Trunc:
1104 case Instruction::ZExt:
1105 case Instruction::SExt:
1106 case Instruction::IntToPtr:
1107 case Instruction::PtrToInt:
1108 case Instruction::BitCast:
1109 case Instruction::PHI:
1110 case Instruction::Call:
1111 case Instruction::Select:
1112 case Instruction::Ret:
1113 case Instruction::Load:
1114 break;
1115 }
1116
1117 if (Idx == ImmIdx) {
1118 unsigned NumConstants = (BitSize + 63) / 64;
1119 unsigned Cost = X86TTI::getIntImmCost(Imm, Ty);
1120 return (Cost <= NumConstants * TCC_Basic)
1121 ? static_cast<unsigned>(TCC_Free)
1122 : Cost;
1123 }
1124
1125 return X86TTI::getIntImmCost(Imm, Ty);
1126 }
1127
getIntImmCost(Intrinsic::ID IID,unsigned Idx,const APInt & Imm,Type * Ty) const1128 unsigned X86TTI::getIntImmCost(Intrinsic::ID IID, unsigned Idx,
1129 const APInt &Imm, Type *Ty) const {
1130 assert(Ty->isIntegerTy());
1131
1132 unsigned BitSize = Ty->getPrimitiveSizeInBits();
1133 // There is no cost model for constants with a bit size of 0. Return TCC_Free
1134 // here, so that constant hoisting will ignore this constant.
1135 if (BitSize == 0)
1136 return TCC_Free;
1137
1138 switch (IID) {
1139 default: return TCC_Free;
1140 case Intrinsic::sadd_with_overflow:
1141 case Intrinsic::uadd_with_overflow:
1142 case Intrinsic::ssub_with_overflow:
1143 case Intrinsic::usub_with_overflow:
1144 case Intrinsic::smul_with_overflow:
1145 case Intrinsic::umul_with_overflow:
1146 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
1147 return TCC_Free;
1148 break;
1149 case Intrinsic::experimental_stackmap:
1150 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
1151 return TCC_Free;
1152 break;
1153 case Intrinsic::experimental_patchpoint_void:
1154 case Intrinsic::experimental_patchpoint_i64:
1155 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
1156 return TCC_Free;
1157 break;
1158 }
1159 return X86TTI::getIntImmCost(Imm, Ty);
1160 }
1161
isLegalMaskedLoad(Type * DataTy,int Consecutive) const1162 bool X86TTI::isLegalMaskedLoad(Type *DataTy, int Consecutive) const {
1163 int DataWidth = DataTy->getPrimitiveSizeInBits();
1164
1165 // Todo: AVX512 allows gather/scatter, works with strided and random as well
1166 if ((DataWidth < 32) || (Consecutive == 0))
1167 return false;
1168 if (ST->hasAVX512() || ST->hasAVX2())
1169 return true;
1170 return false;
1171 }
1172
isLegalMaskedStore(Type * DataType,int Consecutive) const1173 bool X86TTI::isLegalMaskedStore(Type *DataType, int Consecutive) const {
1174 return isLegalMaskedLoad(DataType, Consecutive);
1175 }
1176
1177