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#define DEBUG_TYPE "x86tti"
18#include "X86.h"
19#include "X86TargetMachine.h"
20#include "llvm/Analysis/TargetTransformInfo.h"
21#include "llvm/Support/Debug.h"
22#include "llvm/Target/TargetLowering.h"
23#include "llvm/Target/CostTable.h"
24using namespace llvm;
25
26// Declare the pass initialization routine locally as target-specific passes
27// don't havve a target-wide initialization entry point, and so we rely on the
28// pass constructor initialization.
29namespace llvm {
30void initializeX86TTIPass(PassRegistry &);
31}
32
33namespace {
34
35class X86TTI : public ImmutablePass, public TargetTransformInfo {
36  const X86Subtarget *ST;
37  const X86TargetLowering *TLI;
38
39  /// Estimate the overhead of scalarizing an instruction. Insert and Extract
40  /// are set if the result needs to be inserted and/or extracted from vectors.
41  unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const;
42
43public:
44  X86TTI() : ImmutablePass(ID), ST(0), TLI(0) {
45    llvm_unreachable("This pass cannot be directly constructed");
46  }
47
48  X86TTI(const X86TargetMachine *TM)
49      : ImmutablePass(ID), ST(TM->getSubtargetImpl()),
50        TLI(TM->getTargetLowering()) {
51    initializeX86TTIPass(*PassRegistry::getPassRegistry());
52  }
53
54  virtual void initializePass() {
55    pushTTIStack(this);
56  }
57
58  virtual void finalizePass() {
59    popTTIStack();
60  }
61
62  virtual void getAnalysisUsage(AnalysisUsage &AU) const {
63    TargetTransformInfo::getAnalysisUsage(AU);
64  }
65
66  /// Pass identification.
67  static char ID;
68
69  /// Provide necessary pointer adjustments for the two base classes.
70  virtual void *getAdjustedAnalysisPointer(const void *ID) {
71    if (ID == &TargetTransformInfo::ID)
72      return (TargetTransformInfo*)this;
73    return this;
74  }
75
76  /// \name Scalar TTI Implementations
77  /// @{
78  virtual PopcntSupportKind getPopcntSupport(unsigned TyWidth) const;
79
80  /// @}
81
82  /// \name Vector TTI Implementations
83  /// @{
84
85  virtual unsigned getNumberOfRegisters(bool Vector) const;
86  virtual unsigned getRegisterBitWidth(bool Vector) const;
87  virtual unsigned getMaximumUnrollFactor() const;
88  virtual unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
89                                          OperandValueKind,
90                                          OperandValueKind) const;
91  virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp,
92                                  int Index, Type *SubTp) const;
93  virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
94                                    Type *Src) const;
95  virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
96                                      Type *CondTy) const;
97  virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
98                                      unsigned Index) const;
99  virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
100                                   unsigned Alignment,
101                                   unsigned AddressSpace) const;
102
103  virtual unsigned getAddressComputationCost(Type *PtrTy, bool IsComplex) const;
104
105  /// @}
106};
107
108} // end anonymous namespace
109
110INITIALIZE_AG_PASS(X86TTI, TargetTransformInfo, "x86tti",
111                   "X86 Target Transform Info", true, true, false)
112char X86TTI::ID = 0;
113
114ImmutablePass *
115llvm::createX86TargetTransformInfoPass(const X86TargetMachine *TM) {
116  return new X86TTI(TM);
117}
118
119
120//===----------------------------------------------------------------------===//
121//
122// X86 cost model.
123//
124//===----------------------------------------------------------------------===//
125
126X86TTI::PopcntSupportKind X86TTI::getPopcntSupport(unsigned TyWidth) const {
127  assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
128  // TODO: Currently the __builtin_popcount() implementation using SSE3
129  //   instructions is inefficient. Once the problem is fixed, we should
130  //   call ST->hasSSE3() instead of ST->hasSSE4().
131  return ST->hasSSE41() ? PSK_FastHardware : PSK_Software;
132}
133
134unsigned X86TTI::getNumberOfRegisters(bool Vector) const {
135  if (Vector && !ST->hasSSE1())
136    return 0;
137
138  if (ST->is64Bit())
139    return 16;
140  return 8;
141}
142
143unsigned X86TTI::getRegisterBitWidth(bool Vector) const {
144  if (Vector) {
145    if (ST->hasAVX()) return 256;
146    if (ST->hasSSE1()) return 128;
147    return 0;
148  }
149
150  if (ST->is64Bit())
151    return 64;
152  return 32;
153
154}
155
156unsigned X86TTI::getMaximumUnrollFactor() const {
157  if (ST->isAtom())
158    return 1;
159
160  // Sandybridge and Haswell have multiple execution ports and pipelined
161  // vector units.
162  if (ST->hasAVX())
163    return 4;
164
165  return 2;
166}
167
168unsigned X86TTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
169                                        OperandValueKind Op1Info,
170                                        OperandValueKind Op2Info) const {
171  // Legalize the type.
172  std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty);
173
174  int ISD = TLI->InstructionOpcodeToISD(Opcode);
175  assert(ISD && "Invalid opcode");
176
177  static const CostTblEntry<MVT> AVX2CostTable[] = {
178    // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
179    // customize them to detect the cases where shift amount is a scalar one.
180    { ISD::SHL,     MVT::v4i32,    1 },
181    { ISD::SRL,     MVT::v4i32,    1 },
182    { ISD::SRA,     MVT::v4i32,    1 },
183    { ISD::SHL,     MVT::v8i32,    1 },
184    { ISD::SRL,     MVT::v8i32,    1 },
185    { ISD::SRA,     MVT::v8i32,    1 },
186    { ISD::SHL,     MVT::v2i64,    1 },
187    { ISD::SRL,     MVT::v2i64,    1 },
188    { ISD::SHL,     MVT::v4i64,    1 },
189    { ISD::SRL,     MVT::v4i64,    1 },
190
191    { ISD::SHL,  MVT::v32i8,  42 }, // cmpeqb sequence.
192    { ISD::SHL,  MVT::v16i16,  16*10 }, // Scalarized.
193
194    { ISD::SRL,  MVT::v32i8,  32*10 }, // Scalarized.
195    { ISD::SRL,  MVT::v16i16,  8*10 }, // Scalarized.
196
197    { ISD::SRA,  MVT::v32i8,  32*10 }, // Scalarized.
198    { ISD::SRA,  MVT::v16i16,  16*10 }, // Scalarized.
199    { ISD::SRA,  MVT::v4i64,  4*10 }, // Scalarized.
200
201    // Vectorizing division is a bad idea. See the SSE2 table for more comments.
202    { ISD::SDIV,  MVT::v32i8,  32*20 },
203    { ISD::SDIV,  MVT::v16i16, 16*20 },
204    { ISD::SDIV,  MVT::v8i32,  8*20 },
205    { ISD::SDIV,  MVT::v4i64,  4*20 },
206    { ISD::UDIV,  MVT::v32i8,  32*20 },
207    { ISD::UDIV,  MVT::v16i16, 16*20 },
208    { ISD::UDIV,  MVT::v8i32,  8*20 },
209    { ISD::UDIV,  MVT::v4i64,  4*20 },
210  };
211
212  // Look for AVX2 lowering tricks.
213  if (ST->hasAVX2()) {
214    int Idx = CostTableLookup<MVT>(AVX2CostTable, array_lengthof(AVX2CostTable),
215                                   ISD, LT.second);
216    if (Idx != -1)
217      return LT.first * AVX2CostTable[Idx].Cost;
218  }
219
220  static const CostTblEntry<MVT> SSE2UniformConstCostTable[] = {
221    // We don't correctly identify costs of casts because they are marked as
222    // custom.
223    // Constant splats are cheaper for the following instructions.
224    { ISD::SHL,  MVT::v16i8,  1 }, // psllw.
225    { ISD::SHL,  MVT::v8i16,  1 }, // psllw.
226    { ISD::SHL,  MVT::v4i32,  1 }, // pslld
227    { ISD::SHL,  MVT::v2i64,  1 }, // psllq.
228
229    { ISD::SRL,  MVT::v16i8,  1 }, // psrlw.
230    { ISD::SRL,  MVT::v8i16,  1 }, // psrlw.
231    { ISD::SRL,  MVT::v4i32,  1 }, // psrld.
232    { ISD::SRL,  MVT::v2i64,  1 }, // psrlq.
233
234    { ISD::SRA,  MVT::v16i8,  4 }, // psrlw, pand, pxor, psubb.
235    { ISD::SRA,  MVT::v8i16,  1 }, // psraw.
236    { ISD::SRA,  MVT::v4i32,  1 }, // psrad.
237  };
238
239  if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
240      ST->hasSSE2()) {
241    int Idx = CostTableLookup<MVT>(SSE2UniformConstCostTable,
242                                   array_lengthof(SSE2UniformConstCostTable),
243                                   ISD, LT.second);
244    if (Idx != -1)
245      return LT.first * SSE2UniformConstCostTable[Idx].Cost;
246  }
247
248
249  static const CostTblEntry<MVT> SSE2CostTable[] = {
250    // We don't correctly identify costs of casts because they are marked as
251    // custom.
252    // For some cases, where the shift amount is a scalar we would be able
253    // to generate better code. Unfortunately, when this is the case the value
254    // (the splat) will get hoisted out of the loop, thereby making it invisible
255    // to ISel. The cost model must return worst case assumptions because it is
256    // used for vectorization and we don't want to make vectorized code worse
257    // than scalar code.
258    { ISD::SHL,  MVT::v16i8,  30 }, // cmpeqb sequence.
259    { ISD::SHL,  MVT::v8i16,  8*10 }, // Scalarized.
260    { ISD::SHL,  MVT::v4i32,  2*5 }, // We optimized this using mul.
261    { ISD::SHL,  MVT::v2i64,  2*10 }, // Scalarized.
262
263    { ISD::SRL,  MVT::v16i8,  16*10 }, // Scalarized.
264    { ISD::SRL,  MVT::v8i16,  8*10 }, // Scalarized.
265    { ISD::SRL,  MVT::v4i32,  4*10 }, // Scalarized.
266    { ISD::SRL,  MVT::v2i64,  2*10 }, // Scalarized.
267
268    { ISD::SRA,  MVT::v16i8,  16*10 }, // Scalarized.
269    { ISD::SRA,  MVT::v8i16,  8*10 }, // Scalarized.
270    { ISD::SRA,  MVT::v4i32,  4*10 }, // Scalarized.
271    { ISD::SRA,  MVT::v2i64,  2*10 }, // Scalarized.
272
273    // It is not a good idea to vectorize division. We have to scalarize it and
274    // in the process we will often end up having to spilling regular
275    // registers. The overhead of division is going to dominate most kernels
276    // anyways so try hard to prevent vectorization of division - it is
277    // generally a bad idea. Assume somewhat arbitrarily that we have to be able
278    // to hide "20 cycles" for each lane.
279    { ISD::SDIV,  MVT::v16i8,  16*20 },
280    { ISD::SDIV,  MVT::v8i16,  8*20 },
281    { ISD::SDIV,  MVT::v4i32,  4*20 },
282    { ISD::SDIV,  MVT::v2i64,  2*20 },
283    { ISD::UDIV,  MVT::v16i8,  16*20 },
284    { ISD::UDIV,  MVT::v8i16,  8*20 },
285    { ISD::UDIV,  MVT::v4i32,  4*20 },
286    { ISD::UDIV,  MVT::v2i64,  2*20 },
287  };
288
289  if (ST->hasSSE2()) {
290    int Idx = CostTableLookup<MVT>(SSE2CostTable, array_lengthof(SSE2CostTable),
291                                   ISD, LT.second);
292    if (Idx != -1)
293      return LT.first * SSE2CostTable[Idx].Cost;
294  }
295
296  static const CostTblEntry<MVT> AVX1CostTable[] = {
297    // We don't have to scalarize unsupported ops. We can issue two half-sized
298    // operations and we only need to extract the upper YMM half.
299    // Two ops + 1 extract + 1 insert = 4.
300    { ISD::MUL,     MVT::v8i32,    4 },
301    { ISD::SUB,     MVT::v8i32,    4 },
302    { ISD::ADD,     MVT::v8i32,    4 },
303    { ISD::SUB,     MVT::v4i64,    4 },
304    { ISD::ADD,     MVT::v4i64,    4 },
305    // A v4i64 multiply is custom lowered as two split v2i64 vectors that then
306    // are lowered as a series of long multiplies(3), shifts(4) and adds(2)
307    // Because we believe v4i64 to be a legal type, we must also include the
308    // split factor of two in the cost table. Therefore, the cost here is 18
309    // instead of 9.
310    { ISD::MUL,     MVT::v4i64,    18 },
311  };
312
313  // Look for AVX1 lowering tricks.
314  if (ST->hasAVX() && !ST->hasAVX2()) {
315    int Idx = CostTableLookup<MVT>(AVX1CostTable, array_lengthof(AVX1CostTable),
316                                   ISD, LT.second);
317    if (Idx != -1)
318      return LT.first * AVX1CostTable[Idx].Cost;
319  }
320
321  // Custom lowering of vectors.
322  static const CostTblEntry<MVT> CustomLowered[] = {
323    // A v2i64/v4i64 and multiply is custom lowered as a series of long
324    // multiplies(3), shifts(4) and adds(2).
325    { ISD::MUL,     MVT::v2i64,    9 },
326    { ISD::MUL,     MVT::v4i64,    9 },
327  };
328  int Idx = CostTableLookup<MVT>(CustomLowered, array_lengthof(CustomLowered),
329                                 ISD, LT.second);
330  if (Idx != -1)
331    return LT.first * CustomLowered[Idx].Cost;
332
333  // Special lowering of v4i32 mul on sse2, sse3: Lower v4i32 mul as 2x shuffle,
334  // 2x pmuludq, 2x shuffle.
335  if (ISD == ISD::MUL && LT.second == MVT::v4i32 && ST->hasSSE2() &&
336      !ST->hasSSE41())
337    return 6;
338
339  // Fallback to the default implementation.
340  return TargetTransformInfo::getArithmeticInstrCost(Opcode, Ty, Op1Info,
341                                                     Op2Info);
342}
343
344unsigned X86TTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
345                                Type *SubTp) const {
346  // We only estimate the cost of reverse shuffles.
347  if (Kind != SK_Reverse)
348    return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
349
350  std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
351  unsigned Cost = 1;
352  if (LT.second.getSizeInBits() > 128)
353    Cost = 3; // Extract + insert + copy.
354
355  // Multiple by the number of parts.
356  return Cost * LT.first;
357}
358
359unsigned X86TTI::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const {
360  int ISD = TLI->InstructionOpcodeToISD(Opcode);
361  assert(ISD && "Invalid opcode");
362
363  std::pair<unsigned, MVT> LTSrc = TLI->getTypeLegalizationCost(Src);
364  std::pair<unsigned, MVT> LTDest = TLI->getTypeLegalizationCost(Dst);
365
366  static const TypeConversionCostTblEntry<MVT> SSE2ConvTbl[] = {
367    // These are somewhat magic numbers justified by looking at the output of
368    // Intel's IACA, running some kernels and making sure when we take
369    // legalization into account the throughput will be overestimated.
370    { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
371    { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
372    { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
373    { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
374    { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
375    { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
376    { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
377    { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
378    // There are faster sequences for float conversions.
379    { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
380    { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
381    { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
382    { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
383    { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
384    { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
385    { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
386    { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
387  };
388
389  if (ST->hasSSE2() && !ST->hasAVX()) {
390    int Idx = ConvertCostTableLookup<MVT>(SSE2ConvTbl,
391                                          array_lengthof(SSE2ConvTbl),
392                                          ISD, LTDest.second, LTSrc.second);
393    if (Idx != -1)
394      return LTSrc.first * SSE2ConvTbl[Idx].Cost;
395  }
396
397  EVT SrcTy = TLI->getValueType(Src);
398  EVT DstTy = TLI->getValueType(Dst);
399
400  // The function getSimpleVT only handles simple value types.
401  if (!SrcTy.isSimple() || !DstTy.isSimple())
402    return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
403
404  static const TypeConversionCostTblEntry<MVT> AVXConversionTbl[] = {
405    { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
406    { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
407    { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
408    { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
409    { ISD::TRUNCATE,    MVT::v4i32, MVT::v4i64, 1 },
410    { ISD::TRUNCATE,    MVT::v8i16, MVT::v8i32, 1 },
411
412    { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i1,  8 },
413    { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i8,  8 },
414    { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
415    { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i32, 1 },
416    { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i1,  3 },
417    { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i8,  3 },
418    { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i16, 3 },
419    { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },
420    { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i1,  3 },
421    { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i8,  3 },
422    { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i16, 3 },
423    { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i32, 1 },
424
425    { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i1,  6 },
426    { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i8,  5 },
427    { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
428    { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i32, 9 },
429    { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i1,  7 },
430    { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i8,  2 },
431    { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
432    { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i32, 6 },
433    { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i1,  7 },
434    { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i8,  2 },
435    { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i16, 2 },
436    { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i32, 6 },
437
438    { ISD::FP_TO_SINT,  MVT::v8i8,  MVT::v8f32, 1 },
439    { ISD::FP_TO_SINT,  MVT::v4i8,  MVT::v4f32, 1 },
440    { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1,  6 },
441    { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1,  9 },
442    { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1,  8 },
443    { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8,  6 },
444    { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 6 },
445    { ISD::TRUNCATE,    MVT::v8i32, MVT::v8i64, 3 },
446  };
447
448  if (ST->hasAVX()) {
449    int Idx = ConvertCostTableLookup<MVT>(AVXConversionTbl,
450                                 array_lengthof(AVXConversionTbl),
451                                 ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT());
452    if (Idx != -1)
453      return AVXConversionTbl[Idx].Cost;
454  }
455
456  return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
457}
458
459unsigned X86TTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
460                                    Type *CondTy) const {
461  // Legalize the type.
462  std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
463
464  MVT MTy = LT.second;
465
466  int ISD = TLI->InstructionOpcodeToISD(Opcode);
467  assert(ISD && "Invalid opcode");
468
469  static const CostTblEntry<MVT> SSE42CostTbl[] = {
470    { ISD::SETCC,   MVT::v2f64,   1 },
471    { ISD::SETCC,   MVT::v4f32,   1 },
472    { ISD::SETCC,   MVT::v2i64,   1 },
473    { ISD::SETCC,   MVT::v4i32,   1 },
474    { ISD::SETCC,   MVT::v8i16,   1 },
475    { ISD::SETCC,   MVT::v16i8,   1 },
476  };
477
478  static const CostTblEntry<MVT> AVX1CostTbl[] = {
479    { ISD::SETCC,   MVT::v4f64,   1 },
480    { ISD::SETCC,   MVT::v8f32,   1 },
481    // AVX1 does not support 8-wide integer compare.
482    { ISD::SETCC,   MVT::v4i64,   4 },
483    { ISD::SETCC,   MVT::v8i32,   4 },
484    { ISD::SETCC,   MVT::v16i16,  4 },
485    { ISD::SETCC,   MVT::v32i8,   4 },
486  };
487
488  static const CostTblEntry<MVT> AVX2CostTbl[] = {
489    { ISD::SETCC,   MVT::v4i64,   1 },
490    { ISD::SETCC,   MVT::v8i32,   1 },
491    { ISD::SETCC,   MVT::v16i16,  1 },
492    { ISD::SETCC,   MVT::v32i8,   1 },
493  };
494
495  if (ST->hasAVX2()) {
496    int Idx = CostTableLookup<MVT>(AVX2CostTbl, array_lengthof(AVX2CostTbl),
497                                   ISD, MTy);
498    if (Idx != -1)
499      return LT.first * AVX2CostTbl[Idx].Cost;
500  }
501
502  if (ST->hasAVX()) {
503    int Idx = CostTableLookup<MVT>(AVX1CostTbl, array_lengthof(AVX1CostTbl),
504                                   ISD, MTy);
505    if (Idx != -1)
506      return LT.first * AVX1CostTbl[Idx].Cost;
507  }
508
509  if (ST->hasSSE42()) {
510    int Idx = CostTableLookup<MVT>(SSE42CostTbl, array_lengthof(SSE42CostTbl),
511                                   ISD, MTy);
512    if (Idx != -1)
513      return LT.first * SSE42CostTbl[Idx].Cost;
514  }
515
516  return TargetTransformInfo::getCmpSelInstrCost(Opcode, ValTy, CondTy);
517}
518
519unsigned X86TTI::getVectorInstrCost(unsigned Opcode, Type *Val,
520                                    unsigned Index) const {
521  assert(Val->isVectorTy() && "This must be a vector type");
522
523  if (Index != -1U) {
524    // Legalize the type.
525    std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Val);
526
527    // This type is legalized to a scalar type.
528    if (!LT.second.isVector())
529      return 0;
530
531    // The type may be split. Normalize the index to the new type.
532    unsigned Width = LT.second.getVectorNumElements();
533    Index = Index % Width;
534
535    // Floating point scalars are already located in index #0.
536    if (Val->getScalarType()->isFloatingPointTy() && Index == 0)
537      return 0;
538  }
539
540  return TargetTransformInfo::getVectorInstrCost(Opcode, Val, Index);
541}
542
543unsigned X86TTI::getScalarizationOverhead(Type *Ty, bool Insert,
544                                            bool Extract) const {
545  assert (Ty->isVectorTy() && "Can only scalarize vectors");
546  unsigned Cost = 0;
547
548  for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
549    if (Insert)
550      Cost += TopTTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
551    if (Extract)
552      Cost += TopTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
553  }
554
555  return Cost;
556}
557
558unsigned X86TTI::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
559                                 unsigned AddressSpace) const {
560  // Handle non power of two vectors such as <3 x float>
561  if (VectorType *VTy = dyn_cast<VectorType>(Src)) {
562    unsigned NumElem = VTy->getVectorNumElements();
563
564    // Handle a few common cases:
565    // <3 x float>
566    if (NumElem == 3 && VTy->getScalarSizeInBits() == 32)
567      // Cost = 64 bit store + extract + 32 bit store.
568      return 3;
569
570    // <3 x double>
571    if (NumElem == 3 && VTy->getScalarSizeInBits() == 64)
572      // Cost = 128 bit store + unpack + 64 bit store.
573      return 3;
574
575    // Assume that all other non power-of-two numbers are scalarized.
576    if (!isPowerOf2_32(NumElem)) {
577      unsigned Cost = TargetTransformInfo::getMemoryOpCost(Opcode,
578                                                           VTy->getScalarType(),
579                                                           Alignment,
580                                                           AddressSpace);
581      unsigned SplitCost = getScalarizationOverhead(Src,
582                                                    Opcode == Instruction::Load,
583                                                    Opcode==Instruction::Store);
584      return NumElem * Cost + SplitCost;
585    }
586  }
587
588  // Legalize the type.
589  std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Src);
590  assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
591         "Invalid Opcode");
592
593  // Each load/store unit costs 1.
594  unsigned Cost = LT.first * 1;
595
596  // On Sandybridge 256bit load/stores are double pumped
597  // (but not on Haswell).
598  if (LT.second.getSizeInBits() > 128 && !ST->hasAVX2())
599    Cost*=2;
600
601  return Cost;
602}
603
604unsigned X86TTI::getAddressComputationCost(Type *Ty, bool IsComplex) const {
605  // Address computations in vectorized code with non-consecutive addresses will
606  // likely result in more instructions compared to scalar code where the
607  // computation can more often be merged into the index mode. The resulting
608  // extra micro-ops can significantly decrease throughput.
609  unsigned NumVectorInstToHideOverhead = 10;
610
611  if (Ty->isVectorTy() && IsComplex)
612    return NumVectorInstToHideOverhead;
613
614  return TargetTransformInfo::getAddressComputationCost(Ty, IsComplex);
615}
616