1 //===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 /// \file
9 /// This file implements a TargetTransformInfo analysis pass specific to the
10 /// X86 target machine. It uses the target's detailed information to provide
11 /// more precise answers to certain TTI queries, while letting the target
12 /// independent and default TTI implementations handle the rest.
13 ///
14 //===----------------------------------------------------------------------===//
15 /// About Cost Model numbers used below it's necessary to say the following:
16 /// the numbers correspond to some "generic" X86 CPU instead of usage of
17 /// concrete CPU model. Usually the numbers correspond to CPU where the feature
18 /// apeared at the first time. For example, if we do Subtarget.hasSSE42() in
19 /// the lookups below the cost is based on Nehalem as that was the first CPU
20 /// to support that feature level and thus has most likely the worst case cost.
21 /// Some examples of other technologies/CPUs:
22 ///   SSE 3   - Pentium4 / Athlon64
23 ///   SSE 4.1 - Penryn
24 ///   SSE 4.2 - Nehalem
25 ///   AVX     - Sandy Bridge
26 ///   AVX2    - Haswell
27 ///   AVX-512 - Xeon Phi / Skylake
28 /// And some examples of instruction target dependent costs (latency)
29 ///                   divss     sqrtss          rsqrtss
30 ///   AMD K7            11-16     19              3
31 ///   Piledriver        9-24      13-15           5
32 ///   Jaguar            14        16              2
33 ///   Pentium II,III    18        30              2
34 ///   Nehalem           7-14      7-18            3
35 ///   Haswell           10-13     11              5
36 /// TODO: Develop and implement  the target dependent cost model and
37 /// specialize cost numbers for different Cost Model Targets such as throughput,
38 /// code size, latency and uop count.
39 //===----------------------------------------------------------------------===//
40 
41 #include "X86TargetTransformInfo.h"
42 #include "llvm/Analysis/TargetTransformInfo.h"
43 #include "llvm/CodeGen/BasicTTIImpl.h"
44 #include "llvm/CodeGen/CostTable.h"
45 #include "llvm/CodeGen/TargetLowering.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/Support/Debug.h"
48 
49 using namespace llvm;
50 
51 #define DEBUG_TYPE "x86tti"
52 
53 //===----------------------------------------------------------------------===//
54 //
55 // X86 cost model.
56 //
57 //===----------------------------------------------------------------------===//
58 
59 TargetTransformInfo::PopcntSupportKind
60 X86TTIImpl::getPopcntSupport(unsigned TyWidth) {
61   assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
62   // TODO: Currently the __builtin_popcount() implementation using SSE3
63   //   instructions is inefficient. Once the problem is fixed, we should
64   //   call ST->hasSSE3() instead of ST->hasPOPCNT().
65   return ST->hasPOPCNT() ? TTI::PSK_FastHardware : TTI::PSK_Software;
66 }
67 
68 llvm::Optional<unsigned> X86TTIImpl::getCacheSize(
69   TargetTransformInfo::CacheLevel Level) const {
70   switch (Level) {
71   case TargetTransformInfo::CacheLevel::L1D:
72     //   - Penryn
73     //   - Nehalem
74     //   - Westmere
75     //   - Sandy Bridge
76     //   - Ivy Bridge
77     //   - Haswell
78     //   - Broadwell
79     //   - Skylake
80     //   - Kabylake
81     return 32 * 1024;  //  32 KByte
82   case TargetTransformInfo::CacheLevel::L2D:
83     //   - Penryn
84     //   - Nehalem
85     //   - Westmere
86     //   - Sandy Bridge
87     //   - Ivy Bridge
88     //   - Haswell
89     //   - Broadwell
90     //   - Skylake
91     //   - Kabylake
92     return 256 * 1024; // 256 KByte
93   }
94 
95   llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
96 }
97 
98 llvm::Optional<unsigned> X86TTIImpl::getCacheAssociativity(
99   TargetTransformInfo::CacheLevel Level) const {
100   //   - Penryn
101   //   - Nehalem
102   //   - Westmere
103   //   - Sandy Bridge
104   //   - Ivy Bridge
105   //   - Haswell
106   //   - Broadwell
107   //   - Skylake
108   //   - Kabylake
109   switch (Level) {
110   case TargetTransformInfo::CacheLevel::L1D:
111     LLVM_FALLTHROUGH;
112   case TargetTransformInfo::CacheLevel::L2D:
113     return 8;
114   }
115 
116   llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
117 }
118 
119 unsigned X86TTIImpl::getNumberOfRegisters(unsigned ClassID) const {
120   bool Vector = (ClassID == 1);
121   if (Vector && !ST->hasSSE1())
122     return 0;
123 
124   if (ST->is64Bit()) {
125     if (Vector && ST->hasAVX512())
126       return 32;
127     return 16;
128   }
129   return 8;
130 }
131 
132 TypeSize
133 X86TTIImpl::getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const {
134   unsigned PreferVectorWidth = ST->getPreferVectorWidth();
135   switch (K) {
136   case TargetTransformInfo::RGK_Scalar:
137     return TypeSize::getFixed(ST->is64Bit() ? 64 : 32);
138   case TargetTransformInfo::RGK_FixedWidthVector:
139     if (ST->hasAVX512() && PreferVectorWidth >= 512)
140       return TypeSize::getFixed(512);
141     if (ST->hasAVX() && PreferVectorWidth >= 256)
142       return TypeSize::getFixed(256);
143     if (ST->hasSSE1() && PreferVectorWidth >= 128)
144       return TypeSize::getFixed(128);
145     return TypeSize::getFixed(0);
146   case TargetTransformInfo::RGK_ScalableVector:
147     return TypeSize::getScalable(0);
148   }
149 
150   llvm_unreachable("Unsupported register kind");
151 }
152 
153 unsigned X86TTIImpl::getLoadStoreVecRegBitWidth(unsigned) const {
154   return getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
155       .getFixedSize();
156 }
157 
158 unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) {
159   // If the loop will not be vectorized, don't interleave the loop.
160   // Let regular unroll to unroll the loop, which saves the overflow
161   // check and memory check cost.
162   if (VF == 1)
163     return 1;
164 
165   if (ST->isAtom())
166     return 1;
167 
168   // Sandybridge and Haswell have multiple execution ports and pipelined
169   // vector units.
170   if (ST->hasAVX())
171     return 4;
172 
173   return 2;
174 }
175 
176 InstructionCost X86TTIImpl::getArithmeticInstrCost(
177     unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
178     TTI::OperandValueKind Op1Info, TTI::OperandValueKind Op2Info,
179     TTI::OperandValueProperties Opd1PropInfo,
180     TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args,
181     const Instruction *CxtI) {
182   // TODO: Handle more cost kinds.
183   if (CostKind != TTI::TCK_RecipThroughput)
184     return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
185                                          Op2Info, Opd1PropInfo,
186                                          Opd2PropInfo, Args, CxtI);
187 
188   // vXi8 multiplications are always promoted to vXi16.
189   if (Opcode == Instruction::Mul && Ty->isVectorTy() &&
190       Ty->getScalarSizeInBits() == 8) {
191     Type *WideVecTy =
192         VectorType::getExtendedElementVectorType(cast<VectorType>(Ty));
193     return getCastInstrCost(Instruction::ZExt, WideVecTy, Ty,
194                             TargetTransformInfo::CastContextHint::None,
195                             CostKind) +
196            getCastInstrCost(Instruction::Trunc, Ty, WideVecTy,
197                             TargetTransformInfo::CastContextHint::None,
198                             CostKind) +
199            getArithmeticInstrCost(Opcode, WideVecTy, CostKind, Op1Info, Op2Info,
200                                   Opd1PropInfo, Opd2PropInfo);
201   }
202 
203   // Legalize the type.
204   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
205 
206   int ISD = TLI->InstructionOpcodeToISD(Opcode);
207   assert(ISD && "Invalid opcode");
208 
209   if (ISD == ISD::MUL && Args.size() == 2 && LT.second.isVector() &&
210       LT.second.getScalarType() == MVT::i32) {
211     // Check if the operands can be represented as a smaller datatype.
212     bool Op1Signed = false, Op2Signed = false;
213     unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed);
214     unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed);
215     unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize);
216 
217     // If both are representable as i15 and at least one is zero-extended,
218     // then we can treat this as PMADDWD which has the same costs
219     // as a vXi16 multiply..
220     if (OpMinSize <= 15 && (!Op1Signed || !Op2Signed) && !ST->isPMADDWDSlow())
221       LT.second =
222           MVT::getVectorVT(MVT::i16, 2 * LT.second.getVectorNumElements());
223   }
224 
225   if ((ISD == ISD::SDIV || ISD == ISD::SREM || ISD == ISD::UDIV ||
226        ISD == ISD::UREM) &&
227       (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
228        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
229       Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
230     if (ISD == ISD::SDIV || ISD == ISD::SREM) {
231       // On X86, vector signed division by constants power-of-two are
232       // normally expanded to the sequence SRA + SRL + ADD + SRA.
233       // The OperandValue properties may not be the same as that of the previous
234       // operation; conservatively assume OP_None.
235       InstructionCost Cost =
236           2 * getArithmeticInstrCost(Instruction::AShr, Ty, CostKind, Op1Info,
237                                      Op2Info, TargetTransformInfo::OP_None,
238                                      TargetTransformInfo::OP_None);
239       Cost += getArithmeticInstrCost(Instruction::LShr, Ty, CostKind, Op1Info,
240                                      Op2Info, TargetTransformInfo::OP_None,
241                                      TargetTransformInfo::OP_None);
242       Cost += getArithmeticInstrCost(Instruction::Add, Ty, CostKind, Op1Info,
243                                      Op2Info, TargetTransformInfo::OP_None,
244                                      TargetTransformInfo::OP_None);
245 
246       if (ISD == ISD::SREM) {
247         // For SREM: (X % C) is the equivalent of (X - (X/C)*C)
248         Cost += getArithmeticInstrCost(Instruction::Mul, Ty, CostKind, Op1Info,
249                                        Op2Info);
250         Cost += getArithmeticInstrCost(Instruction::Sub, Ty, CostKind, Op1Info,
251                                        Op2Info);
252       }
253 
254       return Cost;
255     }
256 
257     // Vector unsigned division/remainder will be simplified to shifts/masks.
258     if (ISD == ISD::UDIV)
259       return getArithmeticInstrCost(Instruction::LShr, Ty, CostKind, Op1Info,
260                                     Op2Info, TargetTransformInfo::OP_None,
261                                     TargetTransformInfo::OP_None);
262     // UREM
263     return getArithmeticInstrCost(Instruction::And, Ty, CostKind, Op1Info,
264                                   Op2Info, TargetTransformInfo::OP_None,
265                                   TargetTransformInfo::OP_None);
266   }
267 
268   static const CostTblEntry GLMCostTable[] = {
269     { ISD::FDIV,  MVT::f32,   18 }, // divss
270     { ISD::FDIV,  MVT::v4f32, 35 }, // divps
271     { ISD::FDIV,  MVT::f64,   33 }, // divsd
272     { ISD::FDIV,  MVT::v2f64, 65 }, // divpd
273   };
274 
275   if (ST->useGLMDivSqrtCosts())
276     if (const auto *Entry = CostTableLookup(GLMCostTable, ISD,
277                                             LT.second))
278       return LT.first * Entry->Cost;
279 
280   static const CostTblEntry SLMCostTable[] = {
281     { ISD::MUL,   MVT::v4i32, 11 }, // pmulld
282     { ISD::MUL,   MVT::v8i16, 2  }, // pmullw
283     { ISD::FMUL,  MVT::f64,   2  }, // mulsd
284     { ISD::FMUL,  MVT::v2f64, 4  }, // mulpd
285     { ISD::FMUL,  MVT::v4f32, 2  }, // mulps
286     { ISD::FDIV,  MVT::f32,   17 }, // divss
287     { ISD::FDIV,  MVT::v4f32, 39 }, // divps
288     { ISD::FDIV,  MVT::f64,   32 }, // divsd
289     { ISD::FDIV,  MVT::v2f64, 69 }, // divpd
290     { ISD::FADD,  MVT::v2f64, 2  }, // addpd
291     { ISD::FSUB,  MVT::v2f64, 2  }, // subpd
292     // v2i64/v4i64 mul is custom lowered as a series of long:
293     // multiplies(3), shifts(3) and adds(2)
294     // slm muldq version throughput is 2 and addq throughput 4
295     // thus: 3X2 (muldq throughput) + 3X1 (shift throughput) +
296     //       3X4 (addq throughput) = 17
297     { ISD::MUL,   MVT::v2i64, 17 },
298     // slm addq\subq throughput is 4
299     { ISD::ADD,   MVT::v2i64, 4  },
300     { ISD::SUB,   MVT::v2i64, 4  },
301   };
302 
303   if (ST->isSLM()) {
304     if (Args.size() == 2 && ISD == ISD::MUL && LT.second == MVT::v4i32) {
305       // Check if the operands can be shrinked into a smaller datatype.
306       // TODO: Merge this into generiic vXi32 MUL patterns above.
307       bool Op1Signed = false;
308       unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed);
309       bool Op2Signed = false;
310       unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed);
311 
312       bool SignedMode = Op1Signed || Op2Signed;
313       unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize);
314 
315       if (OpMinSize <= 7)
316         return LT.first * 3; // pmullw/sext
317       if (!SignedMode && OpMinSize <= 8)
318         return LT.first * 3; // pmullw/zext
319       if (OpMinSize <= 15)
320         return LT.first * 5; // pmullw/pmulhw/pshuf
321       if (!SignedMode && OpMinSize <= 16)
322         return LT.first * 5; // pmullw/pmulhw/pshuf
323     }
324 
325     if (const auto *Entry = CostTableLookup(SLMCostTable, ISD,
326                                             LT.second)) {
327       return LT.first * Entry->Cost;
328     }
329   }
330 
331   static const CostTblEntry AVX512BWUniformConstCostTable[] = {
332     { ISD::SHL,  MVT::v64i8,   2 }, // psllw + pand.
333     { ISD::SRL,  MVT::v64i8,   2 }, // psrlw + pand.
334     { ISD::SRA,  MVT::v64i8,   4 }, // psrlw, pand, pxor, psubb.
335   };
336 
337   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
338       ST->hasBWI()) {
339     if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD,
340                                             LT.second))
341       return LT.first * Entry->Cost;
342   }
343 
344   static const CostTblEntry AVX512UniformConstCostTable[] = {
345     { ISD::SRA,  MVT::v2i64,   1 },
346     { ISD::SRA,  MVT::v4i64,   1 },
347     { ISD::SRA,  MVT::v8i64,   1 },
348 
349     { ISD::SHL,  MVT::v64i8,   4 }, // psllw + pand.
350     { ISD::SRL,  MVT::v64i8,   4 }, // psrlw + pand.
351     { ISD::SRA,  MVT::v64i8,   8 }, // psrlw, pand, pxor, psubb.
352 
353     { ISD::SDIV, MVT::v16i32,  6 }, // pmuludq sequence
354     { ISD::SREM, MVT::v16i32,  8 }, // pmuludq+mul+sub sequence
355     { ISD::UDIV, MVT::v16i32,  5 }, // pmuludq sequence
356     { ISD::UREM, MVT::v16i32,  7 }, // pmuludq+mul+sub sequence
357   };
358 
359   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
360       ST->hasAVX512()) {
361     if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD,
362                                             LT.second))
363       return LT.first * Entry->Cost;
364   }
365 
366   static const CostTblEntry AVX2UniformConstCostTable[] = {
367     { ISD::SHL,  MVT::v32i8,   2 }, // psllw + pand.
368     { ISD::SRL,  MVT::v32i8,   2 }, // psrlw + pand.
369     { ISD::SRA,  MVT::v32i8,   4 }, // psrlw, pand, pxor, psubb.
370 
371     { ISD::SRA,  MVT::v4i64,   4 }, // 2 x psrad + shuffle.
372 
373     { ISD::SDIV, MVT::v8i32,   6 }, // pmuludq sequence
374     { ISD::SREM, MVT::v8i32,   8 }, // pmuludq+mul+sub sequence
375     { ISD::UDIV, MVT::v8i32,   5 }, // pmuludq sequence
376     { ISD::UREM, MVT::v8i32,   7 }, // pmuludq+mul+sub sequence
377   };
378 
379   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
380       ST->hasAVX2()) {
381     if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD,
382                                             LT.second))
383       return LT.first * Entry->Cost;
384   }
385 
386   static const CostTblEntry SSE2UniformConstCostTable[] = {
387     { ISD::SHL,  MVT::v16i8,     2 }, // psllw + pand.
388     { ISD::SRL,  MVT::v16i8,     2 }, // psrlw + pand.
389     { ISD::SRA,  MVT::v16i8,     4 }, // psrlw, pand, pxor, psubb.
390 
391     { ISD::SHL,  MVT::v32i8,   4+2 }, // 2*(psllw + pand) + split.
392     { ISD::SRL,  MVT::v32i8,   4+2 }, // 2*(psrlw + pand) + split.
393     { ISD::SRA,  MVT::v32i8,   8+2 }, // 2*(psrlw, pand, pxor, psubb) + split.
394 
395     { ISD::SDIV, MVT::v8i32,  12+2 }, // 2*pmuludq sequence + split.
396     { ISD::SREM, MVT::v8i32,  16+2 }, // 2*pmuludq+mul+sub sequence + split.
397     { ISD::SDIV, MVT::v4i32,     6 }, // pmuludq sequence
398     { ISD::SREM, MVT::v4i32,     8 }, // pmuludq+mul+sub sequence
399     { ISD::UDIV, MVT::v8i32,  10+2 }, // 2*pmuludq sequence + split.
400     { ISD::UREM, MVT::v8i32,  14+2 }, // 2*pmuludq+mul+sub sequence + split.
401     { ISD::UDIV, MVT::v4i32,     5 }, // pmuludq sequence
402     { ISD::UREM, MVT::v4i32,     7 }, // pmuludq+mul+sub sequence
403   };
404 
405   // XOP has faster vXi8 shifts.
406   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
407       ST->hasSSE2() && !ST->hasXOP()) {
408     if (const auto *Entry =
409             CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second))
410       return LT.first * Entry->Cost;
411   }
412 
413   static const CostTblEntry AVX512BWConstCostTable[] = {
414     { ISD::SDIV, MVT::v64i8,  14 }, // 2*ext+2*pmulhw sequence
415     { ISD::SREM, MVT::v64i8,  16 }, // 2*ext+2*pmulhw+mul+sub sequence
416     { ISD::UDIV, MVT::v64i8,  14 }, // 2*ext+2*pmulhw sequence
417     { ISD::UREM, MVT::v64i8,  16 }, // 2*ext+2*pmulhw+mul+sub sequence
418     { ISD::SDIV, MVT::v32i16,  6 }, // vpmulhw sequence
419     { ISD::SREM, MVT::v32i16,  8 }, // vpmulhw+mul+sub sequence
420     { ISD::UDIV, MVT::v32i16,  6 }, // vpmulhuw sequence
421     { ISD::UREM, MVT::v32i16,  8 }, // vpmulhuw+mul+sub sequence
422   };
423 
424   if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
425        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
426       ST->hasBWI()) {
427     if (const auto *Entry =
428             CostTableLookup(AVX512BWConstCostTable, ISD, LT.second))
429       return LT.first * Entry->Cost;
430   }
431 
432   static const CostTblEntry AVX512ConstCostTable[] = {
433     { ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence
434     { ISD::SREM, MVT::v16i32, 17 }, // vpmuldq+mul+sub sequence
435     { ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence
436     { ISD::UREM, MVT::v16i32, 17 }, // vpmuludq+mul+sub sequence
437     { ISD::SDIV, MVT::v64i8,  28 }, // 4*ext+4*pmulhw sequence
438     { ISD::SREM, MVT::v64i8,  32 }, // 4*ext+4*pmulhw+mul+sub sequence
439     { ISD::UDIV, MVT::v64i8,  28 }, // 4*ext+4*pmulhw sequence
440     { ISD::UREM, MVT::v64i8,  32 }, // 4*ext+4*pmulhw+mul+sub sequence
441     { ISD::SDIV, MVT::v32i16, 12 }, // 2*vpmulhw sequence
442     { ISD::SREM, MVT::v32i16, 16 }, // 2*vpmulhw+mul+sub sequence
443     { ISD::UDIV, MVT::v32i16, 12 }, // 2*vpmulhuw sequence
444     { ISD::UREM, MVT::v32i16, 16 }, // 2*vpmulhuw+mul+sub sequence
445   };
446 
447   if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
448        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
449       ST->hasAVX512()) {
450     if (const auto *Entry =
451             CostTableLookup(AVX512ConstCostTable, ISD, LT.second))
452       return LT.first * Entry->Cost;
453   }
454 
455   static const CostTblEntry AVX2ConstCostTable[] = {
456     { ISD::SDIV, MVT::v32i8,  14 }, // 2*ext+2*pmulhw sequence
457     { ISD::SREM, MVT::v32i8,  16 }, // 2*ext+2*pmulhw+mul+sub sequence
458     { ISD::UDIV, MVT::v32i8,  14 }, // 2*ext+2*pmulhw sequence
459     { ISD::UREM, MVT::v32i8,  16 }, // 2*ext+2*pmulhw+mul+sub sequence
460     { ISD::SDIV, MVT::v16i16,  6 }, // vpmulhw sequence
461     { ISD::SREM, MVT::v16i16,  8 }, // vpmulhw+mul+sub sequence
462     { ISD::UDIV, MVT::v16i16,  6 }, // vpmulhuw sequence
463     { ISD::UREM, MVT::v16i16,  8 }, // vpmulhuw+mul+sub sequence
464     { ISD::SDIV, MVT::v8i32,  15 }, // vpmuldq sequence
465     { ISD::SREM, MVT::v8i32,  19 }, // vpmuldq+mul+sub sequence
466     { ISD::UDIV, MVT::v8i32,  15 }, // vpmuludq sequence
467     { ISD::UREM, MVT::v8i32,  19 }, // vpmuludq+mul+sub sequence
468   };
469 
470   if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
471        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
472       ST->hasAVX2()) {
473     if (const auto *Entry = CostTableLookup(AVX2ConstCostTable, ISD, LT.second))
474       return LT.first * Entry->Cost;
475   }
476 
477   static const CostTblEntry SSE2ConstCostTable[] = {
478     { ISD::SDIV, MVT::v32i8,  28+2 }, // 4*ext+4*pmulhw sequence + split.
479     { ISD::SREM, MVT::v32i8,  32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split.
480     { ISD::SDIV, MVT::v16i8,    14 }, // 2*ext+2*pmulhw sequence
481     { ISD::SREM, MVT::v16i8,    16 }, // 2*ext+2*pmulhw+mul+sub sequence
482     { ISD::UDIV, MVT::v32i8,  28+2 }, // 4*ext+4*pmulhw sequence + split.
483     { ISD::UREM, MVT::v32i8,  32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split.
484     { ISD::UDIV, MVT::v16i8,    14 }, // 2*ext+2*pmulhw sequence
485     { ISD::UREM, MVT::v16i8,    16 }, // 2*ext+2*pmulhw+mul+sub sequence
486     { ISD::SDIV, MVT::v16i16, 12+2 }, // 2*pmulhw sequence + split.
487     { ISD::SREM, MVT::v16i16, 16+2 }, // 2*pmulhw+mul+sub sequence + split.
488     { ISD::SDIV, MVT::v8i16,     6 }, // pmulhw sequence
489     { ISD::SREM, MVT::v8i16,     8 }, // pmulhw+mul+sub sequence
490     { ISD::UDIV, MVT::v16i16, 12+2 }, // 2*pmulhuw sequence + split.
491     { ISD::UREM, MVT::v16i16, 16+2 }, // 2*pmulhuw+mul+sub sequence + split.
492     { ISD::UDIV, MVT::v8i16,     6 }, // pmulhuw sequence
493     { ISD::UREM, MVT::v8i16,     8 }, // pmulhuw+mul+sub sequence
494     { ISD::SDIV, MVT::v8i32,  38+2 }, // 2*pmuludq sequence + split.
495     { ISD::SREM, MVT::v8i32,  48+2 }, // 2*pmuludq+mul+sub sequence + split.
496     { ISD::SDIV, MVT::v4i32,    19 }, // pmuludq sequence
497     { ISD::SREM, MVT::v4i32,    24 }, // pmuludq+mul+sub sequence
498     { ISD::UDIV, MVT::v8i32,  30+2 }, // 2*pmuludq sequence + split.
499     { ISD::UREM, MVT::v8i32,  40+2 }, // 2*pmuludq+mul+sub sequence + split.
500     { ISD::UDIV, MVT::v4i32,    15 }, // pmuludq sequence
501     { ISD::UREM, MVT::v4i32,    20 }, // pmuludq+mul+sub sequence
502   };
503 
504   if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
505        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
506       ST->hasSSE2()) {
507     // pmuldq sequence.
508     if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX())
509       return LT.first * 32;
510     if (ISD == ISD::SREM && LT.second == MVT::v8i32 && ST->hasAVX())
511       return LT.first * 38;
512     if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
513       return LT.first * 15;
514     if (ISD == ISD::SREM && LT.second == MVT::v4i32 && ST->hasSSE41())
515       return LT.first * 20;
516 
517     if (const auto *Entry = CostTableLookup(SSE2ConstCostTable, ISD, LT.second))
518       return LT.first * Entry->Cost;
519   }
520 
521   static const CostTblEntry AVX512BWShiftCostTable[] = {
522     { ISD::SHL,   MVT::v16i8,      4 }, // extend/vpsllvw/pack sequence.
523     { ISD::SRL,   MVT::v16i8,      4 }, // extend/vpsrlvw/pack sequence.
524     { ISD::SRA,   MVT::v16i8,      4 }, // extend/vpsravw/pack sequence.
525     { ISD::SHL,   MVT::v32i8,      4 }, // extend/vpsllvw/pack sequence.
526     { ISD::SRL,   MVT::v32i8,      4 }, // extend/vpsrlvw/pack sequence.
527     { ISD::SRA,   MVT::v32i8,      6 }, // extend/vpsravw/pack sequence.
528     { ISD::SHL,   MVT::v64i8,      6 }, // extend/vpsllvw/pack sequence.
529     { ISD::SRL,   MVT::v64i8,      7 }, // extend/vpsrlvw/pack sequence.
530     { ISD::SRA,   MVT::v64i8,     15 }, // extend/vpsravw/pack sequence.
531 
532     { ISD::SHL,   MVT::v8i16,      1 }, // vpsllvw
533     { ISD::SRL,   MVT::v8i16,      1 }, // vpsrlvw
534     { ISD::SRA,   MVT::v8i16,      1 }, // vpsravw
535     { ISD::SHL,   MVT::v16i16,     1 }, // vpsllvw
536     { ISD::SRL,   MVT::v16i16,     1 }, // vpsrlvw
537     { ISD::SRA,   MVT::v16i16,     1 }, // vpsravw
538     { ISD::SHL,   MVT::v32i16,     1 }, // vpsllvw
539     { ISD::SRL,   MVT::v32i16,     1 }, // vpsrlvw
540     { ISD::SRA,   MVT::v32i16,     1 }, // vpsravw
541   };
542 
543   if (ST->hasBWI())
544     if (const auto *Entry = CostTableLookup(AVX512BWShiftCostTable, ISD, LT.second))
545       return LT.first * Entry->Cost;
546 
547   static const CostTblEntry AVX2UniformCostTable[] = {
548     // Uniform splats are cheaper for the following instructions.
549     { ISD::SHL,  MVT::v16i16, 1 }, // psllw.
550     { ISD::SRL,  MVT::v16i16, 1 }, // psrlw.
551     { ISD::SRA,  MVT::v16i16, 1 }, // psraw.
552     { ISD::SHL,  MVT::v32i16, 2 }, // 2*psllw.
553     { ISD::SRL,  MVT::v32i16, 2 }, // 2*psrlw.
554     { ISD::SRA,  MVT::v32i16, 2 }, // 2*psraw.
555 
556     { ISD::SHL,  MVT::v8i32,  1 }, // pslld
557     { ISD::SRL,  MVT::v8i32,  1 }, // psrld
558     { ISD::SRA,  MVT::v8i32,  1 }, // psrad
559     { ISD::SHL,  MVT::v4i64,  1 }, // psllq
560     { ISD::SRL,  MVT::v4i64,  1 }, // psrlq
561   };
562 
563   if (ST->hasAVX2() &&
564       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
565        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
566     if (const auto *Entry =
567             CostTableLookup(AVX2UniformCostTable, ISD, LT.second))
568       return LT.first * Entry->Cost;
569   }
570 
571   static const CostTblEntry SSE2UniformCostTable[] = {
572     // Uniform splats are cheaper for the following instructions.
573     { ISD::SHL,  MVT::v8i16,  1 }, // psllw.
574     { ISD::SHL,  MVT::v4i32,  1 }, // pslld
575     { ISD::SHL,  MVT::v2i64,  1 }, // psllq.
576 
577     { ISD::SRL,  MVT::v8i16,  1 }, // psrlw.
578     { ISD::SRL,  MVT::v4i32,  1 }, // psrld.
579     { ISD::SRL,  MVT::v2i64,  1 }, // psrlq.
580 
581     { ISD::SRA,  MVT::v8i16,  1 }, // psraw.
582     { ISD::SRA,  MVT::v4i32,  1 }, // psrad.
583   };
584 
585   if (ST->hasSSE2() &&
586       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
587        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
588     if (const auto *Entry =
589             CostTableLookup(SSE2UniformCostTable, ISD, LT.second))
590       return LT.first * Entry->Cost;
591   }
592 
593   static const CostTblEntry AVX512DQCostTable[] = {
594     { ISD::MUL,  MVT::v2i64, 2 }, // pmullq
595     { ISD::MUL,  MVT::v4i64, 2 }, // pmullq
596     { ISD::MUL,  MVT::v8i64, 2 }  // pmullq
597   };
598 
599   // Look for AVX512DQ lowering tricks for custom cases.
600   if (ST->hasDQI())
601     if (const auto *Entry = CostTableLookup(AVX512DQCostTable, ISD, LT.second))
602       return LT.first * Entry->Cost;
603 
604   static const CostTblEntry AVX512BWCostTable[] = {
605     { ISD::SHL,   MVT::v64i8,     11 }, // vpblendvb sequence.
606     { ISD::SRL,   MVT::v64i8,     11 }, // vpblendvb sequence.
607     { ISD::SRA,   MVT::v64i8,     24 }, // vpblendvb sequence.
608   };
609 
610   // Look for AVX512BW lowering tricks for custom cases.
611   if (ST->hasBWI())
612     if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD, LT.second))
613       return LT.first * Entry->Cost;
614 
615   static const CostTblEntry AVX512CostTable[] = {
616     { ISD::SHL,     MVT::v4i32,      1 },
617     { ISD::SRL,     MVT::v4i32,      1 },
618     { ISD::SRA,     MVT::v4i32,      1 },
619     { ISD::SHL,     MVT::v8i32,      1 },
620     { ISD::SRL,     MVT::v8i32,      1 },
621     { ISD::SRA,     MVT::v8i32,      1 },
622     { ISD::SHL,     MVT::v16i32,     1 },
623     { ISD::SRL,     MVT::v16i32,     1 },
624     { ISD::SRA,     MVT::v16i32,     1 },
625 
626     { ISD::SHL,     MVT::v2i64,      1 },
627     { ISD::SRL,     MVT::v2i64,      1 },
628     { ISD::SHL,     MVT::v4i64,      1 },
629     { ISD::SRL,     MVT::v4i64,      1 },
630     { ISD::SHL,     MVT::v8i64,      1 },
631     { ISD::SRL,     MVT::v8i64,      1 },
632 
633     { ISD::SRA,     MVT::v2i64,      1 },
634     { ISD::SRA,     MVT::v4i64,      1 },
635     { ISD::SRA,     MVT::v8i64,      1 },
636 
637     { ISD::MUL,     MVT::v16i32,     1 }, // pmulld (Skylake from agner.org)
638     { ISD::MUL,     MVT::v8i32,      1 }, // pmulld (Skylake from agner.org)
639     { ISD::MUL,     MVT::v4i32,      1 }, // pmulld (Skylake from agner.org)
640     { ISD::MUL,     MVT::v8i64,      6 }, // 3*pmuludq/3*shift/2*add
641 
642     { ISD::FNEG,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
643     { ISD::FADD,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
644     { ISD::FSUB,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
645     { ISD::FMUL,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
646     { ISD::FDIV,    MVT::f64,        4 }, // Skylake from http://www.agner.org/
647     { ISD::FDIV,    MVT::v2f64,      4 }, // Skylake from http://www.agner.org/
648     { ISD::FDIV,    MVT::v4f64,      8 }, // Skylake from http://www.agner.org/
649     { ISD::FDIV,    MVT::v8f64,     16 }, // Skylake from http://www.agner.org/
650 
651     { ISD::FNEG,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
652     { ISD::FADD,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
653     { ISD::FSUB,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
654     { ISD::FMUL,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
655     { ISD::FDIV,    MVT::f32,        3 }, // Skylake from http://www.agner.org/
656     { ISD::FDIV,    MVT::v4f32,      3 }, // Skylake from http://www.agner.org/
657     { ISD::FDIV,    MVT::v8f32,      5 }, // Skylake from http://www.agner.org/
658     { ISD::FDIV,    MVT::v16f32,    10 }, // Skylake from http://www.agner.org/
659   };
660 
661   if (ST->hasAVX512())
662     if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second))
663       return LT.first * Entry->Cost;
664 
665   static const CostTblEntry AVX2ShiftCostTable[] = {
666     // Shifts on vXi64/vXi32 on AVX2 is legal even though we declare to
667     // customize them to detect the cases where shift amount is a scalar one.
668     { ISD::SHL,     MVT::v4i32,    2 }, // vpsllvd (Haswell from agner.org)
669     { ISD::SRL,     MVT::v4i32,    2 }, // vpsrlvd (Haswell from agner.org)
670     { ISD::SRA,     MVT::v4i32,    2 }, // vpsravd (Haswell from agner.org)
671     { ISD::SHL,     MVT::v8i32,    2 }, // vpsllvd (Haswell from agner.org)
672     { ISD::SRL,     MVT::v8i32,    2 }, // vpsrlvd (Haswell from agner.org)
673     { ISD::SRA,     MVT::v8i32,    2 }, // vpsravd (Haswell from agner.org)
674     { ISD::SHL,     MVT::v2i64,    1 }, // vpsllvq (Haswell from agner.org)
675     { ISD::SRL,     MVT::v2i64,    1 }, // vpsrlvq (Haswell from agner.org)
676     { ISD::SHL,     MVT::v4i64,    1 }, // vpsllvq (Haswell from agner.org)
677     { ISD::SRL,     MVT::v4i64,    1 }, // vpsrlvq (Haswell from agner.org)
678   };
679 
680   if (ST->hasAVX512()) {
681     if (ISD == ISD::SHL && LT.second == MVT::v32i16 &&
682         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
683          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
684       // On AVX512, a packed v32i16 shift left by a constant build_vector
685       // is lowered into a vector multiply (vpmullw).
686       return getArithmeticInstrCost(Instruction::Mul, Ty, CostKind,
687                                     Op1Info, Op2Info,
688                                     TargetTransformInfo::OP_None,
689                                     TargetTransformInfo::OP_None);
690   }
691 
692   // Look for AVX2 lowering tricks (XOP is always better at v4i32 shifts).
693   if (ST->hasAVX2() && !(ST->hasXOP() && LT.second == MVT::v4i32)) {
694     if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
695         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
696          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
697       // On AVX2, a packed v16i16 shift left by a constant build_vector
698       // is lowered into a vector multiply (vpmullw).
699       return getArithmeticInstrCost(Instruction::Mul, Ty, CostKind,
700                                     Op1Info, Op2Info,
701                                     TargetTransformInfo::OP_None,
702                                     TargetTransformInfo::OP_None);
703 
704     if (const auto *Entry = CostTableLookup(AVX2ShiftCostTable, ISD, LT.second))
705       return LT.first * Entry->Cost;
706   }
707 
708   static const CostTblEntry XOPShiftCostTable[] = {
709     // 128bit shifts take 1cy, but right shifts require negation beforehand.
710     { ISD::SHL,     MVT::v16i8,    1 },
711     { ISD::SRL,     MVT::v16i8,    2 },
712     { ISD::SRA,     MVT::v16i8,    2 },
713     { ISD::SHL,     MVT::v8i16,    1 },
714     { ISD::SRL,     MVT::v8i16,    2 },
715     { ISD::SRA,     MVT::v8i16,    2 },
716     { ISD::SHL,     MVT::v4i32,    1 },
717     { ISD::SRL,     MVT::v4i32,    2 },
718     { ISD::SRA,     MVT::v4i32,    2 },
719     { ISD::SHL,     MVT::v2i64,    1 },
720     { ISD::SRL,     MVT::v2i64,    2 },
721     { ISD::SRA,     MVT::v2i64,    2 },
722     // 256bit shifts require splitting if AVX2 didn't catch them above.
723     { ISD::SHL,     MVT::v32i8,  2+2 },
724     { ISD::SRL,     MVT::v32i8,  4+2 },
725     { ISD::SRA,     MVT::v32i8,  4+2 },
726     { ISD::SHL,     MVT::v16i16, 2+2 },
727     { ISD::SRL,     MVT::v16i16, 4+2 },
728     { ISD::SRA,     MVT::v16i16, 4+2 },
729     { ISD::SHL,     MVT::v8i32,  2+2 },
730     { ISD::SRL,     MVT::v8i32,  4+2 },
731     { ISD::SRA,     MVT::v8i32,  4+2 },
732     { ISD::SHL,     MVT::v4i64,  2+2 },
733     { ISD::SRL,     MVT::v4i64,  4+2 },
734     { ISD::SRA,     MVT::v4i64,  4+2 },
735   };
736 
737   // Look for XOP lowering tricks.
738   if (ST->hasXOP()) {
739     // If the right shift is constant then we'll fold the negation so
740     // it's as cheap as a left shift.
741     int ShiftISD = ISD;
742     if ((ShiftISD == ISD::SRL || ShiftISD == ISD::SRA) &&
743         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
744          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
745       ShiftISD = ISD::SHL;
746     if (const auto *Entry =
747             CostTableLookup(XOPShiftCostTable, ShiftISD, LT.second))
748       return LT.first * Entry->Cost;
749   }
750 
751   static const CostTblEntry SSE2UniformShiftCostTable[] = {
752     // Uniform splats are cheaper for the following instructions.
753     { ISD::SHL,  MVT::v16i16, 2+2 }, // 2*psllw + split.
754     { ISD::SHL,  MVT::v8i32,  2+2 }, // 2*pslld + split.
755     { ISD::SHL,  MVT::v4i64,  2+2 }, // 2*psllq + split.
756 
757     { ISD::SRL,  MVT::v16i16, 2+2 }, // 2*psrlw + split.
758     { ISD::SRL,  MVT::v8i32,  2+2 }, // 2*psrld + split.
759     { ISD::SRL,  MVT::v4i64,  2+2 }, // 2*psrlq + split.
760 
761     { ISD::SRA,  MVT::v16i16, 2+2 }, // 2*psraw + split.
762     { ISD::SRA,  MVT::v8i32,  2+2 }, // 2*psrad + split.
763     { ISD::SRA,  MVT::v2i64,    4 }, // 2*psrad + shuffle.
764     { ISD::SRA,  MVT::v4i64,  8+2 }, // 2*(2*psrad + shuffle) + split.
765   };
766 
767   if (ST->hasSSE2() &&
768       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
769        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
770 
771     // Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table.
772     if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2())
773       return LT.first * 4; // 2*psrad + shuffle.
774 
775     if (const auto *Entry =
776             CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second))
777       return LT.first * Entry->Cost;
778   }
779 
780   if (ISD == ISD::SHL &&
781       Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
782     MVT VT = LT.second;
783     // Vector shift left by non uniform constant can be lowered
784     // into vector multiply.
785     if (((VT == MVT::v8i16 || VT == MVT::v4i32) && ST->hasSSE2()) ||
786         ((VT == MVT::v16i16 || VT == MVT::v8i32) && ST->hasAVX()))
787       ISD = ISD::MUL;
788   }
789 
790   static const CostTblEntry AVX2CostTable[] = {
791     { ISD::SHL,  MVT::v16i8,      6 }, // vpblendvb sequence.
792     { ISD::SHL,  MVT::v32i8,      6 }, // vpblendvb sequence.
793     { ISD::SHL,  MVT::v64i8,     12 }, // 2*vpblendvb sequence.
794     { ISD::SHL,  MVT::v8i16,      5 }, // extend/vpsrlvd/pack sequence.
795     { ISD::SHL,  MVT::v16i16,     7 }, // extend/vpsrlvd/pack sequence.
796     { ISD::SHL,  MVT::v32i16,    14 }, // 2*extend/vpsrlvd/pack sequence.
797 
798     { ISD::SRL,  MVT::v16i8,      6 }, // vpblendvb sequence.
799     { ISD::SRL,  MVT::v32i8,      6 }, // vpblendvb sequence.
800     { ISD::SRL,  MVT::v64i8,     12 }, // 2*vpblendvb sequence.
801     { ISD::SRL,  MVT::v8i16,      5 }, // extend/vpsrlvd/pack sequence.
802     { ISD::SRL,  MVT::v16i16,     7 }, // extend/vpsrlvd/pack sequence.
803     { ISD::SRL,  MVT::v32i16,    14 }, // 2*extend/vpsrlvd/pack sequence.
804 
805     { ISD::SRA,  MVT::v16i8,     17 }, // vpblendvb sequence.
806     { ISD::SRA,  MVT::v32i8,     17 }, // vpblendvb sequence.
807     { ISD::SRA,  MVT::v64i8,     34 }, // 2*vpblendvb sequence.
808     { ISD::SRA,  MVT::v8i16,      5 }, // extend/vpsravd/pack sequence.
809     { ISD::SRA,  MVT::v16i16,     7 }, // extend/vpsravd/pack sequence.
810     { ISD::SRA,  MVT::v32i16,    14 }, // 2*extend/vpsravd/pack sequence.
811     { ISD::SRA,  MVT::v2i64,      2 }, // srl/xor/sub sequence.
812     { ISD::SRA,  MVT::v4i64,      2 }, // srl/xor/sub sequence.
813 
814     { ISD::SUB,  MVT::v32i8,      1 }, // psubb
815     { ISD::ADD,  MVT::v32i8,      1 }, // paddb
816     { ISD::SUB,  MVT::v16i16,     1 }, // psubw
817     { ISD::ADD,  MVT::v16i16,     1 }, // paddw
818     { ISD::SUB,  MVT::v8i32,      1 }, // psubd
819     { ISD::ADD,  MVT::v8i32,      1 }, // paddd
820     { ISD::SUB,  MVT::v4i64,      1 }, // psubq
821     { ISD::ADD,  MVT::v4i64,      1 }, // paddq
822 
823     { ISD::MUL,  MVT::v16i16,     1 }, // pmullw
824     { ISD::MUL,  MVT::v8i32,      2 }, // pmulld (Haswell from agner.org)
825     { ISD::MUL,  MVT::v4i64,      6 }, // 3*pmuludq/3*shift/2*add
826 
827     { ISD::FNEG, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
828     { ISD::FNEG, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
829     { ISD::FADD, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
830     { ISD::FADD, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
831     { ISD::FSUB, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
832     { ISD::FSUB, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
833     { ISD::FMUL, MVT::f64,        1 }, // Haswell from http://www.agner.org/
834     { ISD::FMUL, MVT::v2f64,      1 }, // Haswell from http://www.agner.org/
835     { ISD::FMUL, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
836     { ISD::FMUL, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
837 
838     { ISD::FDIV, MVT::f32,        7 }, // Haswell from http://www.agner.org/
839     { ISD::FDIV, MVT::v4f32,      7 }, // Haswell from http://www.agner.org/
840     { ISD::FDIV, MVT::v8f32,     14 }, // Haswell from http://www.agner.org/
841     { ISD::FDIV, MVT::f64,       14 }, // Haswell from http://www.agner.org/
842     { ISD::FDIV, MVT::v2f64,     14 }, // Haswell from http://www.agner.org/
843     { ISD::FDIV, MVT::v4f64,     28 }, // Haswell from http://www.agner.org/
844   };
845 
846   // Look for AVX2 lowering tricks for custom cases.
847   if (ST->hasAVX2())
848     if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second))
849       return LT.first * Entry->Cost;
850 
851   static const CostTblEntry AVX1CostTable[] = {
852     // We don't have to scalarize unsupported ops. We can issue two half-sized
853     // operations and we only need to extract the upper YMM half.
854     // Two ops + 1 extract + 1 insert = 4.
855     { ISD::MUL,     MVT::v16i16,     4 },
856     { ISD::MUL,     MVT::v8i32,      5 }, // BTVER2 from http://www.agner.org/
857     { ISD::MUL,     MVT::v4i64,     12 },
858 
859     { ISD::SUB,     MVT::v32i8,      4 },
860     { ISD::ADD,     MVT::v32i8,      4 },
861     { ISD::SUB,     MVT::v16i16,     4 },
862     { ISD::ADD,     MVT::v16i16,     4 },
863     { ISD::SUB,     MVT::v8i32,      4 },
864     { ISD::ADD,     MVT::v8i32,      4 },
865     { ISD::SUB,     MVT::v4i64,      4 },
866     { ISD::ADD,     MVT::v4i64,      4 },
867 
868     { ISD::SHL,     MVT::v32i8,     22 }, // pblendvb sequence + split.
869     { ISD::SHL,     MVT::v8i16,      6 }, // pblendvb sequence.
870     { ISD::SHL,     MVT::v16i16,    13 }, // pblendvb sequence + split.
871     { ISD::SHL,     MVT::v4i32,      3 }, // pslld/paddd/cvttps2dq/pmulld
872     { ISD::SHL,     MVT::v8i32,      9 }, // pslld/paddd/cvttps2dq/pmulld + split
873     { ISD::SHL,     MVT::v2i64,      2 }, // Shift each lane + blend.
874     { ISD::SHL,     MVT::v4i64,      6 }, // Shift each lane + blend + split.
875 
876     { ISD::SRL,     MVT::v32i8,     23 }, // pblendvb sequence + split.
877     { ISD::SRL,     MVT::v16i16,    28 }, // pblendvb sequence + split.
878     { ISD::SRL,     MVT::v4i32,      6 }, // Shift each lane + blend.
879     { ISD::SRL,     MVT::v8i32,     14 }, // Shift each lane + blend + split.
880     { ISD::SRL,     MVT::v2i64,      2 }, // Shift each lane + blend.
881     { ISD::SRL,     MVT::v4i64,      6 }, // Shift each lane + blend + split.
882 
883     { ISD::SRA,     MVT::v32i8,     44 }, // pblendvb sequence + split.
884     { ISD::SRA,     MVT::v16i16,    28 }, // pblendvb sequence + split.
885     { ISD::SRA,     MVT::v4i32,      6 }, // Shift each lane + blend.
886     { ISD::SRA,     MVT::v8i32,     14 }, // Shift each lane + blend + split.
887     { ISD::SRA,     MVT::v2i64,      5 }, // Shift each lane + blend.
888     { ISD::SRA,     MVT::v4i64,     12 }, // Shift each lane + blend + split.
889 
890     { ISD::FNEG,    MVT::v4f64,      2 }, // BTVER2 from http://www.agner.org/
891     { ISD::FNEG,    MVT::v8f32,      2 }, // BTVER2 from http://www.agner.org/
892 
893     { ISD::FMUL,    MVT::f64,        2 }, // BTVER2 from http://www.agner.org/
894     { ISD::FMUL,    MVT::v2f64,      2 }, // BTVER2 from http://www.agner.org/
895     { ISD::FMUL,    MVT::v4f64,      4 }, // BTVER2 from http://www.agner.org/
896 
897     { ISD::FDIV,    MVT::f32,       14 }, // SNB from http://www.agner.org/
898     { ISD::FDIV,    MVT::v4f32,     14 }, // SNB from http://www.agner.org/
899     { ISD::FDIV,    MVT::v8f32,     28 }, // SNB from http://www.agner.org/
900     { ISD::FDIV,    MVT::f64,       22 }, // SNB from http://www.agner.org/
901     { ISD::FDIV,    MVT::v2f64,     22 }, // SNB from http://www.agner.org/
902     { ISD::FDIV,    MVT::v4f64,     44 }, // SNB from http://www.agner.org/
903   };
904 
905   if (ST->hasAVX())
906     if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, LT.second))
907       return LT.first * Entry->Cost;
908 
909   static const CostTblEntry SSE42CostTable[] = {
910     { ISD::FADD, MVT::f64,     1 }, // Nehalem from http://www.agner.org/
911     { ISD::FADD, MVT::f32,     1 }, // Nehalem from http://www.agner.org/
912     { ISD::FADD, MVT::v2f64,   1 }, // Nehalem from http://www.agner.org/
913     { ISD::FADD, MVT::v4f32,   1 }, // Nehalem from http://www.agner.org/
914 
915     { ISD::FSUB, MVT::f64,     1 }, // Nehalem from http://www.agner.org/
916     { ISD::FSUB, MVT::f32 ,    1 }, // Nehalem from http://www.agner.org/
917     { ISD::FSUB, MVT::v2f64,   1 }, // Nehalem from http://www.agner.org/
918     { ISD::FSUB, MVT::v4f32,   1 }, // Nehalem from http://www.agner.org/
919 
920     { ISD::FMUL, MVT::f64,     1 }, // Nehalem from http://www.agner.org/
921     { ISD::FMUL, MVT::f32,     1 }, // Nehalem from http://www.agner.org/
922     { ISD::FMUL, MVT::v2f64,   1 }, // Nehalem from http://www.agner.org/
923     { ISD::FMUL, MVT::v4f32,   1 }, // Nehalem from http://www.agner.org/
924 
925     { ISD::FDIV,  MVT::f32,   14 }, // Nehalem from http://www.agner.org/
926     { ISD::FDIV,  MVT::v4f32, 14 }, // Nehalem from http://www.agner.org/
927     { ISD::FDIV,  MVT::f64,   22 }, // Nehalem from http://www.agner.org/
928     { ISD::FDIV,  MVT::v2f64, 22 }, // Nehalem from http://www.agner.org/
929 
930     { ISD::MUL,   MVT::v2i64,  6 }  // 3*pmuludq/3*shift/2*add
931   };
932 
933   if (ST->hasSSE42())
934     if (const auto *Entry = CostTableLookup(SSE42CostTable, ISD, LT.second))
935       return LT.first * Entry->Cost;
936 
937   static const CostTblEntry SSE41CostTable[] = {
938     { ISD::SHL,  MVT::v16i8,      10 }, // pblendvb sequence.
939     { ISD::SHL,  MVT::v8i16,      11 }, // pblendvb sequence.
940     { ISD::SHL,  MVT::v4i32,       4 }, // pslld/paddd/cvttps2dq/pmulld
941 
942     { ISD::SRL,  MVT::v16i8,      11 }, // pblendvb sequence.
943     { ISD::SRL,  MVT::v8i16,      13 }, // pblendvb sequence.
944     { ISD::SRL,  MVT::v4i32,      16 }, // Shift each lane + blend.
945 
946     { ISD::SRA,  MVT::v16i8,      21 }, // pblendvb sequence.
947     { ISD::SRA,  MVT::v8i16,      13 }, // pblendvb sequence.
948 
949     { ISD::MUL,  MVT::v4i32,       2 }  // pmulld (Nehalem from agner.org)
950   };
951 
952   if (ST->hasSSE41())
953     if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second))
954       return LT.first * Entry->Cost;
955 
956   static const CostTblEntry SSE2CostTable[] = {
957     // We don't correctly identify costs of casts because they are marked as
958     // custom.
959     { ISD::SHL,  MVT::v16i8,      13 }, // cmpgtb sequence.
960     { ISD::SHL,  MVT::v8i16,      25 }, // cmpgtw sequence.
961     { ISD::SHL,  MVT::v4i32,      16 }, // pslld/paddd/cvttps2dq/pmuludq.
962     { ISD::SHL,  MVT::v2i64,       4 }, // splat+shuffle sequence.
963 
964     { ISD::SRL,  MVT::v16i8,      14 }, // cmpgtb sequence.
965     { ISD::SRL,  MVT::v8i16,      16 }, // cmpgtw sequence.
966     { ISD::SRL,  MVT::v4i32,      12 }, // Shift each lane + blend.
967     { ISD::SRL,  MVT::v2i64,       4 }, // splat+shuffle sequence.
968 
969     { ISD::SRA,  MVT::v16i8,      27 }, // unpacked cmpgtb sequence.
970     { ISD::SRA,  MVT::v8i16,      16 }, // cmpgtw sequence.
971     { ISD::SRA,  MVT::v4i32,      12 }, // Shift each lane + blend.
972     { ISD::SRA,  MVT::v2i64,       8 }, // srl/xor/sub splat+shuffle sequence.
973 
974     { ISD::MUL,  MVT::v8i16,       1 }, // pmullw
975     { ISD::MUL,  MVT::v4i32,       6 }, // 3*pmuludq/4*shuffle
976     { ISD::MUL,  MVT::v2i64,       8 }, // 3*pmuludq/3*shift/2*add
977 
978     { ISD::FDIV, MVT::f32,        23 }, // Pentium IV from http://www.agner.org/
979     { ISD::FDIV, MVT::v4f32,      39 }, // Pentium IV from http://www.agner.org/
980     { ISD::FDIV, MVT::f64,        38 }, // Pentium IV from http://www.agner.org/
981     { ISD::FDIV, MVT::v2f64,      69 }, // Pentium IV from http://www.agner.org/
982 
983     { ISD::FNEG, MVT::f32,         1 }, // Pentium IV from http://www.agner.org/
984     { ISD::FNEG, MVT::f64,         1 }, // Pentium IV from http://www.agner.org/
985     { ISD::FNEG, MVT::v4f32,       1 }, // Pentium IV from http://www.agner.org/
986     { ISD::FNEG, MVT::v2f64,       1 }, // Pentium IV from http://www.agner.org/
987 
988     { ISD::FADD, MVT::f32,         2 }, // Pentium IV from http://www.agner.org/
989     { ISD::FADD, MVT::f64,         2 }, // Pentium IV from http://www.agner.org/
990 
991     { ISD::FSUB, MVT::f32,         2 }, // Pentium IV from http://www.agner.org/
992     { ISD::FSUB, MVT::f64,         2 }, // Pentium IV from http://www.agner.org/
993   };
994 
995   if (ST->hasSSE2())
996     if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second))
997       return LT.first * Entry->Cost;
998 
999   static const CostTblEntry SSE1CostTable[] = {
1000     { ISD::FDIV, MVT::f32,   17 }, // Pentium III from http://www.agner.org/
1001     { ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/
1002 
1003     { ISD::FNEG, MVT::f32,    2 }, // Pentium III from http://www.agner.org/
1004     { ISD::FNEG, MVT::v4f32,  2 }, // Pentium III from http://www.agner.org/
1005 
1006     { ISD::FADD, MVT::f32,    1 }, // Pentium III from http://www.agner.org/
1007     { ISD::FADD, MVT::v4f32,  2 }, // Pentium III from http://www.agner.org/
1008 
1009     { ISD::FSUB, MVT::f32,    1 }, // Pentium III from http://www.agner.org/
1010     { ISD::FSUB, MVT::v4f32,  2 }, // Pentium III from http://www.agner.org/
1011   };
1012 
1013   if (ST->hasSSE1())
1014     if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second))
1015       return LT.first * Entry->Cost;
1016 
1017   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
1018     { ISD::ADD,  MVT::i64,    1 }, // Core (Merom) from http://www.agner.org/
1019     { ISD::SUB,  MVT::i64,    1 }, // Core (Merom) from http://www.agner.org/
1020   };
1021 
1022   if (ST->is64Bit())
1023     if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, LT.second))
1024       return LT.first * Entry->Cost;
1025 
1026   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
1027     { ISD::ADD,  MVT::i8,    1 }, // Pentium III from http://www.agner.org/
1028     { ISD::ADD,  MVT::i16,   1 }, // Pentium III from http://www.agner.org/
1029     { ISD::ADD,  MVT::i32,   1 }, // Pentium III from http://www.agner.org/
1030 
1031     { ISD::SUB,  MVT::i8,    1 }, // Pentium III from http://www.agner.org/
1032     { ISD::SUB,  MVT::i16,   1 }, // Pentium III from http://www.agner.org/
1033     { ISD::SUB,  MVT::i32,   1 }, // Pentium III from http://www.agner.org/
1034   };
1035 
1036   if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, LT.second))
1037     return LT.first * Entry->Cost;
1038 
1039   // It is not a good idea to vectorize division. We have to scalarize it and
1040   // in the process we will often end up having to spilling regular
1041   // registers. The overhead of division is going to dominate most kernels
1042   // anyways so try hard to prevent vectorization of division - it is
1043   // generally a bad idea. Assume somewhat arbitrarily that we have to be able
1044   // to hide "20 cycles" for each lane.
1045   if (LT.second.isVector() && (ISD == ISD::SDIV || ISD == ISD::SREM ||
1046                                ISD == ISD::UDIV || ISD == ISD::UREM)) {
1047     InstructionCost ScalarCost = getArithmeticInstrCost(
1048         Opcode, Ty->getScalarType(), CostKind, Op1Info, Op2Info,
1049         TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
1050     return 20 * LT.first * LT.second.getVectorNumElements() * ScalarCost;
1051   }
1052 
1053   // Fallback to the default implementation.
1054   return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info);
1055 }
1056 
1057 InstructionCost X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
1058                                            VectorType *BaseTp,
1059                                            ArrayRef<int> Mask, int Index,
1060                                            VectorType *SubTp) {
1061   // 64-bit packed float vectors (v2f32) are widened to type v4f32.
1062   // 64-bit packed integer vectors (v2i32) are widened to type v4i32.
1063   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, BaseTp);
1064 
1065   Kind = improveShuffleKindFromMask(Kind, Mask);
1066   // Treat Transpose as 2-op shuffles - there's no difference in lowering.
1067   if (Kind == TTI::SK_Transpose)
1068     Kind = TTI::SK_PermuteTwoSrc;
1069 
1070   // For Broadcasts we are splatting the first element from the first input
1071   // register, so only need to reference that input and all the output
1072   // registers are the same.
1073   if (Kind == TTI::SK_Broadcast)
1074     LT.first = 1;
1075 
1076   // Subvector extractions are free if they start at the beginning of a
1077   // vector and cheap if the subvectors are aligned.
1078   if (Kind == TTI::SK_ExtractSubvector && LT.second.isVector()) {
1079     int NumElts = LT.second.getVectorNumElements();
1080     if ((Index % NumElts) == 0)
1081       return 0;
1082     std::pair<InstructionCost, MVT> SubLT =
1083         TLI->getTypeLegalizationCost(DL, SubTp);
1084     if (SubLT.second.isVector()) {
1085       int NumSubElts = SubLT.second.getVectorNumElements();
1086       if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0)
1087         return SubLT.first;
1088       // Handle some cases for widening legalization. For now we only handle
1089       // cases where the original subvector was naturally aligned and evenly
1090       // fit in its legalized subvector type.
1091       // FIXME: Remove some of the alignment restrictions.
1092       // FIXME: We can use permq for 64-bit or larger extracts from 256-bit
1093       // vectors.
1094       int OrigSubElts = cast<FixedVectorType>(SubTp)->getNumElements();
1095       if (NumSubElts > OrigSubElts && (Index % OrigSubElts) == 0 &&
1096           (NumSubElts % OrigSubElts) == 0 &&
1097           LT.second.getVectorElementType() ==
1098               SubLT.second.getVectorElementType() &&
1099           LT.second.getVectorElementType().getSizeInBits() ==
1100               BaseTp->getElementType()->getPrimitiveSizeInBits()) {
1101         assert(NumElts >= NumSubElts && NumElts > OrigSubElts &&
1102                "Unexpected number of elements!");
1103         auto *VecTy = FixedVectorType::get(BaseTp->getElementType(),
1104                                            LT.second.getVectorNumElements());
1105         auto *SubTy = FixedVectorType::get(BaseTp->getElementType(),
1106                                            SubLT.second.getVectorNumElements());
1107         int ExtractIndex = alignDown((Index % NumElts), NumSubElts);
1108         InstructionCost ExtractCost = getShuffleCost(
1109             TTI::SK_ExtractSubvector, VecTy, None, ExtractIndex, SubTy);
1110 
1111         // If the original size is 32-bits or more, we can use pshufd. Otherwise
1112         // if we have SSSE3 we can use pshufb.
1113         if (SubTp->getPrimitiveSizeInBits() >= 32 || ST->hasSSSE3())
1114           return ExtractCost + 1; // pshufd or pshufb
1115 
1116         assert(SubTp->getPrimitiveSizeInBits() == 16 &&
1117                "Unexpected vector size");
1118 
1119         return ExtractCost + 2; // worst case pshufhw + pshufd
1120       }
1121     }
1122   }
1123 
1124   // Subvector insertions are cheap if the subvectors are aligned.
1125   // Note that in general, the insertion starting at the beginning of a vector
1126   // isn't free, because we need to preserve the rest of the wide vector.
1127   if (Kind == TTI::SK_InsertSubvector && LT.second.isVector()) {
1128     int NumElts = LT.second.getVectorNumElements();
1129     std::pair<InstructionCost, MVT> SubLT =
1130         TLI->getTypeLegalizationCost(DL, SubTp);
1131     if (SubLT.second.isVector()) {
1132       int NumSubElts = SubLT.second.getVectorNumElements();
1133       if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0)
1134         return SubLT.first;
1135     }
1136 
1137     // If the insertion isn't aligned, treat it like a 2-op shuffle.
1138     Kind = TTI::SK_PermuteTwoSrc;
1139   }
1140 
1141   // Handle some common (illegal) sub-vector types as they are often very cheap
1142   // to shuffle even on targets without PSHUFB.
1143   EVT VT = TLI->getValueType(DL, BaseTp);
1144   if (VT.isSimple() && VT.isVector() && VT.getSizeInBits() < 128 &&
1145       !ST->hasSSSE3()) {
1146      static const CostTblEntry SSE2SubVectorShuffleTbl[] = {
1147       {TTI::SK_Broadcast,        MVT::v4i16, 1}, // pshuflw
1148       {TTI::SK_Broadcast,        MVT::v2i16, 1}, // pshuflw
1149       {TTI::SK_Broadcast,        MVT::v8i8,  2}, // punpck/pshuflw
1150       {TTI::SK_Broadcast,        MVT::v4i8,  2}, // punpck/pshuflw
1151       {TTI::SK_Broadcast,        MVT::v2i8,  1}, // punpck
1152 
1153       {TTI::SK_Reverse,          MVT::v4i16, 1}, // pshuflw
1154       {TTI::SK_Reverse,          MVT::v2i16, 1}, // pshuflw
1155       {TTI::SK_Reverse,          MVT::v4i8,  3}, // punpck/pshuflw/packus
1156       {TTI::SK_Reverse,          MVT::v2i8,  1}, // punpck
1157 
1158       {TTI::SK_PermuteTwoSrc,    MVT::v4i16, 2}, // punpck/pshuflw
1159       {TTI::SK_PermuteTwoSrc,    MVT::v2i16, 2}, // punpck/pshuflw
1160       {TTI::SK_PermuteTwoSrc,    MVT::v8i8,  7}, // punpck/pshuflw
1161       {TTI::SK_PermuteTwoSrc,    MVT::v4i8,  4}, // punpck/pshuflw
1162       {TTI::SK_PermuteTwoSrc,    MVT::v2i8,  2}, // punpck
1163 
1164       {TTI::SK_PermuteSingleSrc, MVT::v4i16, 1}, // pshuflw
1165       {TTI::SK_PermuteSingleSrc, MVT::v2i16, 1}, // pshuflw
1166       {TTI::SK_PermuteSingleSrc, MVT::v8i8,  5}, // punpck/pshuflw
1167       {TTI::SK_PermuteSingleSrc, MVT::v4i8,  3}, // punpck/pshuflw
1168       {TTI::SK_PermuteSingleSrc, MVT::v2i8,  1}, // punpck
1169     };
1170 
1171     if (ST->hasSSE2())
1172       if (const auto *Entry =
1173               CostTableLookup(SSE2SubVectorShuffleTbl, Kind, VT.getSimpleVT()))
1174         return Entry->Cost;
1175   }
1176 
1177   // We are going to permute multiple sources and the result will be in multiple
1178   // destinations. Providing an accurate cost only for splits where the element
1179   // type remains the same.
1180   if (Kind == TTI::SK_PermuteSingleSrc && LT.first != 1) {
1181     MVT LegalVT = LT.second;
1182     if (LegalVT.isVector() &&
1183         LegalVT.getVectorElementType().getSizeInBits() ==
1184             BaseTp->getElementType()->getPrimitiveSizeInBits() &&
1185         LegalVT.getVectorNumElements() <
1186             cast<FixedVectorType>(BaseTp)->getNumElements()) {
1187 
1188       unsigned VecTySize = DL.getTypeStoreSize(BaseTp);
1189       unsigned LegalVTSize = LegalVT.getStoreSize();
1190       // Number of source vectors after legalization:
1191       unsigned NumOfSrcs = (VecTySize + LegalVTSize - 1) / LegalVTSize;
1192       // Number of destination vectors after legalization:
1193       InstructionCost NumOfDests = LT.first;
1194 
1195       auto *SingleOpTy = FixedVectorType::get(BaseTp->getElementType(),
1196                                               LegalVT.getVectorNumElements());
1197 
1198       InstructionCost NumOfShuffles = (NumOfSrcs - 1) * NumOfDests;
1199       return NumOfShuffles * getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy,
1200                                             None, 0, nullptr);
1201     }
1202 
1203     return BaseT::getShuffleCost(Kind, BaseTp, Mask, Index, SubTp);
1204   }
1205 
1206   // For 2-input shuffles, we must account for splitting the 2 inputs into many.
1207   if (Kind == TTI::SK_PermuteTwoSrc && LT.first != 1) {
1208     // We assume that source and destination have the same vector type.
1209     InstructionCost NumOfDests = LT.first;
1210     InstructionCost NumOfShufflesPerDest = LT.first * 2 - 1;
1211     LT.first = NumOfDests * NumOfShufflesPerDest;
1212   }
1213 
1214   static const CostTblEntry AVX512FP16ShuffleTbl[] = {
1215       {TTI::SK_Broadcast, MVT::v32f16, 1}, // vpbroadcastw
1216       {TTI::SK_Broadcast, MVT::v16f16, 1}, // vpbroadcastw
1217       {TTI::SK_Broadcast, MVT::v8f16, 1},  // vpbroadcastw
1218 
1219       {TTI::SK_Reverse, MVT::v32f16, 2}, // vpermw
1220       {TTI::SK_Reverse, MVT::v16f16, 2}, // vpermw
1221       {TTI::SK_Reverse, MVT::v8f16, 1},  // vpshufb
1222 
1223       {TTI::SK_PermuteSingleSrc, MVT::v32f16, 2}, // vpermw
1224       {TTI::SK_PermuteSingleSrc, MVT::v16f16, 2}, // vpermw
1225       {TTI::SK_PermuteSingleSrc, MVT::v8f16, 1},  // vpshufb
1226 
1227       {TTI::SK_PermuteTwoSrc, MVT::v32f16, 2}, // vpermt2w
1228       {TTI::SK_PermuteTwoSrc, MVT::v16f16, 2}, // vpermt2w
1229       {TTI::SK_PermuteTwoSrc, MVT::v8f16, 2}   // vpermt2w
1230   };
1231 
1232   if (!ST->useSoftFloat() && ST->hasFP16())
1233     if (const auto *Entry =
1234             CostTableLookup(AVX512FP16ShuffleTbl, Kind, LT.second))
1235       return LT.first * Entry->Cost;
1236 
1237   static const CostTblEntry AVX512VBMIShuffleTbl[] = {
1238       {TTI::SK_Reverse, MVT::v64i8, 1}, // vpermb
1239       {TTI::SK_Reverse, MVT::v32i8, 1}, // vpermb
1240 
1241       {TTI::SK_PermuteSingleSrc, MVT::v64i8, 1}, // vpermb
1242       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 1}, // vpermb
1243 
1244       {TTI::SK_PermuteTwoSrc, MVT::v64i8, 2}, // vpermt2b
1245       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 2}, // vpermt2b
1246       {TTI::SK_PermuteTwoSrc, MVT::v16i8, 2}  // vpermt2b
1247   };
1248 
1249   if (ST->hasVBMI())
1250     if (const auto *Entry =
1251             CostTableLookup(AVX512VBMIShuffleTbl, Kind, LT.second))
1252       return LT.first * Entry->Cost;
1253 
1254   static const CostTblEntry AVX512BWShuffleTbl[] = {
1255       {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw
1256       {TTI::SK_Broadcast, MVT::v64i8, 1},  // vpbroadcastb
1257 
1258       {TTI::SK_Reverse, MVT::v32i16, 2}, // vpermw
1259       {TTI::SK_Reverse, MVT::v16i16, 2}, // vpermw
1260       {TTI::SK_Reverse, MVT::v64i8, 2},  // pshufb + vshufi64x2
1261 
1262       {TTI::SK_PermuteSingleSrc, MVT::v32i16, 2}, // vpermw
1263       {TTI::SK_PermuteSingleSrc, MVT::v16i16, 2}, // vpermw
1264       {TTI::SK_PermuteSingleSrc, MVT::v64i8, 8},  // extend to v32i16
1265 
1266       {TTI::SK_PermuteTwoSrc, MVT::v32i16, 2}, // vpermt2w
1267       {TTI::SK_PermuteTwoSrc, MVT::v16i16, 2}, // vpermt2w
1268       {TTI::SK_PermuteTwoSrc, MVT::v8i16, 2},  // vpermt2w
1269       {TTI::SK_PermuteTwoSrc, MVT::v64i8, 19}, // 6 * v32i8 + 1
1270 
1271       {TTI::SK_Select, MVT::v32i16, 1}, // vblendmw
1272       {TTI::SK_Select, MVT::v64i8,  1}, // vblendmb
1273   };
1274 
1275   if (ST->hasBWI())
1276     if (const auto *Entry =
1277             CostTableLookup(AVX512BWShuffleTbl, Kind, LT.second))
1278       return LT.first * Entry->Cost;
1279 
1280   static const CostTblEntry AVX512ShuffleTbl[] = {
1281       {TTI::SK_Broadcast, MVT::v8f64, 1},  // vbroadcastpd
1282       {TTI::SK_Broadcast, MVT::v16f32, 1}, // vbroadcastps
1283       {TTI::SK_Broadcast, MVT::v8i64, 1},  // vpbroadcastq
1284       {TTI::SK_Broadcast, MVT::v16i32, 1}, // vpbroadcastd
1285       {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw
1286       {TTI::SK_Broadcast, MVT::v64i8, 1},  // vpbroadcastb
1287 
1288       {TTI::SK_Reverse, MVT::v8f64, 1},  // vpermpd
1289       {TTI::SK_Reverse, MVT::v16f32, 1}, // vpermps
1290       {TTI::SK_Reverse, MVT::v8i64, 1},  // vpermq
1291       {TTI::SK_Reverse, MVT::v16i32, 1}, // vpermd
1292       {TTI::SK_Reverse, MVT::v32i16, 7}, // per mca
1293       {TTI::SK_Reverse, MVT::v64i8,  7}, // per mca
1294 
1295       {TTI::SK_PermuteSingleSrc, MVT::v8f64, 1},  // vpermpd
1296       {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1},  // vpermpd
1297       {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1},  // vpermpd
1298       {TTI::SK_PermuteSingleSrc, MVT::v16f32, 1}, // vpermps
1299       {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1},  // vpermps
1300       {TTI::SK_PermuteSingleSrc, MVT::v4f32, 1},  // vpermps
1301       {TTI::SK_PermuteSingleSrc, MVT::v8i64, 1},  // vpermq
1302       {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1},  // vpermq
1303       {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1},  // vpermq
1304       {TTI::SK_PermuteSingleSrc, MVT::v16i32, 1}, // vpermd
1305       {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1},  // vpermd
1306       {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1},  // vpermd
1307       {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1},  // pshufb
1308 
1309       {TTI::SK_PermuteTwoSrc, MVT::v8f64, 1},  // vpermt2pd
1310       {TTI::SK_PermuteTwoSrc, MVT::v16f32, 1}, // vpermt2ps
1311       {TTI::SK_PermuteTwoSrc, MVT::v8i64, 1},  // vpermt2q
1312       {TTI::SK_PermuteTwoSrc, MVT::v16i32, 1}, // vpermt2d
1313       {TTI::SK_PermuteTwoSrc, MVT::v4f64, 1},  // vpermt2pd
1314       {TTI::SK_PermuteTwoSrc, MVT::v8f32, 1},  // vpermt2ps
1315       {TTI::SK_PermuteTwoSrc, MVT::v4i64, 1},  // vpermt2q
1316       {TTI::SK_PermuteTwoSrc, MVT::v8i32, 1},  // vpermt2d
1317       {TTI::SK_PermuteTwoSrc, MVT::v2f64, 1},  // vpermt2pd
1318       {TTI::SK_PermuteTwoSrc, MVT::v4f32, 1},  // vpermt2ps
1319       {TTI::SK_PermuteTwoSrc, MVT::v2i64, 1},  // vpermt2q
1320       {TTI::SK_PermuteTwoSrc, MVT::v4i32, 1},  // vpermt2d
1321 
1322       // FIXME: This just applies the type legalization cost rules above
1323       // assuming these completely split.
1324       {TTI::SK_PermuteSingleSrc, MVT::v32i16, 14},
1325       {TTI::SK_PermuteSingleSrc, MVT::v64i8,  14},
1326       {TTI::SK_PermuteTwoSrc,    MVT::v32i16, 42},
1327       {TTI::SK_PermuteTwoSrc,    MVT::v64i8,  42},
1328 
1329       {TTI::SK_Select, MVT::v32i16, 1}, // vpternlogq
1330       {TTI::SK_Select, MVT::v64i8,  1}, // vpternlogq
1331       {TTI::SK_Select, MVT::v8f64,  1}, // vblendmpd
1332       {TTI::SK_Select, MVT::v16f32, 1}, // vblendmps
1333       {TTI::SK_Select, MVT::v8i64,  1}, // vblendmq
1334       {TTI::SK_Select, MVT::v16i32, 1}, // vblendmd
1335   };
1336 
1337   if (ST->hasAVX512())
1338     if (const auto *Entry = CostTableLookup(AVX512ShuffleTbl, Kind, LT.second))
1339       return LT.first * Entry->Cost;
1340 
1341   static const CostTblEntry AVX2ShuffleTbl[] = {
1342       {TTI::SK_Broadcast, MVT::v4f64, 1},  // vbroadcastpd
1343       {TTI::SK_Broadcast, MVT::v8f32, 1},  // vbroadcastps
1344       {TTI::SK_Broadcast, MVT::v4i64, 1},  // vpbroadcastq
1345       {TTI::SK_Broadcast, MVT::v8i32, 1},  // vpbroadcastd
1346       {TTI::SK_Broadcast, MVT::v16i16, 1}, // vpbroadcastw
1347       {TTI::SK_Broadcast, MVT::v32i8, 1},  // vpbroadcastb
1348 
1349       {TTI::SK_Reverse, MVT::v4f64, 1},  // vpermpd
1350       {TTI::SK_Reverse, MVT::v8f32, 1},  // vpermps
1351       {TTI::SK_Reverse, MVT::v4i64, 1},  // vpermq
1352       {TTI::SK_Reverse, MVT::v8i32, 1},  // vpermd
1353       {TTI::SK_Reverse, MVT::v16i16, 2}, // vperm2i128 + pshufb
1354       {TTI::SK_Reverse, MVT::v32i8, 2},  // vperm2i128 + pshufb
1355 
1356       {TTI::SK_Select, MVT::v16i16, 1}, // vpblendvb
1357       {TTI::SK_Select, MVT::v32i8, 1},  // vpblendvb
1358 
1359       {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1},  // vpermpd
1360       {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1},  // vpermps
1361       {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1},  // vpermq
1362       {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1},  // vpermd
1363       {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vperm2i128 + 2*vpshufb
1364                                                   // + vpblendvb
1365       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4},  // vperm2i128 + 2*vpshufb
1366                                                   // + vpblendvb
1367 
1368       {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3},  // 2*vpermpd + vblendpd
1369       {TTI::SK_PermuteTwoSrc, MVT::v8f32, 3},  // 2*vpermps + vblendps
1370       {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3},  // 2*vpermq + vpblendd
1371       {TTI::SK_PermuteTwoSrc, MVT::v8i32, 3},  // 2*vpermd + vpblendd
1372       {TTI::SK_PermuteTwoSrc, MVT::v16i16, 7}, // 2*vperm2i128 + 4*vpshufb
1373                                                // + vpblendvb
1374       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 7},  // 2*vperm2i128 + 4*vpshufb
1375                                                // + vpblendvb
1376   };
1377 
1378   if (ST->hasAVX2())
1379     if (const auto *Entry = CostTableLookup(AVX2ShuffleTbl, Kind, LT.second))
1380       return LT.first * Entry->Cost;
1381 
1382   static const CostTblEntry XOPShuffleTbl[] = {
1383       {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2},  // vperm2f128 + vpermil2pd
1384       {TTI::SK_PermuteSingleSrc, MVT::v8f32, 2},  // vperm2f128 + vpermil2ps
1385       {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2},  // vperm2f128 + vpermil2pd
1386       {TTI::SK_PermuteSingleSrc, MVT::v8i32, 2},  // vperm2f128 + vpermil2ps
1387       {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vextractf128 + 2*vpperm
1388                                                   // + vinsertf128
1389       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4},  // vextractf128 + 2*vpperm
1390                                                   // + vinsertf128
1391 
1392       {TTI::SK_PermuteTwoSrc, MVT::v16i16, 9}, // 2*vextractf128 + 6*vpperm
1393                                                // + vinsertf128
1394       {TTI::SK_PermuteTwoSrc, MVT::v8i16, 1},  // vpperm
1395       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 9},  // 2*vextractf128 + 6*vpperm
1396                                                // + vinsertf128
1397       {TTI::SK_PermuteTwoSrc, MVT::v16i8, 1},  // vpperm
1398   };
1399 
1400   if (ST->hasXOP())
1401     if (const auto *Entry = CostTableLookup(XOPShuffleTbl, Kind, LT.second))
1402       return LT.first * Entry->Cost;
1403 
1404   static const CostTblEntry AVX1ShuffleTbl[] = {
1405       {TTI::SK_Broadcast, MVT::v4f64, 2},  // vperm2f128 + vpermilpd
1406       {TTI::SK_Broadcast, MVT::v8f32, 2},  // vperm2f128 + vpermilps
1407       {TTI::SK_Broadcast, MVT::v4i64, 2},  // vperm2f128 + vpermilpd
1408       {TTI::SK_Broadcast, MVT::v8i32, 2},  // vperm2f128 + vpermilps
1409       {TTI::SK_Broadcast, MVT::v16i16, 3}, // vpshuflw + vpshufd + vinsertf128
1410       {TTI::SK_Broadcast, MVT::v32i8, 2},  // vpshufb + vinsertf128
1411 
1412       {TTI::SK_Reverse, MVT::v4f64, 2},  // vperm2f128 + vpermilpd
1413       {TTI::SK_Reverse, MVT::v8f32, 2},  // vperm2f128 + vpermilps
1414       {TTI::SK_Reverse, MVT::v4i64, 2},  // vperm2f128 + vpermilpd
1415       {TTI::SK_Reverse, MVT::v8i32, 2},  // vperm2f128 + vpermilps
1416       {TTI::SK_Reverse, MVT::v16i16, 4}, // vextractf128 + 2*pshufb
1417                                          // + vinsertf128
1418       {TTI::SK_Reverse, MVT::v32i8, 4},  // vextractf128 + 2*pshufb
1419                                          // + vinsertf128
1420 
1421       {TTI::SK_Select, MVT::v4i64, 1},  // vblendpd
1422       {TTI::SK_Select, MVT::v4f64, 1},  // vblendpd
1423       {TTI::SK_Select, MVT::v8i32, 1},  // vblendps
1424       {TTI::SK_Select, MVT::v8f32, 1},  // vblendps
1425       {TTI::SK_Select, MVT::v16i16, 3}, // vpand + vpandn + vpor
1426       {TTI::SK_Select, MVT::v32i8, 3},  // vpand + vpandn + vpor
1427 
1428       {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2},  // vperm2f128 + vshufpd
1429       {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2},  // vperm2f128 + vshufpd
1430       {TTI::SK_PermuteSingleSrc, MVT::v8f32, 4},  // 2*vperm2f128 + 2*vshufps
1431       {TTI::SK_PermuteSingleSrc, MVT::v8i32, 4},  // 2*vperm2f128 + 2*vshufps
1432       {TTI::SK_PermuteSingleSrc, MVT::v16i16, 8}, // vextractf128 + 4*pshufb
1433                                                   // + 2*por + vinsertf128
1434       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 8},  // vextractf128 + 4*pshufb
1435                                                   // + 2*por + vinsertf128
1436 
1437       {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3},   // 2*vperm2f128 + vshufpd
1438       {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3},   // 2*vperm2f128 + vshufpd
1439       {TTI::SK_PermuteTwoSrc, MVT::v8f32, 4},   // 2*vperm2f128 + 2*vshufps
1440       {TTI::SK_PermuteTwoSrc, MVT::v8i32, 4},   // 2*vperm2f128 + 2*vshufps
1441       {TTI::SK_PermuteTwoSrc, MVT::v16i16, 15}, // 2*vextractf128 + 8*pshufb
1442                                                 // + 4*por + vinsertf128
1443       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 15},  // 2*vextractf128 + 8*pshufb
1444                                                 // + 4*por + vinsertf128
1445   };
1446 
1447   if (ST->hasAVX())
1448     if (const auto *Entry = CostTableLookup(AVX1ShuffleTbl, Kind, LT.second))
1449       return LT.first * Entry->Cost;
1450 
1451   static const CostTblEntry SSE41ShuffleTbl[] = {
1452       {TTI::SK_Select, MVT::v2i64, 1}, // pblendw
1453       {TTI::SK_Select, MVT::v2f64, 1}, // movsd
1454       {TTI::SK_Select, MVT::v4i32, 1}, // pblendw
1455       {TTI::SK_Select, MVT::v4f32, 1}, // blendps
1456       {TTI::SK_Select, MVT::v8i16, 1}, // pblendw
1457       {TTI::SK_Select, MVT::v16i8, 1}  // pblendvb
1458   };
1459 
1460   if (ST->hasSSE41())
1461     if (const auto *Entry = CostTableLookup(SSE41ShuffleTbl, Kind, LT.second))
1462       return LT.first * Entry->Cost;
1463 
1464   static const CostTblEntry SSSE3ShuffleTbl[] = {
1465       {TTI::SK_Broadcast, MVT::v8i16, 1}, // pshufb
1466       {TTI::SK_Broadcast, MVT::v16i8, 1}, // pshufb
1467 
1468       {TTI::SK_Reverse, MVT::v8i16, 1}, // pshufb
1469       {TTI::SK_Reverse, MVT::v16i8, 1}, // pshufb
1470 
1471       {TTI::SK_Select, MVT::v8i16, 3}, // 2*pshufb + por
1472       {TTI::SK_Select, MVT::v16i8, 3}, // 2*pshufb + por
1473 
1474       {TTI::SK_PermuteSingleSrc, MVT::v8i16, 1}, // pshufb
1475       {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb
1476 
1477       {TTI::SK_PermuteTwoSrc, MVT::v8i16, 3}, // 2*pshufb + por
1478       {TTI::SK_PermuteTwoSrc, MVT::v16i8, 3}, // 2*pshufb + por
1479   };
1480 
1481   if (ST->hasSSSE3())
1482     if (const auto *Entry = CostTableLookup(SSSE3ShuffleTbl, Kind, LT.second))
1483       return LT.first * Entry->Cost;
1484 
1485   static const CostTblEntry SSE2ShuffleTbl[] = {
1486       {TTI::SK_Broadcast, MVT::v2f64, 1}, // shufpd
1487       {TTI::SK_Broadcast, MVT::v2i64, 1}, // pshufd
1488       {TTI::SK_Broadcast, MVT::v4i32, 1}, // pshufd
1489       {TTI::SK_Broadcast, MVT::v8i16, 2}, // pshuflw + pshufd
1490       {TTI::SK_Broadcast, MVT::v16i8, 3}, // unpck + pshuflw + pshufd
1491 
1492       {TTI::SK_Reverse, MVT::v2f64, 1}, // shufpd
1493       {TTI::SK_Reverse, MVT::v2i64, 1}, // pshufd
1494       {TTI::SK_Reverse, MVT::v4i32, 1}, // pshufd
1495       {TTI::SK_Reverse, MVT::v8i16, 3}, // pshuflw + pshufhw + pshufd
1496       {TTI::SK_Reverse, MVT::v16i8, 9}, // 2*pshuflw + 2*pshufhw
1497                                         // + 2*pshufd + 2*unpck + packus
1498 
1499       {TTI::SK_Select, MVT::v2i64, 1}, // movsd
1500       {TTI::SK_Select, MVT::v2f64, 1}, // movsd
1501       {TTI::SK_Select, MVT::v4i32, 2}, // 2*shufps
1502       {TTI::SK_Select, MVT::v8i16, 3}, // pand + pandn + por
1503       {TTI::SK_Select, MVT::v16i8, 3}, // pand + pandn + por
1504 
1505       {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // shufpd
1506       {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // pshufd
1507       {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // pshufd
1508       {TTI::SK_PermuteSingleSrc, MVT::v8i16, 5}, // 2*pshuflw + 2*pshufhw
1509                                                   // + pshufd/unpck
1510     { TTI::SK_PermuteSingleSrc, MVT::v16i8, 10 }, // 2*pshuflw + 2*pshufhw
1511                                                   // + 2*pshufd + 2*unpck + 2*packus
1512 
1513     { TTI::SK_PermuteTwoSrc,    MVT::v2f64,  1 }, // shufpd
1514     { TTI::SK_PermuteTwoSrc,    MVT::v2i64,  1 }, // shufpd
1515     { TTI::SK_PermuteTwoSrc,    MVT::v4i32,  2 }, // 2*{unpck,movsd,pshufd}
1516     { TTI::SK_PermuteTwoSrc,    MVT::v8i16,  8 }, // blend+permute
1517     { TTI::SK_PermuteTwoSrc,    MVT::v16i8, 13 }, // blend+permute
1518   };
1519 
1520   if (ST->hasSSE2())
1521     if (const auto *Entry = CostTableLookup(SSE2ShuffleTbl, Kind, LT.second))
1522       return LT.first * Entry->Cost;
1523 
1524   static const CostTblEntry SSE1ShuffleTbl[] = {
1525     { TTI::SK_Broadcast,        MVT::v4f32, 1 }, // shufps
1526     { TTI::SK_Reverse,          MVT::v4f32, 1 }, // shufps
1527     { TTI::SK_Select,           MVT::v4f32, 2 }, // 2*shufps
1528     { TTI::SK_PermuteSingleSrc, MVT::v4f32, 1 }, // shufps
1529     { TTI::SK_PermuteTwoSrc,    MVT::v4f32, 2 }, // 2*shufps
1530   };
1531 
1532   if (ST->hasSSE1())
1533     if (const auto *Entry = CostTableLookup(SSE1ShuffleTbl, Kind, LT.second))
1534       return LT.first * Entry->Cost;
1535 
1536   return BaseT::getShuffleCost(Kind, BaseTp, Mask, Index, SubTp);
1537 }
1538 
1539 InstructionCost X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
1540                                              Type *Src,
1541                                              TTI::CastContextHint CCH,
1542                                              TTI::TargetCostKind CostKind,
1543                                              const Instruction *I) {
1544   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1545   assert(ISD && "Invalid opcode");
1546 
1547   // TODO: Allow non-throughput costs that aren't binary.
1548   auto AdjustCost = [&CostKind](InstructionCost Cost) -> InstructionCost {
1549     if (CostKind != TTI::TCK_RecipThroughput)
1550       return Cost == 0 ? 0 : 1;
1551     return Cost;
1552   };
1553 
1554   // The cost tables include both specific, custom (non-legal) src/dst type
1555   // conversions and generic, legalized types. We test for customs first, before
1556   // falling back to legalization.
1557   // FIXME: Need a better design of the cost table to handle non-simple types of
1558   // potential massive combinations (elem_num x src_type x dst_type).
1559   static const TypeConversionCostTblEntry AVX512BWConversionTbl[] {
1560     { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8, 1 },
1561     { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8, 1 },
1562 
1563     // Mask sign extend has an instruction.
1564     { ISD::SIGN_EXTEND, MVT::v2i8,   MVT::v2i1,  1 },
1565     { ISD::SIGN_EXTEND, MVT::v2i16,  MVT::v2i1,  1 },
1566     { ISD::SIGN_EXTEND, MVT::v4i8,   MVT::v4i1,  1 },
1567     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i1,  1 },
1568     { ISD::SIGN_EXTEND, MVT::v8i8,   MVT::v8i1,  1 },
1569     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i1,  1 },
1570     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v16i1, 1 },
1571     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 },
1572     { ISD::SIGN_EXTEND, MVT::v32i8,  MVT::v32i1, 1 },
1573     { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i1, 1 },
1574     { ISD::SIGN_EXTEND, MVT::v64i8,  MVT::v64i1, 1 },
1575 
1576     // Mask zero extend is a sext + shift.
1577     { ISD::ZERO_EXTEND, MVT::v2i8,   MVT::v2i1,  2 },
1578     { ISD::ZERO_EXTEND, MVT::v2i16,  MVT::v2i1,  2 },
1579     { ISD::ZERO_EXTEND, MVT::v4i8,   MVT::v4i1,  2 },
1580     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i1,  2 },
1581     { ISD::ZERO_EXTEND, MVT::v8i8,   MVT::v8i1,  2 },
1582     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i1,  2 },
1583     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v16i1, 2 },
1584     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 2 },
1585     { ISD::ZERO_EXTEND, MVT::v32i8,  MVT::v32i1, 2 },
1586     { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i1, 2 },
1587     { ISD::ZERO_EXTEND, MVT::v64i8,  MVT::v64i1, 2 },
1588 
1589     { ISD::TRUNCATE,    MVT::v32i8,  MVT::v32i16, 2 },
1590     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 2 }, // widen to zmm
1591     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i8,   2 }, // widen to zmm
1592     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i16,  2 }, // widen to zmm
1593     { ISD::TRUNCATE,    MVT::v2i8,   MVT::v2i16,  2 }, // vpmovwb
1594     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i8,   2 }, // widen to zmm
1595     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i16,  2 }, // widen to zmm
1596     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  2 }, // vpmovwb
1597     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i8,   2 }, // widen to zmm
1598     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i16,  2 }, // widen to zmm
1599     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  2 }, // vpmovwb
1600     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i8,  2 }, // widen to zmm
1601     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i16, 2 }, // widen to zmm
1602     { ISD::TRUNCATE,    MVT::v32i1,  MVT::v32i8,  2 }, // widen to zmm
1603     { ISD::TRUNCATE,    MVT::v32i1,  MVT::v32i16, 2 },
1604     { ISD::TRUNCATE,    MVT::v64i1,  MVT::v64i8,  2 },
1605   };
1606 
1607   static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = {
1608     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
1609     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
1610 
1611     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
1612     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
1613 
1614     { ISD::FP_TO_SINT,  MVT::v8i64,  MVT::v8f32,  1 },
1615     { ISD::FP_TO_SINT,  MVT::v8i64,  MVT::v8f64,  1 },
1616 
1617     { ISD::FP_TO_UINT,  MVT::v8i64,  MVT::v8f32,  1 },
1618     { ISD::FP_TO_UINT,  MVT::v8i64,  MVT::v8f64,  1 },
1619   };
1620 
1621   // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and
1622   // 256-bit wide vectors.
1623 
1624   static const TypeConversionCostTblEntry AVX512FConversionTbl[] = {
1625     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v8f32,  1 },
1626     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v16f32, 3 },
1627     { ISD::FP_ROUND,  MVT::v8f32,   MVT::v8f64,  1 },
1628 
1629     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i8,   3 }, // sext+vpslld+vptestmd
1630     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i8,   3 }, // sext+vpslld+vptestmd
1631     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i8,   3 }, // sext+vpslld+vptestmd
1632     { ISD::TRUNCATE,  MVT::v16i1,   MVT::v16i8,  3 }, // sext+vpslld+vptestmd
1633     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i16,  3 }, // sext+vpsllq+vptestmq
1634     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i16,  3 }, // sext+vpsllq+vptestmq
1635     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i16,  3 }, // sext+vpsllq+vptestmq
1636     { ISD::TRUNCATE,  MVT::v16i1,   MVT::v16i16, 3 }, // sext+vpslld+vptestmd
1637     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i32,  2 }, // zmm vpslld+vptestmd
1638     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i32,  2 }, // zmm vpslld+vptestmd
1639     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i32,  2 }, // zmm vpslld+vptestmd
1640     { ISD::TRUNCATE,  MVT::v16i1,   MVT::v16i32, 2 }, // vpslld+vptestmd
1641     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i64,  2 }, // zmm vpsllq+vptestmq
1642     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i64,  2 }, // zmm vpsllq+vptestmq
1643     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i64,  2 }, // vpsllq+vptestmq
1644     { ISD::TRUNCATE,  MVT::v2i8,    MVT::v2i32,  2 }, // vpmovdb
1645     { ISD::TRUNCATE,  MVT::v4i8,    MVT::v4i32,  2 }, // vpmovdb
1646     { ISD::TRUNCATE,  MVT::v16i8,   MVT::v16i32, 2 }, // vpmovdb
1647     { ISD::TRUNCATE,  MVT::v16i16,  MVT::v16i32, 2 }, // vpmovdb
1648     { ISD::TRUNCATE,  MVT::v2i8,    MVT::v2i64,  2 }, // vpmovqb
1649     { ISD::TRUNCATE,  MVT::v2i16,   MVT::v2i64,  1 }, // vpshufb
1650     { ISD::TRUNCATE,  MVT::v8i8,    MVT::v8i64,  2 }, // vpmovqb
1651     { ISD::TRUNCATE,  MVT::v8i16,   MVT::v8i64,  2 }, // vpmovqw
1652     { ISD::TRUNCATE,  MVT::v8i32,   MVT::v8i64,  1 }, // vpmovqd
1653     { ISD::TRUNCATE,  MVT::v4i32,   MVT::v4i64,  1 }, // zmm vpmovqd
1654     { ISD::TRUNCATE,  MVT::v16i8,   MVT::v16i64, 5 },// 2*vpmovqd+concat+vpmovdb
1655 
1656     { ISD::TRUNCATE,  MVT::v16i8,  MVT::v16i16,  3 }, // extend to v16i32
1657     { ISD::TRUNCATE,  MVT::v32i8,  MVT::v32i16,  8 },
1658 
1659     // Sign extend is zmm vpternlogd+vptruncdb.
1660     // Zero extend is zmm broadcast load+vptruncdw.
1661     { ISD::SIGN_EXTEND, MVT::v2i8,   MVT::v2i1,   3 },
1662     { ISD::ZERO_EXTEND, MVT::v2i8,   MVT::v2i1,   4 },
1663     { ISD::SIGN_EXTEND, MVT::v4i8,   MVT::v4i1,   3 },
1664     { ISD::ZERO_EXTEND, MVT::v4i8,   MVT::v4i1,   4 },
1665     { ISD::SIGN_EXTEND, MVT::v8i8,   MVT::v8i1,   3 },
1666     { ISD::ZERO_EXTEND, MVT::v8i8,   MVT::v8i1,   4 },
1667     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v16i1,  3 },
1668     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v16i1,  4 },
1669 
1670     // Sign extend is zmm vpternlogd+vptruncdw.
1671     // Zero extend is zmm vpternlogd+vptruncdw+vpsrlw.
1672     { ISD::SIGN_EXTEND, MVT::v2i16,  MVT::v2i1,   3 },
1673     { ISD::ZERO_EXTEND, MVT::v2i16,  MVT::v2i1,   4 },
1674     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i1,   3 },
1675     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i1,   4 },
1676     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i1,   3 },
1677     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i1,   4 },
1678     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1,  3 },
1679     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1,  4 },
1680 
1681     { ISD::SIGN_EXTEND, MVT::v2i32,  MVT::v2i1,   1 }, // zmm vpternlogd
1682     { ISD::ZERO_EXTEND, MVT::v2i32,  MVT::v2i1,   2 }, // zmm vpternlogd+psrld
1683     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i1,   1 }, // zmm vpternlogd
1684     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i1,   2 }, // zmm vpternlogd+psrld
1685     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   1 }, // zmm vpternlogd
1686     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   2 }, // zmm vpternlogd+psrld
1687     { ISD::SIGN_EXTEND, MVT::v2i64,  MVT::v2i1,   1 }, // zmm vpternlogq
1688     { ISD::ZERO_EXTEND, MVT::v2i64,  MVT::v2i1,   2 }, // zmm vpternlogq+psrlq
1689     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   1 }, // zmm vpternlogq
1690     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   2 }, // zmm vpternlogq+psrlq
1691 
1692     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1,  1 }, // vpternlogd
1693     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1,  2 }, // vpternlogd+psrld
1694     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i1,   1 }, // vpternlogq
1695     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i1,   2 }, // vpternlogq+psrlq
1696 
1697     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
1698     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
1699     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1700     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1701     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i8,   1 },
1702     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i8,   1 },
1703     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
1704     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
1705     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
1706     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
1707 
1708     { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8,  3 }, // FIXME: May not be right
1709     { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8,  3 }, // FIXME: May not be right
1710 
1711     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
1712     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
1713     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v16i8,  2 },
1714     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i8,  1 },
1715     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
1716     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i16, 1 },
1717     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
1718     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
1719 
1720     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
1721     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
1722     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v16i8,  2 },
1723     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i8,  1 },
1724     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
1725     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i16, 1 },
1726     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
1727     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
1728     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64, 26 },
1729     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  5 },
1730 
1731     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v16f32, 2 },
1732     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v16f64, 7 },
1733     { ISD::FP_TO_SINT,  MVT::v32i8,  MVT::v32f64,15 },
1734     { ISD::FP_TO_SINT,  MVT::v64i8,  MVT::v64f32,11 },
1735     { ISD::FP_TO_SINT,  MVT::v64i8,  MVT::v64f64,31 },
1736     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v8f64,  3 },
1737     { ISD::FP_TO_SINT,  MVT::v16i16, MVT::v16f64, 7 },
1738     { ISD::FP_TO_SINT,  MVT::v32i16, MVT::v32f32, 5 },
1739     { ISD::FP_TO_SINT,  MVT::v32i16, MVT::v32f64,15 },
1740     { ISD::FP_TO_SINT,  MVT::v8i32,  MVT::v8f64,  1 },
1741     { ISD::FP_TO_SINT,  MVT::v16i32, MVT::v16f64, 3 },
1742 
1743     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f64,  1 },
1744     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v8f64,  3 },
1745     { ISD::FP_TO_UINT,  MVT::v8i8,   MVT::v8f64,  3 },
1746     { ISD::FP_TO_UINT,  MVT::v16i32, MVT::v16f32, 1 },
1747     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v16f32, 3 },
1748     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v16f32, 3 },
1749   };
1750 
1751   static const TypeConversionCostTblEntry AVX512BWVLConversionTbl[] {
1752     // Mask sign extend has an instruction.
1753     { ISD::SIGN_EXTEND, MVT::v2i8,   MVT::v2i1,  1 },
1754     { ISD::SIGN_EXTEND, MVT::v2i16,  MVT::v2i1,  1 },
1755     { ISD::SIGN_EXTEND, MVT::v4i8,   MVT::v4i1,  1 },
1756     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i1,  1 },
1757     { ISD::SIGN_EXTEND, MVT::v8i8,   MVT::v8i1,  1 },
1758     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i1,  1 },
1759     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v16i1, 1 },
1760     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 },
1761     { ISD::SIGN_EXTEND, MVT::v32i8,  MVT::v32i1, 1 },
1762 
1763     // Mask zero extend is a sext + shift.
1764     { ISD::ZERO_EXTEND, MVT::v2i8,   MVT::v2i1,  2 },
1765     { ISD::ZERO_EXTEND, MVT::v2i16,  MVT::v2i1,  2 },
1766     { ISD::ZERO_EXTEND, MVT::v4i8,   MVT::v4i1,  2 },
1767     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i1,  2 },
1768     { ISD::ZERO_EXTEND, MVT::v8i8,   MVT::v8i1,  2 },
1769     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i1,  2 },
1770     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v16i1, 2 },
1771     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 2 },
1772     { ISD::ZERO_EXTEND, MVT::v32i8,  MVT::v32i1, 2 },
1773 
1774     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 2 },
1775     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i8,   2 }, // vpsllw+vptestmb
1776     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i16,  2 }, // vpsllw+vptestmw
1777     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i8,   2 }, // vpsllw+vptestmb
1778     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i16,  2 }, // vpsllw+vptestmw
1779     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i8,   2 }, // vpsllw+vptestmb
1780     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i16,  2 }, // vpsllw+vptestmw
1781     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i8,  2 }, // vpsllw+vptestmb
1782     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i16, 2 }, // vpsllw+vptestmw
1783     { ISD::TRUNCATE,    MVT::v32i1,  MVT::v32i8,  2 }, // vpsllw+vptestmb
1784   };
1785 
1786   static const TypeConversionCostTblEntry AVX512DQVLConversionTbl[] = {
1787     { ISD::SINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
1788     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
1789     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
1790     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
1791 
1792     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
1793     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
1794     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
1795     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
1796 
1797     { ISD::FP_TO_SINT,  MVT::v2i64,  MVT::v4f32,  1 },
1798     { ISD::FP_TO_SINT,  MVT::v4i64,  MVT::v4f32,  1 },
1799     { ISD::FP_TO_SINT,  MVT::v2i64,  MVT::v2f64,  1 },
1800     { ISD::FP_TO_SINT,  MVT::v4i64,  MVT::v4f64,  1 },
1801 
1802     { ISD::FP_TO_UINT,  MVT::v2i64,  MVT::v4f32,  1 },
1803     { ISD::FP_TO_UINT,  MVT::v4i64,  MVT::v4f32,  1 },
1804     { ISD::FP_TO_UINT,  MVT::v2i64,  MVT::v2f64,  1 },
1805     { ISD::FP_TO_UINT,  MVT::v4i64,  MVT::v4f64,  1 },
1806   };
1807 
1808   static const TypeConversionCostTblEntry AVX512VLConversionTbl[] = {
1809     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i8,   3 }, // sext+vpslld+vptestmd
1810     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i8,   3 }, // sext+vpslld+vptestmd
1811     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i8,   3 }, // sext+vpslld+vptestmd
1812     { ISD::TRUNCATE,  MVT::v16i1,   MVT::v16i8,  8 }, // split+2*v8i8
1813     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i16,  3 }, // sext+vpsllq+vptestmq
1814     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i16,  3 }, // sext+vpsllq+vptestmq
1815     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i16,  3 }, // sext+vpsllq+vptestmq
1816     { ISD::TRUNCATE,  MVT::v16i1,   MVT::v16i16, 8 }, // split+2*v8i16
1817     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i32,  2 }, // vpslld+vptestmd
1818     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i32,  2 }, // vpslld+vptestmd
1819     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i32,  2 }, // vpslld+vptestmd
1820     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i64,  2 }, // vpsllq+vptestmq
1821     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i64,  2 }, // vpsllq+vptestmq
1822     { ISD::TRUNCATE,  MVT::v4i32,   MVT::v4i64,  1 }, // vpmovqd
1823     { ISD::TRUNCATE,  MVT::v4i8,    MVT::v4i64,  2 }, // vpmovqb
1824     { ISD::TRUNCATE,  MVT::v4i16,   MVT::v4i64,  2 }, // vpmovqw
1825     { ISD::TRUNCATE,  MVT::v8i8,    MVT::v8i32,  2 }, // vpmovwb
1826 
1827     // sign extend is vpcmpeq+maskedmove+vpmovdw+vpacksswb
1828     // zero extend is vpcmpeq+maskedmove+vpmovdw+vpsrlw+vpackuswb
1829     { ISD::SIGN_EXTEND, MVT::v2i8,   MVT::v2i1,   5 },
1830     { ISD::ZERO_EXTEND, MVT::v2i8,   MVT::v2i1,   6 },
1831     { ISD::SIGN_EXTEND, MVT::v4i8,   MVT::v4i1,   5 },
1832     { ISD::ZERO_EXTEND, MVT::v4i8,   MVT::v4i1,   6 },
1833     { ISD::SIGN_EXTEND, MVT::v8i8,   MVT::v8i1,   5 },
1834     { ISD::ZERO_EXTEND, MVT::v8i8,   MVT::v8i1,   6 },
1835     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v16i1, 10 },
1836     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v16i1, 12 },
1837 
1838     // sign extend is vpcmpeq+maskedmove+vpmovdw
1839     // zero extend is vpcmpeq+maskedmove+vpmovdw+vpsrlw
1840     { ISD::SIGN_EXTEND, MVT::v2i16,  MVT::v2i1,   4 },
1841     { ISD::ZERO_EXTEND, MVT::v2i16,  MVT::v2i1,   5 },
1842     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i1,   4 },
1843     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i1,   5 },
1844     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i1,   4 },
1845     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i1,   5 },
1846     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 10 },
1847     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 12 },
1848 
1849     { ISD::SIGN_EXTEND, MVT::v2i32,  MVT::v2i1,   1 }, // vpternlogd
1850     { ISD::ZERO_EXTEND, MVT::v2i32,  MVT::v2i1,   2 }, // vpternlogd+psrld
1851     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i1,   1 }, // vpternlogd
1852     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i1,   2 }, // vpternlogd+psrld
1853     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   1 }, // vpternlogd
1854     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   2 }, // vpternlogd+psrld
1855     { ISD::SIGN_EXTEND, MVT::v2i64,  MVT::v2i1,   1 }, // vpternlogq
1856     { ISD::ZERO_EXTEND, MVT::v2i64,  MVT::v2i1,   2 }, // vpternlogq+psrlq
1857     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   1 }, // vpternlogq
1858     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   2 }, // vpternlogq+psrlq
1859 
1860     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v16i8,  1 },
1861     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v16i8,  1 },
1862     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v16i8,  1 },
1863     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v16i8,  1 },
1864     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
1865     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
1866     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v8i16,  1 },
1867     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v8i16,  1 },
1868     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
1869     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
1870     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
1871     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
1872 
1873     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  1 },
1874     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v16i8,  1 },
1875     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  1 },
1876     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  1 },
1877 
1878     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i64,    1 },
1879     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i64,    1 },
1880     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  1 },
1881     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v16i8,  1 },
1882     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  1 },
1883     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  1 },
1884     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  1 },
1885     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  1 },
1886     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  1 },
1887     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  1 },
1888     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  5 },
1889     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  5 },
1890     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  5 },
1891 
1892     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v8f32,  2 },
1893     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v16f32, 2 },
1894     { ISD::FP_TO_SINT,  MVT::v32i8,  MVT::v32f32, 5 },
1895 
1896     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f32,    1 },
1897     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f64,    1 },
1898     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  1 },
1899     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v2f64,  1 },
1900     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f64,  1 },
1901     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f32,  1 },
1902     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f64,  1 },
1903   };
1904 
1905   static const TypeConversionCostTblEntry AVX2ConversionTbl[] = {
1906     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
1907     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
1908     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
1909     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
1910     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1,  1 },
1911     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1,  1 },
1912 
1913     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v16i8,  2 },
1914     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v16i8,  2 },
1915     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v16i8,  2 },
1916     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v16i8,  2 },
1917     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
1918     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
1919     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v8i16,  2 },
1920     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v8i16,  2 },
1921     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
1922     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
1923     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 3 },
1924     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 3 },
1925     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  2 },
1926     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  2 },
1927 
1928     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i32,  2 },
1929 
1930     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v8i16,  1 },
1931     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v4i32,  1 },
1932     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v2i64,  1 },
1933     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v8i32,  4 },
1934     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v4i64,  4 },
1935     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v4i32,  1 },
1936     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v2i64,  1 },
1937     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v4i64,  5 },
1938     { ISD::TRUNCATE,    MVT::v4i32,  MVT::v4i64,  1 },
1939     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  2 },
1940 
1941     { ISD::FP_EXTEND,   MVT::v8f64,  MVT::v8f32,  3 },
1942     { ISD::FP_ROUND,    MVT::v8f32,  MVT::v8f64,  3 },
1943 
1944     { ISD::FP_TO_SINT,  MVT::v16i16, MVT::v8f32,  1 },
1945     { ISD::FP_TO_SINT,  MVT::v4i32,  MVT::v4f64,  1 },
1946     { ISD::FP_TO_SINT,  MVT::v8i32,  MVT::v8f32,  1 },
1947     { ISD::FP_TO_SINT,  MVT::v8i32,  MVT::v8f64,  3 },
1948 
1949     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f32,    3 },
1950     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f64,    3 },
1951     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v8f32,  1 },
1952     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  3 },
1953     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v2f64,  4 },
1954     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f64,  4 },
1955     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f32,  3 },
1956     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v4f64,  4 },
1957 
1958     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  2 },
1959     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v16i8,  2 },
1960     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  2 },
1961     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  2 },
1962     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  1 },
1963     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  1 },
1964     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  3 },
1965 
1966     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  2 },
1967     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v16i8,  2 },
1968     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  2 },
1969     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  2 },
1970     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  2 },
1971     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i32,  1 },
1972     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  2 },
1973     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  2 },
1974     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  2 },
1975     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  4 },
1976   };
1977 
1978   static const TypeConversionCostTblEntry AVXConversionTbl[] = {
1979     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   6 },
1980     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   4 },
1981     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   7 },
1982     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   4 },
1983     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1,  4 },
1984     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1,  4 },
1985 
1986     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v16i8,  3 },
1987     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v16i8,  3 },
1988     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v16i8,  3 },
1989     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v16i8,  3 },
1990     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  3 },
1991     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  3 },
1992     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v8i16,  3 },
1993     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v8i16,  3 },
1994     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  3 },
1995     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  3 },
1996     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  3 },
1997     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  3 },
1998 
1999     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i64,  4 },
2000     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i32,  5 },
2001     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i16, 4 },
2002     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i64,  9 },
2003     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i64, 11 },
2004 
2005     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 2 }, // and+extract+packuswb
2006     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v8i32,  5 },
2007     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  5 },
2008     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v4i64,  5 },
2009     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v4i64,  3 }, // and+extract+2*packusdw
2010     { ISD::TRUNCATE,    MVT::v4i32,  MVT::v4i64,  2 },
2011 
2012     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i1,   3 },
2013     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i1,   3 },
2014     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i1,   8 },
2015     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v16i8,  4 },
2016     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v16i8,  2 },
2017     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  4 },
2018     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v8i16,  2 },
2019     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  2 },
2020     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  2 },
2021     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  4 },
2022     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v2i64,  5 },
2023     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  8 },
2024 
2025     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i1,   7 },
2026     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i1,   7 },
2027     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i1,   6 },
2028     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v16i8,  4 },
2029     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v16i8,  2 },
2030     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  4 },
2031     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v8i16,  2 },
2032     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  4 },
2033     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i32,  4 },
2034     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  5 },
2035     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  6 },
2036     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  8 },
2037     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i32, 10 },
2038     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64, 10 },
2039     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i64, 18 },
2040     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  5 },
2041     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64, 10 },
2042 
2043     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v8f32,  2 },
2044     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v4f64,  2 },
2045     { ISD::FP_TO_SINT,  MVT::v32i8,  MVT::v8f32,  2 },
2046     { ISD::FP_TO_SINT,  MVT::v32i8,  MVT::v4f64,  2 },
2047     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v8f32,  2 },
2048     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v4f64,  2 },
2049     { ISD::FP_TO_SINT,  MVT::v16i16, MVT::v8f32,  2 },
2050     { ISD::FP_TO_SINT,  MVT::v16i16, MVT::v4f64,  2 },
2051     { ISD::FP_TO_SINT,  MVT::v4i32,  MVT::v4f64,  2 },
2052     { ISD::FP_TO_SINT,  MVT::v8i32,  MVT::v8f32,  2 },
2053     { ISD::FP_TO_SINT,  MVT::v8i32,  MVT::v8f64,  5 },
2054 
2055     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v8f32,  2 },
2056     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v4f64,  2 },
2057     { ISD::FP_TO_UINT,  MVT::v32i8,  MVT::v8f32,  2 },
2058     { ISD::FP_TO_UINT,  MVT::v32i8,  MVT::v4f64,  2 },
2059     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v8f32,  2 },
2060     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v4f64,  2 },
2061     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v8f32,  2 },
2062     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v4f64,  2 },
2063     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  3 },
2064     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v2f64,  4 },
2065     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f64,  6 },
2066     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f32,  7 },
2067     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v4f64,  7 },
2068 
2069     { ISD::FP_EXTEND,   MVT::v4f64,  MVT::v4f32,  1 },
2070     { ISD::FP_ROUND,    MVT::v4f32,  MVT::v4f64,  1 },
2071   };
2072 
2073   static const TypeConversionCostTblEntry SSE41ConversionTbl[] = {
2074     { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v16i8,   1 },
2075     { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v16i8,   1 },
2076     { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v16i8,   1 },
2077     { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v16i8,   1 },
2078     { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v16i8,   1 },
2079     { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v16i8,   1 },
2080     { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v8i16,   1 },
2081     { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v8i16,   1 },
2082     { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v8i16,   1 },
2083     { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v8i16,   1 },
2084     { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v4i32,   1 },
2085     { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v4i32,   1 },
2086 
2087     // These truncates end up widening elements.
2088     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i8,   1 }, // PMOVXZBQ
2089     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i16,  1 }, // PMOVXZWQ
2090     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i8,   1 }, // PMOVXZBD
2091 
2092     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v4i32,  2 },
2093     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v4i32,  2 },
2094     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v2i64,  2 },
2095 
2096     { ISD::SINT_TO_FP,  MVT::f32,    MVT::i32,    1 },
2097     { ISD::SINT_TO_FP,  MVT::f64,    MVT::i32,    1 },
2098     { ISD::SINT_TO_FP,  MVT::f32,    MVT::i64,    1 },
2099     { ISD::SINT_TO_FP,  MVT::f64,    MVT::i64,    1 },
2100     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v16i8,  1 },
2101     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  1 },
2102     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v8i16,  1 },
2103     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  1 },
2104     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  1 },
2105     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v4i32,  1 },
2106     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  2 },
2107 
2108     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i32,    1 },
2109     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i32,    1 },
2110     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i64,    4 },
2111     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i64,    4 },
2112     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v16i8,  1 },
2113     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  1 },
2114     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v8i16,  1 },
2115     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  1 },
2116     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  3 },
2117     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  3 },
2118     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v4i32,  2 },
2119     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v2i64, 12 },
2120     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i64, 22 },
2121     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  4 },
2122 
2123     { ISD::FP_TO_SINT,  MVT::i32,    MVT::f32,    1 },
2124     { ISD::FP_TO_SINT,  MVT::i64,    MVT::f32,    1 },
2125     { ISD::FP_TO_SINT,  MVT::i32,    MVT::f64,    1 },
2126     { ISD::FP_TO_SINT,  MVT::i64,    MVT::f64,    1 },
2127     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v4f32,  2 },
2128     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v2f64,  2 },
2129     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v4f32,  1 },
2130     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v2f64,  1 },
2131     { ISD::FP_TO_SINT,  MVT::v4i32,  MVT::v4f32,  1 },
2132     { ISD::FP_TO_SINT,  MVT::v4i32,  MVT::v2f64,  1 },
2133 
2134     { ISD::FP_TO_UINT,  MVT::i32,    MVT::f32,    1 },
2135     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f32,    4 },
2136     { ISD::FP_TO_UINT,  MVT::i32,    MVT::f64,    1 },
2137     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f64,    4 },
2138     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v4f32,  2 },
2139     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v2f64,  2 },
2140     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v4f32,  1 },
2141     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v2f64,  1 },
2142     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  4 },
2143     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v2f64,  4 },
2144   };
2145 
2146   static const TypeConversionCostTblEntry SSE2ConversionTbl[] = {
2147     // These are somewhat magic numbers justified by comparing the
2148     // output of llvm-mca for our various supported scheduler models
2149     // and basing it off the worst case scenario.
2150     { ISD::SINT_TO_FP,  MVT::f32,    MVT::i32,    3 },
2151     { ISD::SINT_TO_FP,  MVT::f64,    MVT::i32,    3 },
2152     { ISD::SINT_TO_FP,  MVT::f32,    MVT::i64,    3 },
2153     { ISD::SINT_TO_FP,  MVT::f64,    MVT::i64,    3 },
2154     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v16i8,  3 },
2155     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  4 },
2156     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v8i16,  3 },
2157     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  4 },
2158     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  3 },
2159     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v4i32,  4 },
2160     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v2i64,  8 },
2161     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  8 },
2162 
2163     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i32,    3 },
2164     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i32,    3 },
2165     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i64,    8 },
2166     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i64,    9 },
2167     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  4 },
2168     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v16i8,  4 },
2169     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v8i16,  4 },
2170     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  4 },
2171     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  7 },
2172     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v4i32,  7 },
2173     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  5 },
2174     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64, 15 },
2175     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v2i64, 18 },
2176 
2177     { ISD::FP_TO_SINT,  MVT::i32,    MVT::f32,    4 },
2178     { ISD::FP_TO_SINT,  MVT::i64,    MVT::f32,    4 },
2179     { ISD::FP_TO_SINT,  MVT::i32,    MVT::f64,    4 },
2180     { ISD::FP_TO_SINT,  MVT::i64,    MVT::f64,    4 },
2181     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v4f32,  6 },
2182     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v2f64,  6 },
2183     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v4f32,  5 },
2184     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v2f64,  5 },
2185     { ISD::FP_TO_SINT,  MVT::v4i32,  MVT::v4f32,  4 },
2186     { ISD::FP_TO_SINT,  MVT::v4i32,  MVT::v2f64,  4 },
2187 
2188     { ISD::FP_TO_UINT,  MVT::i32,    MVT::f32,    4 },
2189     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f32,    4 },
2190     { ISD::FP_TO_UINT,  MVT::i32,    MVT::f64,    4 },
2191     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f64,   15 },
2192     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v4f32,  6 },
2193     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v2f64,  6 },
2194     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v4f32,  5 },
2195     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v2f64,  5 },
2196     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  8 },
2197     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v2f64,  8 },
2198 
2199     { ISD::ZERO_EXTEND, MVT::v2i64,  MVT::v16i8,  4 },
2200     { ISD::SIGN_EXTEND, MVT::v2i64,  MVT::v16i8,  4 },
2201     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v16i8,  2 },
2202     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v16i8,  3 },
2203     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v16i8,  1 },
2204     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v16i8,  2 },
2205     { ISD::ZERO_EXTEND, MVT::v2i64,  MVT::v8i16,  2 },
2206     { ISD::SIGN_EXTEND, MVT::v2i64,  MVT::v8i16,  3 },
2207     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v8i16,  1 },
2208     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v8i16,  2 },
2209     { ISD::ZERO_EXTEND, MVT::v2i64,  MVT::v4i32,  1 },
2210     { ISD::SIGN_EXTEND, MVT::v2i64,  MVT::v4i32,  2 },
2211 
2212     // These truncates are really widening elements.
2213     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i32,  1 }, // PSHUFD
2214     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i16,  2 }, // PUNPCKLWD+DQ
2215     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i8,   3 }, // PUNPCKLBW+WD+PSHUFD
2216     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i16,  1 }, // PUNPCKLWD
2217     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i8,   2 }, // PUNPCKLBW+WD
2218     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i8,   1 }, // PUNPCKLBW
2219 
2220     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v8i16,  2 }, // PAND+PACKUSWB
2221     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 3 },
2222     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v4i32,  3 }, // PAND+2*PACKUSWB
2223     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i32, 7 },
2224     { ISD::TRUNCATE,    MVT::v2i16,  MVT::v2i32,  1 },
2225     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v4i32,  3 },
2226     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  5 },
2227     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32,10 },
2228     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v2i64,  4 }, // PAND+3*PACKUSWB
2229     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v2i64,  2 }, // PSHUFD+PSHUFLW
2230     { ISD::TRUNCATE,    MVT::v4i32,  MVT::v2i64,  1 }, // PSHUFD
2231   };
2232 
2233   // Attempt to map directly to (simple) MVT types to let us match custom entries.
2234   EVT SrcTy = TLI->getValueType(DL, Src);
2235   EVT DstTy = TLI->getValueType(DL, Dst);
2236 
2237   // The function getSimpleVT only handles simple value types.
2238   if (SrcTy.isSimple() && DstTy.isSimple()) {
2239     MVT SimpleSrcTy = SrcTy.getSimpleVT();
2240     MVT SimpleDstTy = DstTy.getSimpleVT();
2241 
2242     if (ST->useAVX512Regs()) {
2243       if (ST->hasBWI())
2244         if (const auto *Entry = ConvertCostTableLookup(
2245                 AVX512BWConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2246           return AdjustCost(Entry->Cost);
2247 
2248       if (ST->hasDQI())
2249         if (const auto *Entry = ConvertCostTableLookup(
2250                 AVX512DQConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2251           return AdjustCost(Entry->Cost);
2252 
2253       if (ST->hasAVX512())
2254         if (const auto *Entry = ConvertCostTableLookup(
2255                 AVX512FConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2256           return AdjustCost(Entry->Cost);
2257     }
2258 
2259     if (ST->hasBWI())
2260       if (const auto *Entry = ConvertCostTableLookup(
2261               AVX512BWVLConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2262         return AdjustCost(Entry->Cost);
2263 
2264     if (ST->hasDQI())
2265       if (const auto *Entry = ConvertCostTableLookup(
2266               AVX512DQVLConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2267         return AdjustCost(Entry->Cost);
2268 
2269     if (ST->hasAVX512())
2270       if (const auto *Entry = ConvertCostTableLookup(AVX512VLConversionTbl, ISD,
2271                                                      SimpleDstTy, SimpleSrcTy))
2272         return AdjustCost(Entry->Cost);
2273 
2274     if (ST->hasAVX2()) {
2275       if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
2276                                                      SimpleDstTy, SimpleSrcTy))
2277         return AdjustCost(Entry->Cost);
2278     }
2279 
2280     if (ST->hasAVX()) {
2281       if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
2282                                                      SimpleDstTy, SimpleSrcTy))
2283         return AdjustCost(Entry->Cost);
2284     }
2285 
2286     if (ST->hasSSE41()) {
2287       if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
2288                                                      SimpleDstTy, SimpleSrcTy))
2289         return AdjustCost(Entry->Cost);
2290     }
2291 
2292     if (ST->hasSSE2()) {
2293       if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
2294                                                      SimpleDstTy, SimpleSrcTy))
2295         return AdjustCost(Entry->Cost);
2296     }
2297   }
2298 
2299   // Fall back to legalized types.
2300   std::pair<InstructionCost, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src);
2301   std::pair<InstructionCost, MVT> LTDest =
2302       TLI->getTypeLegalizationCost(DL, Dst);
2303 
2304   if (ST->useAVX512Regs()) {
2305     if (ST->hasBWI())
2306       if (const auto *Entry = ConvertCostTableLookup(
2307               AVX512BWConversionTbl, ISD, LTDest.second, LTSrc.second))
2308         return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2309 
2310     if (ST->hasDQI())
2311       if (const auto *Entry = ConvertCostTableLookup(
2312               AVX512DQConversionTbl, ISD, LTDest.second, LTSrc.second))
2313         return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2314 
2315     if (ST->hasAVX512())
2316       if (const auto *Entry = ConvertCostTableLookup(
2317               AVX512FConversionTbl, ISD, LTDest.second, LTSrc.second))
2318         return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2319   }
2320 
2321   if (ST->hasBWI())
2322     if (const auto *Entry = ConvertCostTableLookup(AVX512BWVLConversionTbl, ISD,
2323                                                    LTDest.second, LTSrc.second))
2324       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2325 
2326   if (ST->hasDQI())
2327     if (const auto *Entry = ConvertCostTableLookup(AVX512DQVLConversionTbl, ISD,
2328                                                    LTDest.second, LTSrc.second))
2329       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2330 
2331   if (ST->hasAVX512())
2332     if (const auto *Entry = ConvertCostTableLookup(AVX512VLConversionTbl, ISD,
2333                                                    LTDest.second, LTSrc.second))
2334       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2335 
2336   if (ST->hasAVX2())
2337     if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
2338                                                    LTDest.second, LTSrc.second))
2339       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2340 
2341   if (ST->hasAVX())
2342     if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
2343                                                    LTDest.second, LTSrc.second))
2344       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2345 
2346   if (ST->hasSSE41())
2347     if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
2348                                                    LTDest.second, LTSrc.second))
2349       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2350 
2351   if (ST->hasSSE2())
2352     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
2353                                                    LTDest.second, LTSrc.second))
2354       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2355 
2356   // Fallback, for i8/i16 sitofp/uitofp cases we need to extend to i32 for
2357   // sitofp.
2358   if ((ISD == ISD::SINT_TO_FP || ISD == ISD::UINT_TO_FP) &&
2359       1 < Src->getScalarSizeInBits() && Src->getScalarSizeInBits() < 32) {
2360     Type *ExtSrc = Src->getWithNewBitWidth(32);
2361     unsigned ExtOpc =
2362         (ISD == ISD::SINT_TO_FP) ? Instruction::SExt : Instruction::ZExt;
2363 
2364     // For scalar loads the extend would be free.
2365     InstructionCost ExtCost = 0;
2366     if (!(Src->isIntegerTy() && I && isa<LoadInst>(I->getOperand(0))))
2367       ExtCost = getCastInstrCost(ExtOpc, ExtSrc, Src, CCH, CostKind);
2368 
2369     return ExtCost + getCastInstrCost(Instruction::SIToFP, Dst, ExtSrc,
2370                                       TTI::CastContextHint::None, CostKind);
2371   }
2372 
2373   // Fallback for fptosi/fptoui i8/i16 cases we need to truncate from fptosi
2374   // i32.
2375   if ((ISD == ISD::FP_TO_SINT || ISD == ISD::FP_TO_UINT) &&
2376       1 < Dst->getScalarSizeInBits() && Dst->getScalarSizeInBits() < 32) {
2377     Type *TruncDst = Dst->getWithNewBitWidth(32);
2378     return getCastInstrCost(Instruction::FPToSI, TruncDst, Src, CCH, CostKind) +
2379            getCastInstrCost(Instruction::Trunc, Dst, TruncDst,
2380                             TTI::CastContextHint::None, CostKind);
2381   }
2382 
2383   return AdjustCost(
2384       BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I));
2385 }
2386 
2387 InstructionCost X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
2388                                                Type *CondTy,
2389                                                CmpInst::Predicate VecPred,
2390                                                TTI::TargetCostKind CostKind,
2391                                                const Instruction *I) {
2392   // TODO: Handle other cost kinds.
2393   if (CostKind != TTI::TCK_RecipThroughput)
2394     return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
2395                                      I);
2396 
2397   // Legalize the type.
2398   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
2399 
2400   MVT MTy = LT.second;
2401 
2402   int ISD = TLI->InstructionOpcodeToISD(Opcode);
2403   assert(ISD && "Invalid opcode");
2404 
2405   unsigned ExtraCost = 0;
2406   if (I && (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)) {
2407     // Some vector comparison predicates cost extra instructions.
2408     if (MTy.isVector() &&
2409         !((ST->hasXOP() && (!ST->hasAVX2() || MTy.is128BitVector())) ||
2410           (ST->hasAVX512() && 32 <= MTy.getScalarSizeInBits()) ||
2411           ST->hasBWI())) {
2412       switch (cast<CmpInst>(I)->getPredicate()) {
2413       case CmpInst::Predicate::ICMP_NE:
2414         // xor(cmpeq(x,y),-1)
2415         ExtraCost = 1;
2416         break;
2417       case CmpInst::Predicate::ICMP_SGE:
2418       case CmpInst::Predicate::ICMP_SLE:
2419         // xor(cmpgt(x,y),-1)
2420         ExtraCost = 1;
2421         break;
2422       case CmpInst::Predicate::ICMP_ULT:
2423       case CmpInst::Predicate::ICMP_UGT:
2424         // cmpgt(xor(x,signbit),xor(y,signbit))
2425         // xor(cmpeq(pmaxu(x,y),x),-1)
2426         ExtraCost = 2;
2427         break;
2428       case CmpInst::Predicate::ICMP_ULE:
2429       case CmpInst::Predicate::ICMP_UGE:
2430         if ((ST->hasSSE41() && MTy.getScalarSizeInBits() == 32) ||
2431             (ST->hasSSE2() && MTy.getScalarSizeInBits() < 32)) {
2432           // cmpeq(psubus(x,y),0)
2433           // cmpeq(pminu(x,y),x)
2434           ExtraCost = 1;
2435         } else {
2436           // xor(cmpgt(xor(x,signbit),xor(y,signbit)),-1)
2437           ExtraCost = 3;
2438         }
2439         break;
2440       default:
2441         break;
2442       }
2443     }
2444   }
2445 
2446   static const CostTblEntry SLMCostTbl[] = {
2447     // slm pcmpeq/pcmpgt throughput is 2
2448     { ISD::SETCC,   MVT::v2i64,   2 },
2449   };
2450 
2451   static const CostTblEntry AVX512BWCostTbl[] = {
2452     { ISD::SETCC,   MVT::v32i16,  1 },
2453     { ISD::SETCC,   MVT::v64i8,   1 },
2454 
2455     { ISD::SELECT,  MVT::v32i16,  1 },
2456     { ISD::SELECT,  MVT::v64i8,   1 },
2457   };
2458 
2459   static const CostTblEntry AVX512CostTbl[] = {
2460     { ISD::SETCC,   MVT::v8i64,   1 },
2461     { ISD::SETCC,   MVT::v16i32,  1 },
2462     { ISD::SETCC,   MVT::v8f64,   1 },
2463     { ISD::SETCC,   MVT::v16f32,  1 },
2464 
2465     { ISD::SELECT,  MVT::v8i64,   1 },
2466     { ISD::SELECT,  MVT::v16i32,  1 },
2467     { ISD::SELECT,  MVT::v8f64,   1 },
2468     { ISD::SELECT,  MVT::v16f32,  1 },
2469 
2470     { ISD::SETCC,   MVT::v32i16,  2 }, // FIXME: should probably be 4
2471     { ISD::SETCC,   MVT::v64i8,   2 }, // FIXME: should probably be 4
2472 
2473     { ISD::SELECT,  MVT::v32i16,  2 }, // FIXME: should be 3
2474     { ISD::SELECT,  MVT::v64i8,   2 }, // FIXME: should be 3
2475   };
2476 
2477   static const CostTblEntry AVX2CostTbl[] = {
2478     { ISD::SETCC,   MVT::v4i64,   1 },
2479     { ISD::SETCC,   MVT::v8i32,   1 },
2480     { ISD::SETCC,   MVT::v16i16,  1 },
2481     { ISD::SETCC,   MVT::v32i8,   1 },
2482 
2483     { ISD::SELECT,  MVT::v4i64,   1 }, // pblendvb
2484     { ISD::SELECT,  MVT::v8i32,   1 }, // pblendvb
2485     { ISD::SELECT,  MVT::v16i16,  1 }, // pblendvb
2486     { ISD::SELECT,  MVT::v32i8,   1 }, // pblendvb
2487   };
2488 
2489   static const CostTblEntry AVX1CostTbl[] = {
2490     { ISD::SETCC,   MVT::v4f64,   1 },
2491     { ISD::SETCC,   MVT::v8f32,   1 },
2492     // AVX1 does not support 8-wide integer compare.
2493     { ISD::SETCC,   MVT::v4i64,   4 },
2494     { ISD::SETCC,   MVT::v8i32,   4 },
2495     { ISD::SETCC,   MVT::v16i16,  4 },
2496     { ISD::SETCC,   MVT::v32i8,   4 },
2497 
2498     { ISD::SELECT,  MVT::v4f64,   1 }, // vblendvpd
2499     { ISD::SELECT,  MVT::v8f32,   1 }, // vblendvps
2500     { ISD::SELECT,  MVT::v4i64,   1 }, // vblendvpd
2501     { ISD::SELECT,  MVT::v8i32,   1 }, // vblendvps
2502     { ISD::SELECT,  MVT::v16i16,  3 }, // vandps + vandnps + vorps
2503     { ISD::SELECT,  MVT::v32i8,   3 }, // vandps + vandnps + vorps
2504   };
2505 
2506   static const CostTblEntry SSE42CostTbl[] = {
2507     { ISD::SETCC,   MVT::v2f64,   1 },
2508     { ISD::SETCC,   MVT::v4f32,   1 },
2509     { ISD::SETCC,   MVT::v2i64,   1 },
2510   };
2511 
2512   static const CostTblEntry SSE41CostTbl[] = {
2513     { ISD::SELECT,  MVT::v2f64,   1 }, // blendvpd
2514     { ISD::SELECT,  MVT::v4f32,   1 }, // blendvps
2515     { ISD::SELECT,  MVT::v2i64,   1 }, // pblendvb
2516     { ISD::SELECT,  MVT::v4i32,   1 }, // pblendvb
2517     { ISD::SELECT,  MVT::v8i16,   1 }, // pblendvb
2518     { ISD::SELECT,  MVT::v16i8,   1 }, // pblendvb
2519   };
2520 
2521   static const CostTblEntry SSE2CostTbl[] = {
2522     { ISD::SETCC,   MVT::v2f64,   2 },
2523     { ISD::SETCC,   MVT::f64,     1 },
2524     { ISD::SETCC,   MVT::v2i64,   8 },
2525     { ISD::SETCC,   MVT::v4i32,   1 },
2526     { ISD::SETCC,   MVT::v8i16,   1 },
2527     { ISD::SETCC,   MVT::v16i8,   1 },
2528 
2529     { ISD::SELECT,  MVT::v2f64,   3 }, // andpd + andnpd + orpd
2530     { ISD::SELECT,  MVT::v2i64,   3 }, // pand + pandn + por
2531     { ISD::SELECT,  MVT::v4i32,   3 }, // pand + pandn + por
2532     { ISD::SELECT,  MVT::v8i16,   3 }, // pand + pandn + por
2533     { ISD::SELECT,  MVT::v16i8,   3 }, // pand + pandn + por
2534   };
2535 
2536   static const CostTblEntry SSE1CostTbl[] = {
2537     { ISD::SETCC,   MVT::v4f32,   2 },
2538     { ISD::SETCC,   MVT::f32,     1 },
2539 
2540     { ISD::SELECT,  MVT::v4f32,   3 }, // andps + andnps + orps
2541   };
2542 
2543   if (ST->isSLM())
2544     if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy))
2545       return LT.first * (ExtraCost + Entry->Cost);
2546 
2547   if (ST->hasBWI())
2548     if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
2549       return LT.first * (ExtraCost + Entry->Cost);
2550 
2551   if (ST->hasAVX512())
2552     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
2553       return LT.first * (ExtraCost + Entry->Cost);
2554 
2555   if (ST->hasAVX2())
2556     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
2557       return LT.first * (ExtraCost + Entry->Cost);
2558 
2559   if (ST->hasAVX())
2560     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
2561       return LT.first * (ExtraCost + Entry->Cost);
2562 
2563   if (ST->hasSSE42())
2564     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
2565       return LT.first * (ExtraCost + Entry->Cost);
2566 
2567   if (ST->hasSSE41())
2568     if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
2569       return LT.first * (ExtraCost + Entry->Cost);
2570 
2571   if (ST->hasSSE2())
2572     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
2573       return LT.first * (ExtraCost + Entry->Cost);
2574 
2575   if (ST->hasSSE1())
2576     if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
2577       return LT.first * (ExtraCost + Entry->Cost);
2578 
2579   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I);
2580 }
2581 
2582 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; }
2583 
2584 InstructionCost
2585 X86TTIImpl::getTypeBasedIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
2586                                            TTI::TargetCostKind CostKind) {
2587 
2588   // Costs should match the codegen from:
2589   // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll
2590   // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll
2591   // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll
2592   // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll
2593   // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll
2594 
2595   // TODO: Overflow intrinsics (*ADDO, *SUBO, *MULO) with vector types are not
2596   //       specialized in these tables yet.
2597   static const CostTblEntry AVX512BITALGCostTbl[] = {
2598     { ISD::CTPOP,      MVT::v32i16,  1 },
2599     { ISD::CTPOP,      MVT::v64i8,   1 },
2600     { ISD::CTPOP,      MVT::v16i16,  1 },
2601     { ISD::CTPOP,      MVT::v32i8,   1 },
2602     { ISD::CTPOP,      MVT::v8i16,   1 },
2603     { ISD::CTPOP,      MVT::v16i8,   1 },
2604   };
2605   static const CostTblEntry AVX512VPOPCNTDQCostTbl[] = {
2606     { ISD::CTPOP,      MVT::v8i64,   1 },
2607     { ISD::CTPOP,      MVT::v16i32,  1 },
2608     { ISD::CTPOP,      MVT::v4i64,   1 },
2609     { ISD::CTPOP,      MVT::v8i32,   1 },
2610     { ISD::CTPOP,      MVT::v2i64,   1 },
2611     { ISD::CTPOP,      MVT::v4i32,   1 },
2612   };
2613   static const CostTblEntry AVX512CDCostTbl[] = {
2614     { ISD::CTLZ,       MVT::v8i64,   1 },
2615     { ISD::CTLZ,       MVT::v16i32,  1 },
2616     { ISD::CTLZ,       MVT::v32i16,  8 },
2617     { ISD::CTLZ,       MVT::v64i8,  20 },
2618     { ISD::CTLZ,       MVT::v4i64,   1 },
2619     { ISD::CTLZ,       MVT::v8i32,   1 },
2620     { ISD::CTLZ,       MVT::v16i16,  4 },
2621     { ISD::CTLZ,       MVT::v32i8,  10 },
2622     { ISD::CTLZ,       MVT::v2i64,   1 },
2623     { ISD::CTLZ,       MVT::v4i32,   1 },
2624     { ISD::CTLZ,       MVT::v8i16,   4 },
2625     { ISD::CTLZ,       MVT::v16i8,   4 },
2626   };
2627   static const CostTblEntry AVX512BWCostTbl[] = {
2628     { ISD::ABS,        MVT::v32i16,  1 },
2629     { ISD::ABS,        MVT::v64i8,   1 },
2630     { ISD::BITREVERSE, MVT::v8i64,   3 },
2631     { ISD::BITREVERSE, MVT::v16i32,  3 },
2632     { ISD::BITREVERSE, MVT::v32i16,  3 },
2633     { ISD::BITREVERSE, MVT::v64i8,   2 },
2634     { ISD::BSWAP,      MVT::v8i64,   1 },
2635     { ISD::BSWAP,      MVT::v16i32,  1 },
2636     { ISD::BSWAP,      MVT::v32i16,  1 },
2637     { ISD::CTLZ,       MVT::v8i64,  23 },
2638     { ISD::CTLZ,       MVT::v16i32, 22 },
2639     { ISD::CTLZ,       MVT::v32i16, 18 },
2640     { ISD::CTLZ,       MVT::v64i8,  17 },
2641     { ISD::CTPOP,      MVT::v8i64,   7 },
2642     { ISD::CTPOP,      MVT::v16i32, 11 },
2643     { ISD::CTPOP,      MVT::v32i16,  9 },
2644     { ISD::CTPOP,      MVT::v64i8,   6 },
2645     { ISD::CTTZ,       MVT::v8i64,  10 },
2646     { ISD::CTTZ,       MVT::v16i32, 14 },
2647     { ISD::CTTZ,       MVT::v32i16, 12 },
2648     { ISD::CTTZ,       MVT::v64i8,   9 },
2649     { ISD::SADDSAT,    MVT::v32i16,  1 },
2650     { ISD::SADDSAT,    MVT::v64i8,   1 },
2651     { ISD::SMAX,       MVT::v32i16,  1 },
2652     { ISD::SMAX,       MVT::v64i8,   1 },
2653     { ISD::SMIN,       MVT::v32i16,  1 },
2654     { ISD::SMIN,       MVT::v64i8,   1 },
2655     { ISD::SSUBSAT,    MVT::v32i16,  1 },
2656     { ISD::SSUBSAT,    MVT::v64i8,   1 },
2657     { ISD::UADDSAT,    MVT::v32i16,  1 },
2658     { ISD::UADDSAT,    MVT::v64i8,   1 },
2659     { ISD::UMAX,       MVT::v32i16,  1 },
2660     { ISD::UMAX,       MVT::v64i8,   1 },
2661     { ISD::UMIN,       MVT::v32i16,  1 },
2662     { ISD::UMIN,       MVT::v64i8,   1 },
2663     { ISD::USUBSAT,    MVT::v32i16,  1 },
2664     { ISD::USUBSAT,    MVT::v64i8,   1 },
2665   };
2666   static const CostTblEntry AVX512CostTbl[] = {
2667     { ISD::ABS,        MVT::v8i64,   1 },
2668     { ISD::ABS,        MVT::v16i32,  1 },
2669     { ISD::ABS,        MVT::v32i16,  2 }, // FIXME: include split
2670     { ISD::ABS,        MVT::v64i8,   2 }, // FIXME: include split
2671     { ISD::ABS,        MVT::v4i64,   1 },
2672     { ISD::ABS,        MVT::v2i64,   1 },
2673     { ISD::BITREVERSE, MVT::v8i64,  36 },
2674     { ISD::BITREVERSE, MVT::v16i32, 24 },
2675     { ISD::BITREVERSE, MVT::v32i16, 10 },
2676     { ISD::BITREVERSE, MVT::v64i8,  10 },
2677     { ISD::BSWAP,      MVT::v8i64,   4 },
2678     { ISD::BSWAP,      MVT::v16i32,  4 },
2679     { ISD::BSWAP,      MVT::v32i16,  4 },
2680     { ISD::CTLZ,       MVT::v8i64,  29 },
2681     { ISD::CTLZ,       MVT::v16i32, 35 },
2682     { ISD::CTLZ,       MVT::v32i16, 28 },
2683     { ISD::CTLZ,       MVT::v64i8,  18 },
2684     { ISD::CTPOP,      MVT::v8i64,  16 },
2685     { ISD::CTPOP,      MVT::v16i32, 24 },
2686     { ISD::CTPOP,      MVT::v32i16, 18 },
2687     { ISD::CTPOP,      MVT::v64i8,  12 },
2688     { ISD::CTTZ,       MVT::v8i64,  20 },
2689     { ISD::CTTZ,       MVT::v16i32, 28 },
2690     { ISD::CTTZ,       MVT::v32i16, 24 },
2691     { ISD::CTTZ,       MVT::v64i8,  18 },
2692     { ISD::SMAX,       MVT::v8i64,   1 },
2693     { ISD::SMAX,       MVT::v16i32,  1 },
2694     { ISD::SMAX,       MVT::v32i16,  2 }, // FIXME: include split
2695     { ISD::SMAX,       MVT::v64i8,   2 }, // FIXME: include split
2696     { ISD::SMAX,       MVT::v4i64,   1 },
2697     { ISD::SMAX,       MVT::v2i64,   1 },
2698     { ISD::SMIN,       MVT::v8i64,   1 },
2699     { ISD::SMIN,       MVT::v16i32,  1 },
2700     { ISD::SMIN,       MVT::v32i16,  2 }, // FIXME: include split
2701     { ISD::SMIN,       MVT::v64i8,   2 }, // FIXME: include split
2702     { ISD::SMIN,       MVT::v4i64,   1 },
2703     { ISD::SMIN,       MVT::v2i64,   1 },
2704     { ISD::UMAX,       MVT::v8i64,   1 },
2705     { ISD::UMAX,       MVT::v16i32,  1 },
2706     { ISD::UMAX,       MVT::v32i16,  2 }, // FIXME: include split
2707     { ISD::UMAX,       MVT::v64i8,   2 }, // FIXME: include split
2708     { ISD::UMAX,       MVT::v4i64,   1 },
2709     { ISD::UMAX,       MVT::v2i64,   1 },
2710     { ISD::UMIN,       MVT::v8i64,   1 },
2711     { ISD::UMIN,       MVT::v16i32,  1 },
2712     { ISD::UMIN,       MVT::v32i16,  2 }, // FIXME: include split
2713     { ISD::UMIN,       MVT::v64i8,   2 }, // FIXME: include split
2714     { ISD::UMIN,       MVT::v4i64,   1 },
2715     { ISD::UMIN,       MVT::v2i64,   1 },
2716     { ISD::USUBSAT,    MVT::v16i32,  2 }, // pmaxud + psubd
2717     { ISD::USUBSAT,    MVT::v2i64,   2 }, // pmaxuq + psubq
2718     { ISD::USUBSAT,    MVT::v4i64,   2 }, // pmaxuq + psubq
2719     { ISD::USUBSAT,    MVT::v8i64,   2 }, // pmaxuq + psubq
2720     { ISD::UADDSAT,    MVT::v16i32,  3 }, // not + pminud + paddd
2721     { ISD::UADDSAT,    MVT::v2i64,   3 }, // not + pminuq + paddq
2722     { ISD::UADDSAT,    MVT::v4i64,   3 }, // not + pminuq + paddq
2723     { ISD::UADDSAT,    MVT::v8i64,   3 }, // not + pminuq + paddq
2724     { ISD::SADDSAT,    MVT::v32i16,  2 }, // FIXME: include split
2725     { ISD::SADDSAT,    MVT::v64i8,   2 }, // FIXME: include split
2726     { ISD::SSUBSAT,    MVT::v32i16,  2 }, // FIXME: include split
2727     { ISD::SSUBSAT,    MVT::v64i8,   2 }, // FIXME: include split
2728     { ISD::UADDSAT,    MVT::v32i16,  2 }, // FIXME: include split
2729     { ISD::UADDSAT,    MVT::v64i8,   2 }, // FIXME: include split
2730     { ISD::USUBSAT,    MVT::v32i16,  2 }, // FIXME: include split
2731     { ISD::USUBSAT,    MVT::v64i8,   2 }, // FIXME: include split
2732     { ISD::FMAXNUM,    MVT::f32,     2 },
2733     { ISD::FMAXNUM,    MVT::v4f32,   2 },
2734     { ISD::FMAXNUM,    MVT::v8f32,   2 },
2735     { ISD::FMAXNUM,    MVT::v16f32,  2 },
2736     { ISD::FMAXNUM,    MVT::f64,     2 },
2737     { ISD::FMAXNUM,    MVT::v2f64,   2 },
2738     { ISD::FMAXNUM,    MVT::v4f64,   2 },
2739     { ISD::FMAXNUM,    MVT::v8f64,   2 },
2740   };
2741   static const CostTblEntry XOPCostTbl[] = {
2742     { ISD::BITREVERSE, MVT::v4i64,   4 },
2743     { ISD::BITREVERSE, MVT::v8i32,   4 },
2744     { ISD::BITREVERSE, MVT::v16i16,  4 },
2745     { ISD::BITREVERSE, MVT::v32i8,   4 },
2746     { ISD::BITREVERSE, MVT::v2i64,   1 },
2747     { ISD::BITREVERSE, MVT::v4i32,   1 },
2748     { ISD::BITREVERSE, MVT::v8i16,   1 },
2749     { ISD::BITREVERSE, MVT::v16i8,   1 },
2750     { ISD::BITREVERSE, MVT::i64,     3 },
2751     { ISD::BITREVERSE, MVT::i32,     3 },
2752     { ISD::BITREVERSE, MVT::i16,     3 },
2753     { ISD::BITREVERSE, MVT::i8,      3 }
2754   };
2755   static const CostTblEntry AVX2CostTbl[] = {
2756     { ISD::ABS,        MVT::v4i64,   2 }, // VBLENDVPD(X,VPSUBQ(0,X),X)
2757     { ISD::ABS,        MVT::v8i32,   1 },
2758     { ISD::ABS,        MVT::v16i16,  1 },
2759     { ISD::ABS,        MVT::v32i8,   1 },
2760     { ISD::BITREVERSE, MVT::v2i64,   3 },
2761     { ISD::BITREVERSE, MVT::v4i64,   3 },
2762     { ISD::BITREVERSE, MVT::v4i32,   3 },
2763     { ISD::BITREVERSE, MVT::v8i32,   3 },
2764     { ISD::BITREVERSE, MVT::v8i16,   3 },
2765     { ISD::BITREVERSE, MVT::v16i16,  3 },
2766     { ISD::BITREVERSE, MVT::v16i8,   3 },
2767     { ISD::BITREVERSE, MVT::v32i8,   3 },
2768     { ISD::BSWAP,      MVT::v4i64,   1 },
2769     { ISD::BSWAP,      MVT::v8i32,   1 },
2770     { ISD::BSWAP,      MVT::v16i16,  1 },
2771     { ISD::CTLZ,       MVT::v2i64,   7 },
2772     { ISD::CTLZ,       MVT::v4i64,   7 },
2773     { ISD::CTLZ,       MVT::v4i32,   5 },
2774     { ISD::CTLZ,       MVT::v8i32,   5 },
2775     { ISD::CTLZ,       MVT::v8i16,   4 },
2776     { ISD::CTLZ,       MVT::v16i16,  4 },
2777     { ISD::CTLZ,       MVT::v16i8,   3 },
2778     { ISD::CTLZ,       MVT::v32i8,   3 },
2779     { ISD::CTPOP,      MVT::v2i64,   3 },
2780     { ISD::CTPOP,      MVT::v4i64,   3 },
2781     { ISD::CTPOP,      MVT::v4i32,   7 },
2782     { ISD::CTPOP,      MVT::v8i32,   7 },
2783     { ISD::CTPOP,      MVT::v8i16,   3 },
2784     { ISD::CTPOP,      MVT::v16i16,  3 },
2785     { ISD::CTPOP,      MVT::v16i8,   2 },
2786     { ISD::CTPOP,      MVT::v32i8,   2 },
2787     { ISD::CTTZ,       MVT::v2i64,   4 },
2788     { ISD::CTTZ,       MVT::v4i64,   4 },
2789     { ISD::CTTZ,       MVT::v4i32,   7 },
2790     { ISD::CTTZ,       MVT::v8i32,   7 },
2791     { ISD::CTTZ,       MVT::v8i16,   4 },
2792     { ISD::CTTZ,       MVT::v16i16,  4 },
2793     { ISD::CTTZ,       MVT::v16i8,   3 },
2794     { ISD::CTTZ,       MVT::v32i8,   3 },
2795     { ISD::SADDSAT,    MVT::v16i16,  1 },
2796     { ISD::SADDSAT,    MVT::v32i8,   1 },
2797     { ISD::SMAX,       MVT::v8i32,   1 },
2798     { ISD::SMAX,       MVT::v16i16,  1 },
2799     { ISD::SMAX,       MVT::v32i8,   1 },
2800     { ISD::SMIN,       MVT::v8i32,   1 },
2801     { ISD::SMIN,       MVT::v16i16,  1 },
2802     { ISD::SMIN,       MVT::v32i8,   1 },
2803     { ISD::SSUBSAT,    MVT::v16i16,  1 },
2804     { ISD::SSUBSAT,    MVT::v32i8,   1 },
2805     { ISD::UADDSAT,    MVT::v16i16,  1 },
2806     { ISD::UADDSAT,    MVT::v32i8,   1 },
2807     { ISD::UADDSAT,    MVT::v8i32,   3 }, // not + pminud + paddd
2808     { ISD::UMAX,       MVT::v8i32,   1 },
2809     { ISD::UMAX,       MVT::v16i16,  1 },
2810     { ISD::UMAX,       MVT::v32i8,   1 },
2811     { ISD::UMIN,       MVT::v8i32,   1 },
2812     { ISD::UMIN,       MVT::v16i16,  1 },
2813     { ISD::UMIN,       MVT::v32i8,   1 },
2814     { ISD::USUBSAT,    MVT::v16i16,  1 },
2815     { ISD::USUBSAT,    MVT::v32i8,   1 },
2816     { ISD::USUBSAT,    MVT::v8i32,   2 }, // pmaxud + psubd
2817     { ISD::FMAXNUM,    MVT::v8f32,   3 }, // MAXPS + CMPUNORDPS + BLENDVPS
2818     { ISD::FMAXNUM,    MVT::v4f64,   3 }, // MAXPD + CMPUNORDPD + BLENDVPD
2819     { ISD::FSQRT,      MVT::f32,     7 }, // Haswell from http://www.agner.org/
2820     { ISD::FSQRT,      MVT::v4f32,   7 }, // Haswell from http://www.agner.org/
2821     { ISD::FSQRT,      MVT::v8f32,  14 }, // Haswell from http://www.agner.org/
2822     { ISD::FSQRT,      MVT::f64,    14 }, // Haswell from http://www.agner.org/
2823     { ISD::FSQRT,      MVT::v2f64,  14 }, // Haswell from http://www.agner.org/
2824     { ISD::FSQRT,      MVT::v4f64,  28 }, // Haswell from http://www.agner.org/
2825   };
2826   static const CostTblEntry AVX1CostTbl[] = {
2827     { ISD::ABS,        MVT::v4i64,   5 }, // VBLENDVPD(X,VPSUBQ(0,X),X)
2828     { ISD::ABS,        MVT::v8i32,   3 },
2829     { ISD::ABS,        MVT::v16i16,  3 },
2830     { ISD::ABS,        MVT::v32i8,   3 },
2831     { ISD::BITREVERSE, MVT::v4i64,  12 }, // 2 x 128-bit Op + extract/insert
2832     { ISD::BITREVERSE, MVT::v8i32,  12 }, // 2 x 128-bit Op + extract/insert
2833     { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert
2834     { ISD::BITREVERSE, MVT::v32i8,  12 }, // 2 x 128-bit Op + extract/insert
2835     { ISD::BSWAP,      MVT::v4i64,   4 },
2836     { ISD::BSWAP,      MVT::v8i32,   4 },
2837     { ISD::BSWAP,      MVT::v16i16,  4 },
2838     { ISD::CTLZ,       MVT::v4i64,  48 }, // 2 x 128-bit Op + extract/insert
2839     { ISD::CTLZ,       MVT::v8i32,  38 }, // 2 x 128-bit Op + extract/insert
2840     { ISD::CTLZ,       MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert
2841     { ISD::CTLZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
2842     { ISD::CTPOP,      MVT::v4i64,  16 }, // 2 x 128-bit Op + extract/insert
2843     { ISD::CTPOP,      MVT::v8i32,  24 }, // 2 x 128-bit Op + extract/insert
2844     { ISD::CTPOP,      MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert
2845     { ISD::CTPOP,      MVT::v32i8,  14 }, // 2 x 128-bit Op + extract/insert
2846     { ISD::CTTZ,       MVT::v4i64,  22 }, // 2 x 128-bit Op + extract/insert
2847     { ISD::CTTZ,       MVT::v8i32,  30 }, // 2 x 128-bit Op + extract/insert
2848     { ISD::CTTZ,       MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert
2849     { ISD::CTTZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
2850     { ISD::SADDSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
2851     { ISD::SADDSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
2852     { ISD::SMAX,       MVT::v8i32,   4 }, // 2 x 128-bit Op + extract/insert
2853     { ISD::SMAX,       MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
2854     { ISD::SMAX,       MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
2855     { ISD::SMIN,       MVT::v8i32,   4 }, // 2 x 128-bit Op + extract/insert
2856     { ISD::SMIN,       MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
2857     { ISD::SMIN,       MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
2858     { ISD::SSUBSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
2859     { ISD::SSUBSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
2860     { ISD::UADDSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
2861     { ISD::UADDSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
2862     { ISD::UADDSAT,    MVT::v8i32,   8 }, // 2 x 128-bit Op + extract/insert
2863     { ISD::UMAX,       MVT::v8i32,   4 }, // 2 x 128-bit Op + extract/insert
2864     { ISD::UMAX,       MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
2865     { ISD::UMAX,       MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
2866     { ISD::UMIN,       MVT::v8i32,   4 }, // 2 x 128-bit Op + extract/insert
2867     { ISD::UMIN,       MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
2868     { ISD::UMIN,       MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
2869     { ISD::USUBSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
2870     { ISD::USUBSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
2871     { ISD::USUBSAT,    MVT::v8i32,   6 }, // 2 x 128-bit Op + extract/insert
2872     { ISD::FMAXNUM,    MVT::f32,     3 }, // MAXSS + CMPUNORDSS + BLENDVPS
2873     { ISD::FMAXNUM,    MVT::v4f32,   3 }, // MAXPS + CMPUNORDPS + BLENDVPS
2874     { ISD::FMAXNUM,    MVT::v8f32,   5 }, // MAXPS + CMPUNORDPS + BLENDVPS + ?
2875     { ISD::FMAXNUM,    MVT::f64,     3 }, // MAXSD + CMPUNORDSD + BLENDVPD
2876     { ISD::FMAXNUM,    MVT::v2f64,   3 }, // MAXPD + CMPUNORDPD + BLENDVPD
2877     { ISD::FMAXNUM,    MVT::v4f64,   5 }, // MAXPD + CMPUNORDPD + BLENDVPD + ?
2878     { ISD::FSQRT,      MVT::f32,    14 }, // SNB from http://www.agner.org/
2879     { ISD::FSQRT,      MVT::v4f32,  14 }, // SNB from http://www.agner.org/
2880     { ISD::FSQRT,      MVT::v8f32,  28 }, // SNB from http://www.agner.org/
2881     { ISD::FSQRT,      MVT::f64,    21 }, // SNB from http://www.agner.org/
2882     { ISD::FSQRT,      MVT::v2f64,  21 }, // SNB from http://www.agner.org/
2883     { ISD::FSQRT,      MVT::v4f64,  43 }, // SNB from http://www.agner.org/
2884   };
2885   static const CostTblEntry GLMCostTbl[] = {
2886     { ISD::FSQRT, MVT::f32,   19 }, // sqrtss
2887     { ISD::FSQRT, MVT::v4f32, 37 }, // sqrtps
2888     { ISD::FSQRT, MVT::f64,   34 }, // sqrtsd
2889     { ISD::FSQRT, MVT::v2f64, 67 }, // sqrtpd
2890   };
2891   static const CostTblEntry SLMCostTbl[] = {
2892     { ISD::FSQRT, MVT::f32,   20 }, // sqrtss
2893     { ISD::FSQRT, MVT::v4f32, 40 }, // sqrtps
2894     { ISD::FSQRT, MVT::f64,   35 }, // sqrtsd
2895     { ISD::FSQRT, MVT::v2f64, 70 }, // sqrtpd
2896   };
2897   static const CostTblEntry SSE42CostTbl[] = {
2898     { ISD::USUBSAT,    MVT::v4i32,   2 }, // pmaxud + psubd
2899     { ISD::UADDSAT,    MVT::v4i32,   3 }, // not + pminud + paddd
2900     { ISD::FSQRT,      MVT::f32,    18 }, // Nehalem from http://www.agner.org/
2901     { ISD::FSQRT,      MVT::v4f32,  18 }, // Nehalem from http://www.agner.org/
2902   };
2903   static const CostTblEntry SSE41CostTbl[] = {
2904     { ISD::ABS,        MVT::v2i64,   2 }, // BLENDVPD(X,PSUBQ(0,X),X)
2905     { ISD::SMAX,       MVT::v4i32,   1 },
2906     { ISD::SMAX,       MVT::v16i8,   1 },
2907     { ISD::SMIN,       MVT::v4i32,   1 },
2908     { ISD::SMIN,       MVT::v16i8,   1 },
2909     { ISD::UMAX,       MVT::v4i32,   1 },
2910     { ISD::UMAX,       MVT::v8i16,   1 },
2911     { ISD::UMIN,       MVT::v4i32,   1 },
2912     { ISD::UMIN,       MVT::v8i16,   1 },
2913   };
2914   static const CostTblEntry SSSE3CostTbl[] = {
2915     { ISD::ABS,        MVT::v4i32,   1 },
2916     { ISD::ABS,        MVT::v8i16,   1 },
2917     { ISD::ABS,        MVT::v16i8,   1 },
2918     { ISD::BITREVERSE, MVT::v2i64,   5 },
2919     { ISD::BITREVERSE, MVT::v4i32,   5 },
2920     { ISD::BITREVERSE, MVT::v8i16,   5 },
2921     { ISD::BITREVERSE, MVT::v16i8,   5 },
2922     { ISD::BSWAP,      MVT::v2i64,   1 },
2923     { ISD::BSWAP,      MVT::v4i32,   1 },
2924     { ISD::BSWAP,      MVT::v8i16,   1 },
2925     { ISD::CTLZ,       MVT::v2i64,  23 },
2926     { ISD::CTLZ,       MVT::v4i32,  18 },
2927     { ISD::CTLZ,       MVT::v8i16,  14 },
2928     { ISD::CTLZ,       MVT::v16i8,   9 },
2929     { ISD::CTPOP,      MVT::v2i64,   7 },
2930     { ISD::CTPOP,      MVT::v4i32,  11 },
2931     { ISD::CTPOP,      MVT::v8i16,   9 },
2932     { ISD::CTPOP,      MVT::v16i8,   6 },
2933     { ISD::CTTZ,       MVT::v2i64,  10 },
2934     { ISD::CTTZ,       MVT::v4i32,  14 },
2935     { ISD::CTTZ,       MVT::v8i16,  12 },
2936     { ISD::CTTZ,       MVT::v16i8,   9 }
2937   };
2938   static const CostTblEntry SSE2CostTbl[] = {
2939     { ISD::ABS,        MVT::v2i64,   4 },
2940     { ISD::ABS,        MVT::v4i32,   3 },
2941     { ISD::ABS,        MVT::v8i16,   2 },
2942     { ISD::ABS,        MVT::v16i8,   2 },
2943     { ISD::BITREVERSE, MVT::v2i64,  29 },
2944     { ISD::BITREVERSE, MVT::v4i32,  27 },
2945     { ISD::BITREVERSE, MVT::v8i16,  27 },
2946     { ISD::BITREVERSE, MVT::v16i8,  20 },
2947     { ISD::BSWAP,      MVT::v2i64,   7 },
2948     { ISD::BSWAP,      MVT::v4i32,   7 },
2949     { ISD::BSWAP,      MVT::v8i16,   7 },
2950     { ISD::CTLZ,       MVT::v2i64,  25 },
2951     { ISD::CTLZ,       MVT::v4i32,  26 },
2952     { ISD::CTLZ,       MVT::v8i16,  20 },
2953     { ISD::CTLZ,       MVT::v16i8,  17 },
2954     { ISD::CTPOP,      MVT::v2i64,  12 },
2955     { ISD::CTPOP,      MVT::v4i32,  15 },
2956     { ISD::CTPOP,      MVT::v8i16,  13 },
2957     { ISD::CTPOP,      MVT::v16i8,  10 },
2958     { ISD::CTTZ,       MVT::v2i64,  14 },
2959     { ISD::CTTZ,       MVT::v4i32,  18 },
2960     { ISD::CTTZ,       MVT::v8i16,  16 },
2961     { ISD::CTTZ,       MVT::v16i8,  13 },
2962     { ISD::SADDSAT,    MVT::v8i16,   1 },
2963     { ISD::SADDSAT,    MVT::v16i8,   1 },
2964     { ISD::SMAX,       MVT::v8i16,   1 },
2965     { ISD::SMIN,       MVT::v8i16,   1 },
2966     { ISD::SSUBSAT,    MVT::v8i16,   1 },
2967     { ISD::SSUBSAT,    MVT::v16i8,   1 },
2968     { ISD::UADDSAT,    MVT::v8i16,   1 },
2969     { ISD::UADDSAT,    MVT::v16i8,   1 },
2970     { ISD::UMAX,       MVT::v8i16,   2 },
2971     { ISD::UMAX,       MVT::v16i8,   1 },
2972     { ISD::UMIN,       MVT::v8i16,   2 },
2973     { ISD::UMIN,       MVT::v16i8,   1 },
2974     { ISD::USUBSAT,    MVT::v8i16,   1 },
2975     { ISD::USUBSAT,    MVT::v16i8,   1 },
2976     { ISD::FMAXNUM,    MVT::f64,     4 },
2977     { ISD::FMAXNUM,    MVT::v2f64,   4 },
2978     { ISD::FSQRT,      MVT::f64,    32 }, // Nehalem from http://www.agner.org/
2979     { ISD::FSQRT,      MVT::v2f64,  32 }, // Nehalem from http://www.agner.org/
2980   };
2981   static const CostTblEntry SSE1CostTbl[] = {
2982     { ISD::FMAXNUM,    MVT::f32,     4 },
2983     { ISD::FMAXNUM,    MVT::v4f32,   4 },
2984     { ISD::FSQRT,      MVT::f32,    28 }, // Pentium III from http://www.agner.org/
2985     { ISD::FSQRT,      MVT::v4f32,  56 }, // Pentium III from http://www.agner.org/
2986   };
2987   static const CostTblEntry BMI64CostTbl[] = { // 64-bit targets
2988     { ISD::CTTZ,       MVT::i64,     1 },
2989   };
2990   static const CostTblEntry BMI32CostTbl[] = { // 32 or 64-bit targets
2991     { ISD::CTTZ,       MVT::i32,     1 },
2992     { ISD::CTTZ,       MVT::i16,     1 },
2993     { ISD::CTTZ,       MVT::i8,      1 },
2994   };
2995   static const CostTblEntry LZCNT64CostTbl[] = { // 64-bit targets
2996     { ISD::CTLZ,       MVT::i64,     1 },
2997   };
2998   static const CostTblEntry LZCNT32CostTbl[] = { // 32 or 64-bit targets
2999     { ISD::CTLZ,       MVT::i32,     1 },
3000     { ISD::CTLZ,       MVT::i16,     1 },
3001     { ISD::CTLZ,       MVT::i8,      1 },
3002   };
3003   static const CostTblEntry POPCNT64CostTbl[] = { // 64-bit targets
3004     { ISD::CTPOP,      MVT::i64,     1 },
3005   };
3006   static const CostTblEntry POPCNT32CostTbl[] = { // 32 or 64-bit targets
3007     { ISD::CTPOP,      MVT::i32,     1 },
3008     { ISD::CTPOP,      MVT::i16,     1 },
3009     { ISD::CTPOP,      MVT::i8,      1 },
3010   };
3011   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
3012     { ISD::ABS,        MVT::i64,     2 }, // SUB+CMOV
3013     { ISD::BITREVERSE, MVT::i64,    14 },
3014     { ISD::BSWAP,      MVT::i64,     1 },
3015     { ISD::CTLZ,       MVT::i64,     4 }, // BSR+XOR or BSR+XOR+CMOV
3016     { ISD::CTTZ,       MVT::i64,     3 }, // TEST+BSF+CMOV/BRANCH
3017     { ISD::CTPOP,      MVT::i64,    10 },
3018     { ISD::SADDO,      MVT::i64,     1 },
3019     { ISD::UADDO,      MVT::i64,     1 },
3020     { ISD::UMULO,      MVT::i64,     2 }, // mulq + seto
3021   };
3022   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
3023     { ISD::ABS,        MVT::i32,     2 }, // SUB+CMOV
3024     { ISD::ABS,        MVT::i16,     2 }, // SUB+CMOV
3025     { ISD::BITREVERSE, MVT::i32,    14 },
3026     { ISD::BITREVERSE, MVT::i16,    14 },
3027     { ISD::BITREVERSE, MVT::i8,     11 },
3028     { ISD::BSWAP,      MVT::i32,     1 },
3029     { ISD::BSWAP,      MVT::i16,     1 }, // ROL
3030     { ISD::CTLZ,       MVT::i32,     4 }, // BSR+XOR or BSR+XOR+CMOV
3031     { ISD::CTLZ,       MVT::i16,     4 }, // BSR+XOR or BSR+XOR+CMOV
3032     { ISD::CTLZ,       MVT::i8,      4 }, // BSR+XOR or BSR+XOR+CMOV
3033     { ISD::CTTZ,       MVT::i32,     3 }, // TEST+BSF+CMOV/BRANCH
3034     { ISD::CTTZ,       MVT::i16,     3 }, // TEST+BSF+CMOV/BRANCH
3035     { ISD::CTTZ,       MVT::i8,      3 }, // TEST+BSF+CMOV/BRANCH
3036     { ISD::CTPOP,      MVT::i32,     8 },
3037     { ISD::CTPOP,      MVT::i16,     9 },
3038     { ISD::CTPOP,      MVT::i8,      7 },
3039     { ISD::SADDO,      MVT::i32,     1 },
3040     { ISD::SADDO,      MVT::i16,     1 },
3041     { ISD::SADDO,      MVT::i8,      1 },
3042     { ISD::UADDO,      MVT::i32,     1 },
3043     { ISD::UADDO,      MVT::i16,     1 },
3044     { ISD::UADDO,      MVT::i8,      1 },
3045     { ISD::UMULO,      MVT::i32,     2 }, // mul + seto
3046     { ISD::UMULO,      MVT::i16,     2 },
3047     { ISD::UMULO,      MVT::i8,      2 },
3048   };
3049 
3050   Type *RetTy = ICA.getReturnType();
3051   Type *OpTy = RetTy;
3052   Intrinsic::ID IID = ICA.getID();
3053   unsigned ISD = ISD::DELETED_NODE;
3054   switch (IID) {
3055   default:
3056     break;
3057   case Intrinsic::abs:
3058     ISD = ISD::ABS;
3059     break;
3060   case Intrinsic::bitreverse:
3061     ISD = ISD::BITREVERSE;
3062     break;
3063   case Intrinsic::bswap:
3064     ISD = ISD::BSWAP;
3065     break;
3066   case Intrinsic::ctlz:
3067     ISD = ISD::CTLZ;
3068     break;
3069   case Intrinsic::ctpop:
3070     ISD = ISD::CTPOP;
3071     break;
3072   case Intrinsic::cttz:
3073     ISD = ISD::CTTZ;
3074     break;
3075   case Intrinsic::maxnum:
3076   case Intrinsic::minnum:
3077     // FMINNUM has same costs so don't duplicate.
3078     ISD = ISD::FMAXNUM;
3079     break;
3080   case Intrinsic::sadd_sat:
3081     ISD = ISD::SADDSAT;
3082     break;
3083   case Intrinsic::smax:
3084     ISD = ISD::SMAX;
3085     break;
3086   case Intrinsic::smin:
3087     ISD = ISD::SMIN;
3088     break;
3089   case Intrinsic::ssub_sat:
3090     ISD = ISD::SSUBSAT;
3091     break;
3092   case Intrinsic::uadd_sat:
3093     ISD = ISD::UADDSAT;
3094     break;
3095   case Intrinsic::umax:
3096     ISD = ISD::UMAX;
3097     break;
3098   case Intrinsic::umin:
3099     ISD = ISD::UMIN;
3100     break;
3101   case Intrinsic::usub_sat:
3102     ISD = ISD::USUBSAT;
3103     break;
3104   case Intrinsic::sqrt:
3105     ISD = ISD::FSQRT;
3106     break;
3107   case Intrinsic::sadd_with_overflow:
3108   case Intrinsic::ssub_with_overflow:
3109     // SSUBO has same costs so don't duplicate.
3110     ISD = ISD::SADDO;
3111     OpTy = RetTy->getContainedType(0);
3112     break;
3113   case Intrinsic::uadd_with_overflow:
3114   case Intrinsic::usub_with_overflow:
3115     // USUBO has same costs so don't duplicate.
3116     ISD = ISD::UADDO;
3117     OpTy = RetTy->getContainedType(0);
3118     break;
3119   case Intrinsic::umul_with_overflow:
3120   case Intrinsic::smul_with_overflow:
3121     // SMULO has same costs so don't duplicate.
3122     ISD = ISD::UMULO;
3123     OpTy = RetTy->getContainedType(0);
3124     break;
3125   }
3126 
3127   if (ISD != ISD::DELETED_NODE) {
3128     // Legalize the type.
3129     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, OpTy);
3130     MVT MTy = LT.second;
3131 
3132     // Attempt to lookup cost.
3133     if (ISD == ISD::BITREVERSE && ST->hasGFNI() && ST->hasSSSE3() &&
3134         MTy.isVector()) {
3135       // With PSHUFB the code is very similar for all types. If we have integer
3136       // byte operations, we just need a GF2P8AFFINEQB for vXi8. For other types
3137       // we also need a PSHUFB.
3138       unsigned Cost = MTy.getVectorElementType() == MVT::i8 ? 1 : 2;
3139 
3140       // Without byte operations, we need twice as many GF2P8AFFINEQB and PSHUFB
3141       // instructions. We also need an extract and an insert.
3142       if (!(MTy.is128BitVector() || (ST->hasAVX2() && MTy.is256BitVector()) ||
3143             (ST->hasBWI() && MTy.is512BitVector())))
3144         Cost = Cost * 2 + 2;
3145 
3146       return LT.first * Cost;
3147     }
3148 
3149     auto adjustTableCost = [](const CostTblEntry &Entry,
3150                               InstructionCost LegalizationCost,
3151                               FastMathFlags FMF) {
3152       // If there are no NANs to deal with, then these are reduced to a
3153       // single MIN** or MAX** instruction instead of the MIN/CMP/SELECT that we
3154       // assume is used in the non-fast case.
3155       if (Entry.ISD == ISD::FMAXNUM || Entry.ISD == ISD::FMINNUM) {
3156         if (FMF.noNaNs())
3157           return LegalizationCost * 1;
3158       }
3159       return LegalizationCost * (int)Entry.Cost;
3160     };
3161 
3162     if (ST->useGLMDivSqrtCosts())
3163       if (const auto *Entry = CostTableLookup(GLMCostTbl, ISD, MTy))
3164         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3165 
3166     if (ST->isSLM())
3167       if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy))
3168         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3169 
3170     if (ST->hasBITALG())
3171       if (const auto *Entry = CostTableLookup(AVX512BITALGCostTbl, ISD, MTy))
3172         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3173 
3174     if (ST->hasVPOPCNTDQ())
3175       if (const auto *Entry = CostTableLookup(AVX512VPOPCNTDQCostTbl, ISD, MTy))
3176         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3177 
3178     if (ST->hasCDI())
3179       if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy))
3180         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3181 
3182     if (ST->hasBWI())
3183       if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
3184         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3185 
3186     if (ST->hasAVX512())
3187       if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
3188         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3189 
3190     if (ST->hasXOP())
3191       if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
3192         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3193 
3194     if (ST->hasAVX2())
3195       if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
3196         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3197 
3198     if (ST->hasAVX())
3199       if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
3200         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3201 
3202     if (ST->hasSSE42())
3203       if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
3204         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3205 
3206     if (ST->hasSSE41())
3207       if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
3208         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3209 
3210     if (ST->hasSSSE3())
3211       if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy))
3212         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3213 
3214     if (ST->hasSSE2())
3215       if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
3216         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3217 
3218     if (ST->hasSSE1())
3219       if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
3220         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3221 
3222     if (ST->hasBMI()) {
3223       if (ST->is64Bit())
3224         if (const auto *Entry = CostTableLookup(BMI64CostTbl, ISD, MTy))
3225           return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3226 
3227       if (const auto *Entry = CostTableLookup(BMI32CostTbl, ISD, MTy))
3228         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3229     }
3230 
3231     if (ST->hasLZCNT()) {
3232       if (ST->is64Bit())
3233         if (const auto *Entry = CostTableLookup(LZCNT64CostTbl, ISD, MTy))
3234           return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3235 
3236       if (const auto *Entry = CostTableLookup(LZCNT32CostTbl, ISD, MTy))
3237         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3238     }
3239 
3240     if (ST->hasPOPCNT()) {
3241       if (ST->is64Bit())
3242         if (const auto *Entry = CostTableLookup(POPCNT64CostTbl, ISD, MTy))
3243           return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3244 
3245       if (const auto *Entry = CostTableLookup(POPCNT32CostTbl, ISD, MTy))
3246         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3247     }
3248 
3249     if (ISD == ISD::BSWAP && ST->hasMOVBE() && ST->hasFastMOVBE()) {
3250       if (const Instruction *II = ICA.getInst()) {
3251         if (II->hasOneUse() && isa<StoreInst>(II->user_back()))
3252           return TTI::TCC_Free;
3253         if (auto *LI = dyn_cast<LoadInst>(II->getOperand(0))) {
3254           if (LI->hasOneUse())
3255             return TTI::TCC_Free;
3256         }
3257       }
3258     }
3259 
3260     // TODO - add BMI (TZCNT) scalar handling
3261 
3262     if (ST->is64Bit())
3263       if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
3264         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3265 
3266     if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
3267       return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3268   }
3269 
3270   return BaseT::getIntrinsicInstrCost(ICA, CostKind);
3271 }
3272 
3273 InstructionCost
3274 X86TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
3275                                   TTI::TargetCostKind CostKind) {
3276   if (ICA.isTypeBasedOnly())
3277     return getTypeBasedIntrinsicInstrCost(ICA, CostKind);
3278 
3279   static const CostTblEntry AVX512CostTbl[] = {
3280     { ISD::ROTL,       MVT::v8i64,   1 },
3281     { ISD::ROTL,       MVT::v4i64,   1 },
3282     { ISD::ROTL,       MVT::v2i64,   1 },
3283     { ISD::ROTL,       MVT::v16i32,  1 },
3284     { ISD::ROTL,       MVT::v8i32,   1 },
3285     { ISD::ROTL,       MVT::v4i32,   1 },
3286     { ISD::ROTR,       MVT::v8i64,   1 },
3287     { ISD::ROTR,       MVT::v4i64,   1 },
3288     { ISD::ROTR,       MVT::v2i64,   1 },
3289     { ISD::ROTR,       MVT::v16i32,  1 },
3290     { ISD::ROTR,       MVT::v8i32,   1 },
3291     { ISD::ROTR,       MVT::v4i32,   1 }
3292   };
3293   // XOP: ROTL = VPROT(X,Y), ROTR = VPROT(X,SUB(0,Y))
3294   static const CostTblEntry XOPCostTbl[] = {
3295     { ISD::ROTL,       MVT::v4i64,   4 },
3296     { ISD::ROTL,       MVT::v8i32,   4 },
3297     { ISD::ROTL,       MVT::v16i16,  4 },
3298     { ISD::ROTL,       MVT::v32i8,   4 },
3299     { ISD::ROTL,       MVT::v2i64,   1 },
3300     { ISD::ROTL,       MVT::v4i32,   1 },
3301     { ISD::ROTL,       MVT::v8i16,   1 },
3302     { ISD::ROTL,       MVT::v16i8,   1 },
3303     { ISD::ROTR,       MVT::v4i64,   6 },
3304     { ISD::ROTR,       MVT::v8i32,   6 },
3305     { ISD::ROTR,       MVT::v16i16,  6 },
3306     { ISD::ROTR,       MVT::v32i8,   6 },
3307     { ISD::ROTR,       MVT::v2i64,   2 },
3308     { ISD::ROTR,       MVT::v4i32,   2 },
3309     { ISD::ROTR,       MVT::v8i16,   2 },
3310     { ISD::ROTR,       MVT::v16i8,   2 }
3311   };
3312   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
3313     { ISD::ROTL,       MVT::i64,     1 },
3314     { ISD::ROTR,       MVT::i64,     1 },
3315     { ISD::FSHL,       MVT::i64,     4 }
3316   };
3317   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
3318     { ISD::ROTL,       MVT::i32,     1 },
3319     { ISD::ROTL,       MVT::i16,     1 },
3320     { ISD::ROTL,       MVT::i8,      1 },
3321     { ISD::ROTR,       MVT::i32,     1 },
3322     { ISD::ROTR,       MVT::i16,     1 },
3323     { ISD::ROTR,       MVT::i8,      1 },
3324     { ISD::FSHL,       MVT::i32,     4 },
3325     { ISD::FSHL,       MVT::i16,     4 },
3326     { ISD::FSHL,       MVT::i8,      4 }
3327   };
3328 
3329   Intrinsic::ID IID = ICA.getID();
3330   Type *RetTy = ICA.getReturnType();
3331   const SmallVectorImpl<const Value *> &Args = ICA.getArgs();
3332   unsigned ISD = ISD::DELETED_NODE;
3333   switch (IID) {
3334   default:
3335     break;
3336   case Intrinsic::fshl:
3337     ISD = ISD::FSHL;
3338     if (Args[0] == Args[1])
3339       ISD = ISD::ROTL;
3340     break;
3341   case Intrinsic::fshr:
3342     // FSHR has same costs so don't duplicate.
3343     ISD = ISD::FSHL;
3344     if (Args[0] == Args[1])
3345       ISD = ISD::ROTR;
3346     break;
3347   }
3348 
3349   if (ISD != ISD::DELETED_NODE) {
3350     // Legalize the type.
3351     std::pair<InstructionCost, MVT> LT =
3352         TLI->getTypeLegalizationCost(DL, RetTy);
3353     MVT MTy = LT.second;
3354 
3355     // Attempt to lookup cost.
3356     if (ST->hasAVX512())
3357       if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
3358         return LT.first * Entry->Cost;
3359 
3360     if (ST->hasXOP())
3361       if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
3362         return LT.first * Entry->Cost;
3363 
3364     if (ST->is64Bit())
3365       if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
3366         return LT.first * Entry->Cost;
3367 
3368     if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
3369       return LT.first * Entry->Cost;
3370   }
3371 
3372   return BaseT::getIntrinsicInstrCost(ICA, CostKind);
3373 }
3374 
3375 InstructionCost X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
3376                                                unsigned Index) {
3377   static const CostTblEntry SLMCostTbl[] = {
3378      { ISD::EXTRACT_VECTOR_ELT,       MVT::i8,      4 },
3379      { ISD::EXTRACT_VECTOR_ELT,       MVT::i16,     4 },
3380      { ISD::EXTRACT_VECTOR_ELT,       MVT::i32,     4 },
3381      { ISD::EXTRACT_VECTOR_ELT,       MVT::i64,     7 }
3382    };
3383 
3384   assert(Val->isVectorTy() && "This must be a vector type");
3385   Type *ScalarType = Val->getScalarType();
3386   int RegisterFileMoveCost = 0;
3387 
3388   // Non-immediate extraction/insertion can be handled as a sequence of
3389   // aliased loads+stores via the stack.
3390   if (Index == -1U && (Opcode == Instruction::ExtractElement ||
3391                        Opcode == Instruction::InsertElement)) {
3392     // TODO: On some SSE41+ targets, we expand to cmp+splat+select patterns:
3393     // inselt N0, N1, N2 --> select (SplatN2 == {0,1,2...}) ? SplatN1 : N0.
3394 
3395     // TODO: Move this to BasicTTIImpl.h? We'd need better gep + index handling.
3396     assert(isa<FixedVectorType>(Val) && "Fixed vector type expected");
3397     Align VecAlign = DL.getPrefTypeAlign(Val);
3398     Align SclAlign = DL.getPrefTypeAlign(ScalarType);
3399 
3400     // Extract - store vector to stack, load scalar.
3401     if (Opcode == Instruction::ExtractElement) {
3402       return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0,
3403                              TTI::TargetCostKind::TCK_RecipThroughput) +
3404              getMemoryOpCost(Instruction::Load, ScalarType, SclAlign, 0,
3405                              TTI::TargetCostKind::TCK_RecipThroughput);
3406     }
3407     // Insert - store vector to stack, store scalar, load vector.
3408     if (Opcode == Instruction::InsertElement) {
3409       return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0,
3410                              TTI::TargetCostKind::TCK_RecipThroughput) +
3411              getMemoryOpCost(Instruction::Store, ScalarType, SclAlign, 0,
3412                              TTI::TargetCostKind::TCK_RecipThroughput) +
3413              getMemoryOpCost(Instruction::Load, Val, VecAlign, 0,
3414                              TTI::TargetCostKind::TCK_RecipThroughput);
3415     }
3416   }
3417 
3418   if (Index != -1U && (Opcode == Instruction::ExtractElement ||
3419                        Opcode == Instruction::InsertElement)) {
3420     // Legalize the type.
3421     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Val);
3422 
3423     // This type is legalized to a scalar type.
3424     if (!LT.second.isVector())
3425       return 0;
3426 
3427     // The type may be split. Normalize the index to the new type.
3428     unsigned NumElts = LT.second.getVectorNumElements();
3429     unsigned SubNumElts = NumElts;
3430     Index = Index % NumElts;
3431 
3432     // For >128-bit vectors, we need to extract higher 128-bit subvectors.
3433     // For inserts, we also need to insert the subvector back.
3434     if (LT.second.getSizeInBits() > 128) {
3435       assert((LT.second.getSizeInBits() % 128) == 0 && "Illegal vector");
3436       unsigned NumSubVecs = LT.second.getSizeInBits() / 128;
3437       SubNumElts = NumElts / NumSubVecs;
3438       if (SubNumElts <= Index) {
3439         RegisterFileMoveCost += (Opcode == Instruction::InsertElement ? 2 : 1);
3440         Index %= SubNumElts;
3441       }
3442     }
3443 
3444     if (Index == 0) {
3445       // Floating point scalars are already located in index #0.
3446       // Many insertions to #0 can fold away for scalar fp-ops, so let's assume
3447       // true for all.
3448       if (ScalarType->isFloatingPointTy())
3449         return RegisterFileMoveCost;
3450 
3451       // Assume movd/movq XMM -> GPR is relatively cheap on all targets.
3452       if (ScalarType->isIntegerTy() && Opcode == Instruction::ExtractElement)
3453         return 1 + RegisterFileMoveCost;
3454     }
3455 
3456     int ISD = TLI->InstructionOpcodeToISD(Opcode);
3457     assert(ISD && "Unexpected vector opcode");
3458     MVT MScalarTy = LT.second.getScalarType();
3459     if (ST->isSLM())
3460       if (auto *Entry = CostTableLookup(SLMCostTbl, ISD, MScalarTy))
3461         return Entry->Cost + RegisterFileMoveCost;
3462 
3463     // Assume pinsr/pextr XMM <-> GPR is relatively cheap on all targets.
3464     if ((MScalarTy == MVT::i16 && ST->hasSSE2()) ||
3465         (MScalarTy.isInteger() && ST->hasSSE41()))
3466       return 1 + RegisterFileMoveCost;
3467 
3468     // Assume insertps is relatively cheap on all targets.
3469     if (MScalarTy == MVT::f32 && ST->hasSSE41() &&
3470         Opcode == Instruction::InsertElement)
3471       return 1 + RegisterFileMoveCost;
3472 
3473     // For extractions we just need to shuffle the element to index 0, which
3474     // should be very cheap (assume cost = 1). For insertions we need to shuffle
3475     // the elements to its destination. In both cases we must handle the
3476     // subvector move(s).
3477     // If the vector type is already less than 128-bits then don't reduce it.
3478     // TODO: Under what circumstances should we shuffle using the full width?
3479     InstructionCost ShuffleCost = 1;
3480     if (Opcode == Instruction::InsertElement) {
3481       auto *SubTy = cast<VectorType>(Val);
3482       EVT VT = TLI->getValueType(DL, Val);
3483       if (VT.getScalarType() != MScalarTy || VT.getSizeInBits() >= 128)
3484         SubTy = FixedVectorType::get(ScalarType, SubNumElts);
3485       ShuffleCost =
3486           getShuffleCost(TTI::SK_PermuteTwoSrc, SubTy, None, 0, SubTy);
3487     }
3488     int IntOrFpCost = ScalarType->isFloatingPointTy() ? 0 : 1;
3489     return ShuffleCost + IntOrFpCost + RegisterFileMoveCost;
3490   }
3491 
3492   // Add to the base cost if we know that the extracted element of a vector is
3493   // destined to be moved to and used in the integer register file.
3494   if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy())
3495     RegisterFileMoveCost += 1;
3496 
3497   return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost;
3498 }
3499 
3500 InstructionCost X86TTIImpl::getScalarizationOverhead(VectorType *Ty,
3501                                                      const APInt &DemandedElts,
3502                                                      bool Insert,
3503                                                      bool Extract) {
3504   InstructionCost Cost = 0;
3505 
3506   // For insertions, a ISD::BUILD_VECTOR style vector initialization can be much
3507   // cheaper than an accumulation of ISD::INSERT_VECTOR_ELT.
3508   if (Insert) {
3509     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
3510     MVT MScalarTy = LT.second.getScalarType();
3511 
3512     if ((MScalarTy == MVT::i16 && ST->hasSSE2()) ||
3513         (MScalarTy.isInteger() && ST->hasSSE41()) ||
3514         (MScalarTy == MVT::f32 && ST->hasSSE41())) {
3515       // For types we can insert directly, insertion into 128-bit sub vectors is
3516       // cheap, followed by a cheap chain of concatenations.
3517       if (LT.second.getSizeInBits() <= 128) {
3518         Cost +=
3519             BaseT::getScalarizationOverhead(Ty, DemandedElts, Insert, false);
3520       } else {
3521         // In each 128-lane, if at least one index is demanded but not all
3522         // indices are demanded and this 128-lane is not the first 128-lane of
3523         // the legalized-vector, then this 128-lane needs a extracti128; If in
3524         // each 128-lane, there is at least one demanded index, this 128-lane
3525         // needs a inserti128.
3526 
3527         // The following cases will help you build a better understanding:
3528         // Assume we insert several elements into a v8i32 vector in avx2,
3529         // Case#1: inserting into 1th index needs vpinsrd + inserti128.
3530         // Case#2: inserting into 5th index needs extracti128 + vpinsrd +
3531         // inserti128.
3532         // Case#3: inserting into 4,5,6,7 index needs 4*vpinsrd + inserti128.
3533         const int CostValue = *LT.first.getValue();
3534         assert(CostValue >= 0 && "Negative cost!");
3535         unsigned Num128Lanes = LT.second.getSizeInBits() / 128 * CostValue;
3536         unsigned NumElts = LT.second.getVectorNumElements() * CostValue;
3537         APInt WidenedDemandedElts = DemandedElts.zextOrSelf(NumElts);
3538         unsigned Scale = NumElts / Num128Lanes;
3539         // We iterate each 128-lane, and check if we need a
3540         // extracti128/inserti128 for this 128-lane.
3541         for (unsigned I = 0; I < NumElts; I += Scale) {
3542           APInt Mask = WidenedDemandedElts.getBitsSet(NumElts, I, I + Scale);
3543           APInt MaskedDE = Mask & WidenedDemandedElts;
3544           unsigned Population = MaskedDE.countPopulation();
3545           Cost += (Population > 0 && Population != Scale &&
3546                    I % LT.second.getVectorNumElements() != 0);
3547           Cost += Population > 0;
3548         }
3549         Cost += DemandedElts.countPopulation();
3550 
3551         // For vXf32 cases, insertion into the 0'th index in each v4f32
3552         // 128-bit vector is free.
3553         // NOTE: This assumes legalization widens vXf32 vectors.
3554         if (MScalarTy == MVT::f32)
3555           for (unsigned i = 0, e = cast<FixedVectorType>(Ty)->getNumElements();
3556                i < e; i += 4)
3557             if (DemandedElts[i])
3558               Cost--;
3559       }
3560     } else if (LT.second.isVector()) {
3561       // Without fast insertion, we need to use MOVD/MOVQ to pass each demanded
3562       // integer element as a SCALAR_TO_VECTOR, then we build the vector as a
3563       // series of UNPCK followed by CONCAT_VECTORS - all of these can be
3564       // considered cheap.
3565       if (Ty->isIntOrIntVectorTy())
3566         Cost += DemandedElts.countPopulation();
3567 
3568       // Get the smaller of the legalized or original pow2-extended number of
3569       // vector elements, which represents the number of unpacks we'll end up
3570       // performing.
3571       unsigned NumElts = LT.second.getVectorNumElements();
3572       unsigned Pow2Elts =
3573           PowerOf2Ceil(cast<FixedVectorType>(Ty)->getNumElements());
3574       Cost += (std::min<unsigned>(NumElts, Pow2Elts) - 1) * LT.first;
3575     }
3576   }
3577 
3578   // TODO: Use default extraction for now, but we should investigate extending this
3579   // to handle repeated subvector extraction.
3580   if (Extract)
3581     Cost += BaseT::getScalarizationOverhead(Ty, DemandedElts, false, Extract);
3582 
3583   return Cost;
3584 }
3585 
3586 InstructionCost X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
3587                                             MaybeAlign Alignment,
3588                                             unsigned AddressSpace,
3589                                             TTI::TargetCostKind CostKind,
3590                                             const Instruction *I) {
3591   // TODO: Handle other cost kinds.
3592   if (CostKind != TTI::TCK_RecipThroughput) {
3593     if (auto *SI = dyn_cast_or_null<StoreInst>(I)) {
3594       // Store instruction with index and scale costs 2 Uops.
3595       // Check the preceding GEP to identify non-const indices.
3596       if (auto *GEP = dyn_cast<GetElementPtrInst>(SI->getPointerOperand())) {
3597         if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); }))
3598           return TTI::TCC_Basic * 2;
3599       }
3600     }
3601     return TTI::TCC_Basic;
3602   }
3603 
3604   assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
3605          "Invalid Opcode");
3606   // Type legalization can't handle structs
3607   if (TLI->getValueType(DL, Src, true) == MVT::Other)
3608     return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
3609                                   CostKind);
3610 
3611   // Legalize the type.
3612   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
3613 
3614   auto *VTy = dyn_cast<FixedVectorType>(Src);
3615 
3616   // Handle the simple case of non-vectors.
3617   // NOTE: this assumes that legalization never creates vector from scalars!
3618   if (!VTy || !LT.second.isVector())
3619     // Each load/store unit costs 1.
3620     return LT.first * 1;
3621 
3622   bool IsLoad = Opcode == Instruction::Load;
3623 
3624   Type *EltTy = VTy->getElementType();
3625 
3626   const int EltTyBits = DL.getTypeSizeInBits(EltTy);
3627 
3628   InstructionCost Cost = 0;
3629 
3630   // Source of truth: how many elements were there in the original IR vector?
3631   const unsigned SrcNumElt = VTy->getNumElements();
3632 
3633   // How far have we gotten?
3634   int NumEltRemaining = SrcNumElt;
3635   // Note that we intentionally capture by-reference, NumEltRemaining changes.
3636   auto NumEltDone = [&]() { return SrcNumElt - NumEltRemaining; };
3637 
3638   const int MaxLegalOpSizeBytes = divideCeil(LT.second.getSizeInBits(), 8);
3639 
3640   // Note that even if we can store 64 bits of an XMM, we still operate on XMM.
3641   const unsigned XMMBits = 128;
3642   if (XMMBits % EltTyBits != 0)
3643     // Vector size must be a multiple of the element size. I.e. no padding.
3644     return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
3645                                   CostKind);
3646   const int NumEltPerXMM = XMMBits / EltTyBits;
3647 
3648   auto *XMMVecTy = FixedVectorType::get(EltTy, NumEltPerXMM);
3649 
3650   for (int CurrOpSizeBytes = MaxLegalOpSizeBytes, SubVecEltsLeft = 0;
3651        NumEltRemaining > 0; CurrOpSizeBytes /= 2) {
3652     // How many elements would a single op deal with at once?
3653     if ((8 * CurrOpSizeBytes) % EltTyBits != 0)
3654       // Vector size must be a multiple of the element size. I.e. no padding.
3655       return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
3656                                     CostKind);
3657     int CurrNumEltPerOp = (8 * CurrOpSizeBytes) / EltTyBits;
3658 
3659     assert(CurrOpSizeBytes > 0 && CurrNumEltPerOp > 0 && "How'd we get here?");
3660     assert((((NumEltRemaining * EltTyBits) < (2 * 8 * CurrOpSizeBytes)) ||
3661             (CurrOpSizeBytes == MaxLegalOpSizeBytes)) &&
3662            "Unless we haven't halved the op size yet, "
3663            "we have less than two op's sized units of work left.");
3664 
3665     auto *CurrVecTy = CurrNumEltPerOp > NumEltPerXMM
3666                           ? FixedVectorType::get(EltTy, CurrNumEltPerOp)
3667                           : XMMVecTy;
3668 
3669     assert(CurrVecTy->getNumElements() % CurrNumEltPerOp == 0 &&
3670            "After halving sizes, the vector elt count is no longer a multiple "
3671            "of number of elements per operation?");
3672     auto *CoalescedVecTy =
3673         CurrNumEltPerOp == 1
3674             ? CurrVecTy
3675             : FixedVectorType::get(
3676                   IntegerType::get(Src->getContext(),
3677                                    EltTyBits * CurrNumEltPerOp),
3678                   CurrVecTy->getNumElements() / CurrNumEltPerOp);
3679     assert(DL.getTypeSizeInBits(CoalescedVecTy) ==
3680                DL.getTypeSizeInBits(CurrVecTy) &&
3681            "coalesciing elements doesn't change vector width.");
3682 
3683     while (NumEltRemaining > 0) {
3684       assert(SubVecEltsLeft >= 0 && "Subreg element count overconsumtion?");
3685 
3686       // Can we use this vector size, as per the remaining element count?
3687       // Iff the vector is naturally aligned, we can do a wide load regardless.
3688       if (NumEltRemaining < CurrNumEltPerOp &&
3689           (!IsLoad || Alignment.valueOrOne() < CurrOpSizeBytes) &&
3690           CurrOpSizeBytes != 1)
3691         break; // Try smalled vector size.
3692 
3693       bool Is0thSubVec = (NumEltDone() % LT.second.getVectorNumElements()) == 0;
3694 
3695       // If we have fully processed the previous reg, we need to replenish it.
3696       if (SubVecEltsLeft == 0) {
3697         SubVecEltsLeft += CurrVecTy->getNumElements();
3698         // And that's free only for the 0'th subvector of a legalized vector.
3699         if (!Is0thSubVec)
3700           Cost += getShuffleCost(IsLoad ? TTI::ShuffleKind::SK_InsertSubvector
3701                                         : TTI::ShuffleKind::SK_ExtractSubvector,
3702                                  VTy, None, NumEltDone(), CurrVecTy);
3703       }
3704 
3705       // While we can directly load/store ZMM, YMM, and 64-bit halves of XMM,
3706       // for smaller widths (32/16/8) we have to insert/extract them separately.
3707       // Again, it's free for the 0'th subreg (if op is 32/64 bit wide,
3708       // but let's pretend that it is also true for 16/8 bit wide ops...)
3709       if (CurrOpSizeBytes <= 32 / 8 && !Is0thSubVec) {
3710         int NumEltDoneInCurrXMM = NumEltDone() % NumEltPerXMM;
3711         assert(NumEltDoneInCurrXMM % CurrNumEltPerOp == 0 && "");
3712         int CoalescedVecEltIdx = NumEltDoneInCurrXMM / CurrNumEltPerOp;
3713         APInt DemandedElts =
3714             APInt::getBitsSet(CoalescedVecTy->getNumElements(),
3715                               CoalescedVecEltIdx, CoalescedVecEltIdx + 1);
3716         assert(DemandedElts.countPopulation() == 1 && "Inserting single value");
3717         Cost += getScalarizationOverhead(CoalescedVecTy, DemandedElts, IsLoad,
3718                                          !IsLoad);
3719       }
3720 
3721       // This isn't exactly right. We're using slow unaligned 32-byte accesses
3722       // as a proxy for a double-pumped AVX memory interface such as on
3723       // Sandybridge.
3724       if (CurrOpSizeBytes == 32 && ST->isUnalignedMem32Slow())
3725         Cost += 2;
3726       else
3727         Cost += 1;
3728 
3729       SubVecEltsLeft -= CurrNumEltPerOp;
3730       NumEltRemaining -= CurrNumEltPerOp;
3731       Alignment = commonAlignment(Alignment.valueOrOne(), CurrOpSizeBytes);
3732     }
3733   }
3734 
3735   assert(NumEltRemaining <= 0 && "Should have processed all the elements.");
3736 
3737   return Cost;
3738 }
3739 
3740 InstructionCost
3741 X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy, Align Alignment,
3742                                   unsigned AddressSpace,
3743                                   TTI::TargetCostKind CostKind) {
3744   bool IsLoad = (Instruction::Load == Opcode);
3745   bool IsStore = (Instruction::Store == Opcode);
3746 
3747   auto *SrcVTy = dyn_cast<FixedVectorType>(SrcTy);
3748   if (!SrcVTy)
3749     // To calculate scalar take the regular cost, without mask
3750     return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace, CostKind);
3751 
3752   unsigned NumElem = SrcVTy->getNumElements();
3753   auto *MaskTy =
3754       FixedVectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem);
3755   if ((IsLoad && !isLegalMaskedLoad(SrcVTy, Alignment)) ||
3756       (IsStore && !isLegalMaskedStore(SrcVTy, Alignment))) {
3757     // Scalarization
3758     APInt DemandedElts = APInt::getAllOnes(NumElem);
3759     InstructionCost MaskSplitCost =
3760         getScalarizationOverhead(MaskTy, DemandedElts, false, true);
3761     InstructionCost ScalarCompareCost = getCmpSelInstrCost(
3762         Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr,
3763         CmpInst::BAD_ICMP_PREDICATE, CostKind);
3764     InstructionCost BranchCost = getCFInstrCost(Instruction::Br, CostKind);
3765     InstructionCost MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost);
3766     InstructionCost ValueSplitCost =
3767         getScalarizationOverhead(SrcVTy, DemandedElts, IsLoad, IsStore);
3768     InstructionCost MemopCost =
3769         NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
3770                                          Alignment, AddressSpace, CostKind);
3771     return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost;
3772   }
3773 
3774   // Legalize the type.
3775   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy);
3776   auto VT = TLI->getValueType(DL, SrcVTy);
3777   InstructionCost Cost = 0;
3778   if (VT.isSimple() && LT.second != VT.getSimpleVT() &&
3779       LT.second.getVectorNumElements() == NumElem)
3780     // Promotion requires extend/truncate for data and a shuffle for mask.
3781     Cost += getShuffleCost(TTI::SK_PermuteTwoSrc, SrcVTy, None, 0, nullptr) +
3782             getShuffleCost(TTI::SK_PermuteTwoSrc, MaskTy, None, 0, nullptr);
3783 
3784   else if (LT.first * LT.second.getVectorNumElements() > NumElem) {
3785     auto *NewMaskTy = FixedVectorType::get(MaskTy->getElementType(),
3786                                            LT.second.getVectorNumElements());
3787     // Expanding requires fill mask with zeroes
3788     Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, None, 0, MaskTy);
3789   }
3790 
3791   // Pre-AVX512 - each maskmov load costs 2 + store costs ~8.
3792   if (!ST->hasAVX512())
3793     return Cost + LT.first * (IsLoad ? 2 : 8);
3794 
3795   // AVX-512 masked load/store is cheapper
3796   return Cost + LT.first;
3797 }
3798 
3799 InstructionCost X86TTIImpl::getAddressComputationCost(Type *Ty,
3800                                                       ScalarEvolution *SE,
3801                                                       const SCEV *Ptr) {
3802   // Address computations in vectorized code with non-consecutive addresses will
3803   // likely result in more instructions compared to scalar code where the
3804   // computation can more often be merged into the index mode. The resulting
3805   // extra micro-ops can significantly decrease throughput.
3806   const unsigned NumVectorInstToHideOverhead = 10;
3807 
3808   // Cost modeling of Strided Access Computation is hidden by the indexing
3809   // modes of X86 regardless of the stride value. We dont believe that there
3810   // is a difference between constant strided access in gerenal and constant
3811   // strided value which is less than or equal to 64.
3812   // Even in the case of (loop invariant) stride whose value is not known at
3813   // compile time, the address computation will not incur more than one extra
3814   // ADD instruction.
3815   if (Ty->isVectorTy() && SE) {
3816     if (!BaseT::isStridedAccess(Ptr))
3817       return NumVectorInstToHideOverhead;
3818     if (!BaseT::getConstantStrideStep(SE, Ptr))
3819       return 1;
3820   }
3821 
3822   return BaseT::getAddressComputationCost(Ty, SE, Ptr);
3823 }
3824 
3825 InstructionCost
3826 X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy,
3827                                        Optional<FastMathFlags> FMF,
3828                                        TTI::TargetCostKind CostKind) {
3829   if (TTI::requiresOrderedReduction(FMF))
3830     return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
3831 
3832   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
3833   // and make it as the cost.
3834 
3835   static const CostTblEntry SLMCostTblNoPairWise[] = {
3836     { ISD::FADD,  MVT::v2f64,   3 },
3837     { ISD::ADD,   MVT::v2i64,   5 },
3838   };
3839 
3840   static const CostTblEntry SSE2CostTblNoPairWise[] = {
3841     { ISD::FADD,  MVT::v2f64,   2 },
3842     { ISD::FADD,  MVT::v2f32,   2 },
3843     { ISD::FADD,  MVT::v4f32,   4 },
3844     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
3845     { ISD::ADD,   MVT::v2i32,   2 }, // FIXME: chosen to be less than v4i32
3846     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.3".
3847     { ISD::ADD,   MVT::v2i16,   2 },      // The data reported by the IACA tool is "4.3".
3848     { ISD::ADD,   MVT::v4i16,   3 },      // The data reported by the IACA tool is "4.3".
3849     { ISD::ADD,   MVT::v8i16,   4 },      // The data reported by the IACA tool is "4.3".
3850     { ISD::ADD,   MVT::v2i8,    2 },
3851     { ISD::ADD,   MVT::v4i8,    2 },
3852     { ISD::ADD,   MVT::v8i8,    2 },
3853     { ISD::ADD,   MVT::v16i8,   3 },
3854   };
3855 
3856   static const CostTblEntry AVX1CostTblNoPairWise[] = {
3857     { ISD::FADD,  MVT::v4f64,   3 },
3858     { ISD::FADD,  MVT::v4f32,   3 },
3859     { ISD::FADD,  MVT::v8f32,   4 },
3860     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
3861     { ISD::ADD,   MVT::v4i64,   3 },
3862     { ISD::ADD,   MVT::v8i32,   5 },
3863     { ISD::ADD,   MVT::v16i16,  5 },
3864     { ISD::ADD,   MVT::v32i8,   4 },
3865   };
3866 
3867   int ISD = TLI->InstructionOpcodeToISD(Opcode);
3868   assert(ISD && "Invalid opcode");
3869 
3870   // Before legalizing the type, give a chance to look up illegal narrow types
3871   // in the table.
3872   // FIXME: Is there a better way to do this?
3873   EVT VT = TLI->getValueType(DL, ValTy);
3874   if (VT.isSimple()) {
3875     MVT MTy = VT.getSimpleVT();
3876     if (ST->isSLM())
3877       if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy))
3878         return Entry->Cost;
3879 
3880     if (ST->hasAVX())
3881       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
3882         return Entry->Cost;
3883 
3884     if (ST->hasSSE2())
3885       if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
3886         return Entry->Cost;
3887   }
3888 
3889   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
3890 
3891   MVT MTy = LT.second;
3892 
3893   auto *ValVTy = cast<FixedVectorType>(ValTy);
3894 
3895   // Special case: vXi8 mul reductions are performed as vXi16.
3896   if (ISD == ISD::MUL && MTy.getScalarType() == MVT::i8) {
3897     auto *WideSclTy = IntegerType::get(ValVTy->getContext(), 16);
3898     auto *WideVecTy = FixedVectorType::get(WideSclTy, ValVTy->getNumElements());
3899     return getCastInstrCost(Instruction::ZExt, WideVecTy, ValTy,
3900                             TargetTransformInfo::CastContextHint::None,
3901                             CostKind) +
3902            getArithmeticReductionCost(Opcode, WideVecTy, FMF, CostKind);
3903   }
3904 
3905   InstructionCost ArithmeticCost = 0;
3906   if (LT.first != 1 && MTy.isVector() &&
3907       MTy.getVectorNumElements() < ValVTy->getNumElements()) {
3908     // Type needs to be split. We need LT.first - 1 arithmetic ops.
3909     auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(),
3910                                             MTy.getVectorNumElements());
3911     ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind);
3912     ArithmeticCost *= LT.first - 1;
3913   }
3914 
3915   if (ST->isSLM())
3916     if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy))
3917       return ArithmeticCost + Entry->Cost;
3918 
3919   if (ST->hasAVX())
3920     if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
3921       return ArithmeticCost + Entry->Cost;
3922 
3923   if (ST->hasSSE2())
3924     if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
3925       return ArithmeticCost + Entry->Cost;
3926 
3927   // FIXME: These assume a naive kshift+binop lowering, which is probably
3928   // conservative in most cases.
3929   static const CostTblEntry AVX512BoolReduction[] = {
3930     { ISD::AND,  MVT::v2i1,   3 },
3931     { ISD::AND,  MVT::v4i1,   5 },
3932     { ISD::AND,  MVT::v8i1,   7 },
3933     { ISD::AND,  MVT::v16i1,  9 },
3934     { ISD::AND,  MVT::v32i1, 11 },
3935     { ISD::AND,  MVT::v64i1, 13 },
3936     { ISD::OR,   MVT::v2i1,   3 },
3937     { ISD::OR,   MVT::v4i1,   5 },
3938     { ISD::OR,   MVT::v8i1,   7 },
3939     { ISD::OR,   MVT::v16i1,  9 },
3940     { ISD::OR,   MVT::v32i1, 11 },
3941     { ISD::OR,   MVT::v64i1, 13 },
3942   };
3943 
3944   static const CostTblEntry AVX2BoolReduction[] = {
3945     { ISD::AND,  MVT::v16i16,  2 }, // vpmovmskb + cmp
3946     { ISD::AND,  MVT::v32i8,   2 }, // vpmovmskb + cmp
3947     { ISD::OR,   MVT::v16i16,  2 }, // vpmovmskb + cmp
3948     { ISD::OR,   MVT::v32i8,   2 }, // vpmovmskb + cmp
3949   };
3950 
3951   static const CostTblEntry AVX1BoolReduction[] = {
3952     { ISD::AND,  MVT::v4i64,   2 }, // vmovmskpd + cmp
3953     { ISD::AND,  MVT::v8i32,   2 }, // vmovmskps + cmp
3954     { ISD::AND,  MVT::v16i16,  4 }, // vextractf128 + vpand + vpmovmskb + cmp
3955     { ISD::AND,  MVT::v32i8,   4 }, // vextractf128 + vpand + vpmovmskb + cmp
3956     { ISD::OR,   MVT::v4i64,   2 }, // vmovmskpd + cmp
3957     { ISD::OR,   MVT::v8i32,   2 }, // vmovmskps + cmp
3958     { ISD::OR,   MVT::v16i16,  4 }, // vextractf128 + vpor + vpmovmskb + cmp
3959     { ISD::OR,   MVT::v32i8,   4 }, // vextractf128 + vpor + vpmovmskb + cmp
3960   };
3961 
3962   static const CostTblEntry SSE2BoolReduction[] = {
3963     { ISD::AND,  MVT::v2i64,   2 }, // movmskpd + cmp
3964     { ISD::AND,  MVT::v4i32,   2 }, // movmskps + cmp
3965     { ISD::AND,  MVT::v8i16,   2 }, // pmovmskb + cmp
3966     { ISD::AND,  MVT::v16i8,   2 }, // pmovmskb + cmp
3967     { ISD::OR,   MVT::v2i64,   2 }, // movmskpd + cmp
3968     { ISD::OR,   MVT::v4i32,   2 }, // movmskps + cmp
3969     { ISD::OR,   MVT::v8i16,   2 }, // pmovmskb + cmp
3970     { ISD::OR,   MVT::v16i8,   2 }, // pmovmskb + cmp
3971   };
3972 
3973   // Handle bool allof/anyof patterns.
3974   if (ValVTy->getElementType()->isIntegerTy(1)) {
3975     InstructionCost ArithmeticCost = 0;
3976     if (LT.first != 1 && MTy.isVector() &&
3977         MTy.getVectorNumElements() < ValVTy->getNumElements()) {
3978       // Type needs to be split. We need LT.first - 1 arithmetic ops.
3979       auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(),
3980                                               MTy.getVectorNumElements());
3981       ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind);
3982       ArithmeticCost *= LT.first - 1;
3983     }
3984 
3985     if (ST->hasAVX512())
3986       if (const auto *Entry = CostTableLookup(AVX512BoolReduction, ISD, MTy))
3987         return ArithmeticCost + Entry->Cost;
3988     if (ST->hasAVX2())
3989       if (const auto *Entry = CostTableLookup(AVX2BoolReduction, ISD, MTy))
3990         return ArithmeticCost + Entry->Cost;
3991     if (ST->hasAVX())
3992       if (const auto *Entry = CostTableLookup(AVX1BoolReduction, ISD, MTy))
3993         return ArithmeticCost + Entry->Cost;
3994     if (ST->hasSSE2())
3995       if (const auto *Entry = CostTableLookup(SSE2BoolReduction, ISD, MTy))
3996         return ArithmeticCost + Entry->Cost;
3997 
3998     return BaseT::getArithmeticReductionCost(Opcode, ValVTy, FMF, CostKind);
3999   }
4000 
4001   unsigned NumVecElts = ValVTy->getNumElements();
4002   unsigned ScalarSize = ValVTy->getScalarSizeInBits();
4003 
4004   // Special case power of 2 reductions where the scalar type isn't changed
4005   // by type legalization.
4006   if (!isPowerOf2_32(NumVecElts) || ScalarSize != MTy.getScalarSizeInBits())
4007     return BaseT::getArithmeticReductionCost(Opcode, ValVTy, FMF, CostKind);
4008 
4009   InstructionCost ReductionCost = 0;
4010 
4011   auto *Ty = ValVTy;
4012   if (LT.first != 1 && MTy.isVector() &&
4013       MTy.getVectorNumElements() < ValVTy->getNumElements()) {
4014     // Type needs to be split. We need LT.first - 1 arithmetic ops.
4015     Ty = FixedVectorType::get(ValVTy->getElementType(),
4016                               MTy.getVectorNumElements());
4017     ReductionCost = getArithmeticInstrCost(Opcode, Ty, CostKind);
4018     ReductionCost *= LT.first - 1;
4019     NumVecElts = MTy.getVectorNumElements();
4020   }
4021 
4022   // Now handle reduction with the legal type, taking into account size changes
4023   // at each level.
4024   while (NumVecElts > 1) {
4025     // Determine the size of the remaining vector we need to reduce.
4026     unsigned Size = NumVecElts * ScalarSize;
4027     NumVecElts /= 2;
4028     // If we're reducing from 256/512 bits, use an extract_subvector.
4029     if (Size > 128) {
4030       auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts);
4031       ReductionCost +=
4032           getShuffleCost(TTI::SK_ExtractSubvector, Ty, None, NumVecElts, SubTy);
4033       Ty = SubTy;
4034     } else if (Size == 128) {
4035       // Reducing from 128 bits is a permute of v2f64/v2i64.
4036       FixedVectorType *ShufTy;
4037       if (ValVTy->isFloatingPointTy())
4038         ShufTy =
4039             FixedVectorType::get(Type::getDoubleTy(ValVTy->getContext()), 2);
4040       else
4041         ShufTy =
4042             FixedVectorType::get(Type::getInt64Ty(ValVTy->getContext()), 2);
4043       ReductionCost +=
4044           getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4045     } else if (Size == 64) {
4046       // Reducing from 64 bits is a shuffle of v4f32/v4i32.
4047       FixedVectorType *ShufTy;
4048       if (ValVTy->isFloatingPointTy())
4049         ShufTy =
4050             FixedVectorType::get(Type::getFloatTy(ValVTy->getContext()), 4);
4051       else
4052         ShufTy =
4053             FixedVectorType::get(Type::getInt32Ty(ValVTy->getContext()), 4);
4054       ReductionCost +=
4055           getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4056     } else {
4057       // Reducing from smaller size is a shift by immediate.
4058       auto *ShiftTy = FixedVectorType::get(
4059           Type::getIntNTy(ValVTy->getContext(), Size), 128 / Size);
4060       ReductionCost += getArithmeticInstrCost(
4061           Instruction::LShr, ShiftTy, CostKind,
4062           TargetTransformInfo::OK_AnyValue,
4063           TargetTransformInfo::OK_UniformConstantValue,
4064           TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
4065     }
4066 
4067     // Add the arithmetic op for this level.
4068     ReductionCost += getArithmeticInstrCost(Opcode, Ty, CostKind);
4069   }
4070 
4071   // Add the final extract element to the cost.
4072   return ReductionCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
4073 }
4074 
4075 InstructionCost X86TTIImpl::getMinMaxCost(Type *Ty, Type *CondTy,
4076                                           bool IsUnsigned) {
4077   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
4078 
4079   MVT MTy = LT.second;
4080 
4081   int ISD;
4082   if (Ty->isIntOrIntVectorTy()) {
4083     ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN;
4084   } else {
4085     assert(Ty->isFPOrFPVectorTy() &&
4086            "Expected float point or integer vector type.");
4087     ISD = ISD::FMINNUM;
4088   }
4089 
4090   static const CostTblEntry SSE1CostTbl[] = {
4091     {ISD::FMINNUM, MVT::v4f32, 1},
4092   };
4093 
4094   static const CostTblEntry SSE2CostTbl[] = {
4095     {ISD::FMINNUM, MVT::v2f64, 1},
4096     {ISD::SMIN,    MVT::v8i16, 1},
4097     {ISD::UMIN,    MVT::v16i8, 1},
4098   };
4099 
4100   static const CostTblEntry SSE41CostTbl[] = {
4101     {ISD::SMIN,    MVT::v4i32, 1},
4102     {ISD::UMIN,    MVT::v4i32, 1},
4103     {ISD::UMIN,    MVT::v8i16, 1},
4104     {ISD::SMIN,    MVT::v16i8, 1},
4105   };
4106 
4107   static const CostTblEntry SSE42CostTbl[] = {
4108     {ISD::UMIN,    MVT::v2i64, 3}, // xor+pcmpgtq+blendvpd
4109   };
4110 
4111   static const CostTblEntry AVX1CostTbl[] = {
4112     {ISD::FMINNUM, MVT::v8f32,  1},
4113     {ISD::FMINNUM, MVT::v4f64,  1},
4114     {ISD::SMIN,    MVT::v8i32,  3},
4115     {ISD::UMIN,    MVT::v8i32,  3},
4116     {ISD::SMIN,    MVT::v16i16, 3},
4117     {ISD::UMIN,    MVT::v16i16, 3},
4118     {ISD::SMIN,    MVT::v32i8,  3},
4119     {ISD::UMIN,    MVT::v32i8,  3},
4120   };
4121 
4122   static const CostTblEntry AVX2CostTbl[] = {
4123     {ISD::SMIN,    MVT::v8i32,  1},
4124     {ISD::UMIN,    MVT::v8i32,  1},
4125     {ISD::SMIN,    MVT::v16i16, 1},
4126     {ISD::UMIN,    MVT::v16i16, 1},
4127     {ISD::SMIN,    MVT::v32i8,  1},
4128     {ISD::UMIN,    MVT::v32i8,  1},
4129   };
4130 
4131   static const CostTblEntry AVX512CostTbl[] = {
4132     {ISD::FMINNUM, MVT::v16f32, 1},
4133     {ISD::FMINNUM, MVT::v8f64,  1},
4134     {ISD::SMIN,    MVT::v2i64,  1},
4135     {ISD::UMIN,    MVT::v2i64,  1},
4136     {ISD::SMIN,    MVT::v4i64,  1},
4137     {ISD::UMIN,    MVT::v4i64,  1},
4138     {ISD::SMIN,    MVT::v8i64,  1},
4139     {ISD::UMIN,    MVT::v8i64,  1},
4140     {ISD::SMIN,    MVT::v16i32, 1},
4141     {ISD::UMIN,    MVT::v16i32, 1},
4142   };
4143 
4144   static const CostTblEntry AVX512BWCostTbl[] = {
4145     {ISD::SMIN,    MVT::v32i16, 1},
4146     {ISD::UMIN,    MVT::v32i16, 1},
4147     {ISD::SMIN,    MVT::v64i8,  1},
4148     {ISD::UMIN,    MVT::v64i8,  1},
4149   };
4150 
4151   // If we have a native MIN/MAX instruction for this type, use it.
4152   if (ST->hasBWI())
4153     if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
4154       return LT.first * Entry->Cost;
4155 
4156   if (ST->hasAVX512())
4157     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
4158       return LT.first * Entry->Cost;
4159 
4160   if (ST->hasAVX2())
4161     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
4162       return LT.first * Entry->Cost;
4163 
4164   if (ST->hasAVX())
4165     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
4166       return LT.first * Entry->Cost;
4167 
4168   if (ST->hasSSE42())
4169     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
4170       return LT.first * Entry->Cost;
4171 
4172   if (ST->hasSSE41())
4173     if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
4174       return LT.first * Entry->Cost;
4175 
4176   if (ST->hasSSE2())
4177     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
4178       return LT.first * Entry->Cost;
4179 
4180   if (ST->hasSSE1())
4181     if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
4182       return LT.first * Entry->Cost;
4183 
4184   unsigned CmpOpcode;
4185   if (Ty->isFPOrFPVectorTy()) {
4186     CmpOpcode = Instruction::FCmp;
4187   } else {
4188     assert(Ty->isIntOrIntVectorTy() &&
4189            "expecting floating point or integer type for min/max reduction");
4190     CmpOpcode = Instruction::ICmp;
4191   }
4192 
4193   TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
4194   // Otherwise fall back to cmp+select.
4195   InstructionCost Result =
4196       getCmpSelInstrCost(CmpOpcode, Ty, CondTy, CmpInst::BAD_ICMP_PREDICATE,
4197                          CostKind) +
4198       getCmpSelInstrCost(Instruction::Select, Ty, CondTy,
4199                          CmpInst::BAD_ICMP_PREDICATE, CostKind);
4200   return Result;
4201 }
4202 
4203 InstructionCost
4204 X86TTIImpl::getMinMaxReductionCost(VectorType *ValTy, VectorType *CondTy,
4205                                    bool IsUnsigned,
4206                                    TTI::TargetCostKind CostKind) {
4207   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
4208 
4209   MVT MTy = LT.second;
4210 
4211   int ISD;
4212   if (ValTy->isIntOrIntVectorTy()) {
4213     ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN;
4214   } else {
4215     assert(ValTy->isFPOrFPVectorTy() &&
4216            "Expected float point or integer vector type.");
4217     ISD = ISD::FMINNUM;
4218   }
4219 
4220   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
4221   // and make it as the cost.
4222 
4223   static const CostTblEntry SSE2CostTblNoPairWise[] = {
4224       {ISD::UMIN, MVT::v2i16, 5}, // need pxors to use pminsw/pmaxsw
4225       {ISD::UMIN, MVT::v4i16, 7}, // need pxors to use pminsw/pmaxsw
4226       {ISD::UMIN, MVT::v8i16, 9}, // need pxors to use pminsw/pmaxsw
4227   };
4228 
4229   static const CostTblEntry SSE41CostTblNoPairWise[] = {
4230       {ISD::SMIN, MVT::v2i16, 3}, // same as sse2
4231       {ISD::SMIN, MVT::v4i16, 5}, // same as sse2
4232       {ISD::UMIN, MVT::v2i16, 5}, // same as sse2
4233       {ISD::UMIN, MVT::v4i16, 7}, // same as sse2
4234       {ISD::SMIN, MVT::v8i16, 4}, // phminposuw+xor
4235       {ISD::UMIN, MVT::v8i16, 4}, // FIXME: umin is cheaper than umax
4236       {ISD::SMIN, MVT::v2i8,  3}, // pminsb
4237       {ISD::SMIN, MVT::v4i8,  5}, // pminsb
4238       {ISD::SMIN, MVT::v8i8,  7}, // pminsb
4239       {ISD::SMIN, MVT::v16i8, 6},
4240       {ISD::UMIN, MVT::v2i8,  3}, // same as sse2
4241       {ISD::UMIN, MVT::v4i8,  5}, // same as sse2
4242       {ISD::UMIN, MVT::v8i8,  7}, // same as sse2
4243       {ISD::UMIN, MVT::v16i8, 6}, // FIXME: umin is cheaper than umax
4244   };
4245 
4246   static const CostTblEntry AVX1CostTblNoPairWise[] = {
4247       {ISD::SMIN, MVT::v16i16, 6},
4248       {ISD::UMIN, MVT::v16i16, 6}, // FIXME: umin is cheaper than umax
4249       {ISD::SMIN, MVT::v32i8, 8},
4250       {ISD::UMIN, MVT::v32i8, 8},
4251   };
4252 
4253   static const CostTblEntry AVX512BWCostTblNoPairWise[] = {
4254       {ISD::SMIN, MVT::v32i16, 8},
4255       {ISD::UMIN, MVT::v32i16, 8}, // FIXME: umin is cheaper than umax
4256       {ISD::SMIN, MVT::v64i8, 10},
4257       {ISD::UMIN, MVT::v64i8, 10},
4258   };
4259 
4260   // Before legalizing the type, give a chance to look up illegal narrow types
4261   // in the table.
4262   // FIXME: Is there a better way to do this?
4263   EVT VT = TLI->getValueType(DL, ValTy);
4264   if (VT.isSimple()) {
4265     MVT MTy = VT.getSimpleVT();
4266     if (ST->hasBWI())
4267       if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy))
4268         return Entry->Cost;
4269 
4270     if (ST->hasAVX())
4271       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
4272         return Entry->Cost;
4273 
4274     if (ST->hasSSE41())
4275       if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy))
4276         return Entry->Cost;
4277 
4278     if (ST->hasSSE2())
4279       if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
4280         return Entry->Cost;
4281   }
4282 
4283   auto *ValVTy = cast<FixedVectorType>(ValTy);
4284   unsigned NumVecElts = ValVTy->getNumElements();
4285 
4286   auto *Ty = ValVTy;
4287   InstructionCost MinMaxCost = 0;
4288   if (LT.first != 1 && MTy.isVector() &&
4289       MTy.getVectorNumElements() < ValVTy->getNumElements()) {
4290     // Type needs to be split. We need LT.first - 1 operations ops.
4291     Ty = FixedVectorType::get(ValVTy->getElementType(),
4292                               MTy.getVectorNumElements());
4293     auto *SubCondTy = FixedVectorType::get(CondTy->getElementType(),
4294                                            MTy.getVectorNumElements());
4295     MinMaxCost = getMinMaxCost(Ty, SubCondTy, IsUnsigned);
4296     MinMaxCost *= LT.first - 1;
4297     NumVecElts = MTy.getVectorNumElements();
4298   }
4299 
4300   if (ST->hasBWI())
4301     if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy))
4302       return MinMaxCost + Entry->Cost;
4303 
4304   if (ST->hasAVX())
4305     if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
4306       return MinMaxCost + Entry->Cost;
4307 
4308   if (ST->hasSSE41())
4309     if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy))
4310       return MinMaxCost + Entry->Cost;
4311 
4312   if (ST->hasSSE2())
4313     if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
4314       return MinMaxCost + Entry->Cost;
4315 
4316   unsigned ScalarSize = ValTy->getScalarSizeInBits();
4317 
4318   // Special case power of 2 reductions where the scalar type isn't changed
4319   // by type legalization.
4320   if (!isPowerOf2_32(ValVTy->getNumElements()) ||
4321       ScalarSize != MTy.getScalarSizeInBits())
4322     return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsUnsigned, CostKind);
4323 
4324   // Now handle reduction with the legal type, taking into account size changes
4325   // at each level.
4326   while (NumVecElts > 1) {
4327     // Determine the size of the remaining vector we need to reduce.
4328     unsigned Size = NumVecElts * ScalarSize;
4329     NumVecElts /= 2;
4330     // If we're reducing from 256/512 bits, use an extract_subvector.
4331     if (Size > 128) {
4332       auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts);
4333       MinMaxCost +=
4334           getShuffleCost(TTI::SK_ExtractSubvector, Ty, None, NumVecElts, SubTy);
4335       Ty = SubTy;
4336     } else if (Size == 128) {
4337       // Reducing from 128 bits is a permute of v2f64/v2i64.
4338       VectorType *ShufTy;
4339       if (ValTy->isFloatingPointTy())
4340         ShufTy =
4341             FixedVectorType::get(Type::getDoubleTy(ValTy->getContext()), 2);
4342       else
4343         ShufTy = FixedVectorType::get(Type::getInt64Ty(ValTy->getContext()), 2);
4344       MinMaxCost +=
4345           getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4346     } else if (Size == 64) {
4347       // Reducing from 64 bits is a shuffle of v4f32/v4i32.
4348       FixedVectorType *ShufTy;
4349       if (ValTy->isFloatingPointTy())
4350         ShufTy = FixedVectorType::get(Type::getFloatTy(ValTy->getContext()), 4);
4351       else
4352         ShufTy = FixedVectorType::get(Type::getInt32Ty(ValTy->getContext()), 4);
4353       MinMaxCost +=
4354           getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4355     } else {
4356       // Reducing from smaller size is a shift by immediate.
4357       auto *ShiftTy = FixedVectorType::get(
4358           Type::getIntNTy(ValTy->getContext(), Size), 128 / Size);
4359       MinMaxCost += getArithmeticInstrCost(
4360           Instruction::LShr, ShiftTy, TTI::TCK_RecipThroughput,
4361           TargetTransformInfo::OK_AnyValue,
4362           TargetTransformInfo::OK_UniformConstantValue,
4363           TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
4364     }
4365 
4366     // Add the arithmetic op for this level.
4367     auto *SubCondTy =
4368         FixedVectorType::get(CondTy->getElementType(), Ty->getNumElements());
4369     MinMaxCost += getMinMaxCost(Ty, SubCondTy, IsUnsigned);
4370   }
4371 
4372   // Add the final extract element to the cost.
4373   return MinMaxCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
4374 }
4375 
4376 /// Calculate the cost of materializing a 64-bit value. This helper
4377 /// method might only calculate a fraction of a larger immediate. Therefore it
4378 /// is valid to return a cost of ZERO.
4379 InstructionCost X86TTIImpl::getIntImmCost(int64_t Val) {
4380   if (Val == 0)
4381     return TTI::TCC_Free;
4382 
4383   if (isInt<32>(Val))
4384     return TTI::TCC_Basic;
4385 
4386   return 2 * TTI::TCC_Basic;
4387 }
4388 
4389 InstructionCost X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
4390                                           TTI::TargetCostKind CostKind) {
4391   assert(Ty->isIntegerTy());
4392 
4393   unsigned BitSize = Ty->getPrimitiveSizeInBits();
4394   if (BitSize == 0)
4395     return ~0U;
4396 
4397   // Never hoist constants larger than 128bit, because this might lead to
4398   // incorrect code generation or assertions in codegen.
4399   // Fixme: Create a cost model for types larger than i128 once the codegen
4400   // issues have been fixed.
4401   if (BitSize > 128)
4402     return TTI::TCC_Free;
4403 
4404   if (Imm == 0)
4405     return TTI::TCC_Free;
4406 
4407   // Sign-extend all constants to a multiple of 64-bit.
4408   APInt ImmVal = Imm;
4409   if (BitSize % 64 != 0)
4410     ImmVal = Imm.sext(alignTo(BitSize, 64));
4411 
4412   // Split the constant into 64-bit chunks and calculate the cost for each
4413   // chunk.
4414   InstructionCost Cost = 0;
4415   for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
4416     APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
4417     int64_t Val = Tmp.getSExtValue();
4418     Cost += getIntImmCost(Val);
4419   }
4420   // We need at least one instruction to materialize the constant.
4421   return std::max<InstructionCost>(1, Cost);
4422 }
4423 
4424 InstructionCost X86TTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
4425                                               const APInt &Imm, Type *Ty,
4426                                               TTI::TargetCostKind CostKind,
4427                                               Instruction *Inst) {
4428   assert(Ty->isIntegerTy());
4429 
4430   unsigned BitSize = Ty->getPrimitiveSizeInBits();
4431   // There is no cost model for constants with a bit size of 0. Return TCC_Free
4432   // here, so that constant hoisting will ignore this constant.
4433   if (BitSize == 0)
4434     return TTI::TCC_Free;
4435 
4436   unsigned ImmIdx = ~0U;
4437   switch (Opcode) {
4438   default:
4439     return TTI::TCC_Free;
4440   case Instruction::GetElementPtr:
4441     // Always hoist the base address of a GetElementPtr. This prevents the
4442     // creation of new constants for every base constant that gets constant
4443     // folded with the offset.
4444     if (Idx == 0)
4445       return 2 * TTI::TCC_Basic;
4446     return TTI::TCC_Free;
4447   case Instruction::Store:
4448     ImmIdx = 0;
4449     break;
4450   case Instruction::ICmp:
4451     // This is an imperfect hack to prevent constant hoisting of
4452     // compares that might be trying to check if a 64-bit value fits in
4453     // 32-bits. The backend can optimize these cases using a right shift by 32.
4454     // Ideally we would check the compare predicate here. There also other
4455     // similar immediates the backend can use shifts for.
4456     if (Idx == 1 && Imm.getBitWidth() == 64) {
4457       uint64_t ImmVal = Imm.getZExtValue();
4458       if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff)
4459         return TTI::TCC_Free;
4460     }
4461     ImmIdx = 1;
4462     break;
4463   case Instruction::And:
4464     // We support 64-bit ANDs with immediates with 32-bits of leading zeroes
4465     // by using a 32-bit operation with implicit zero extension. Detect such
4466     // immediates here as the normal path expects bit 31 to be sign extended.
4467     if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue()))
4468       return TTI::TCC_Free;
4469     ImmIdx = 1;
4470     break;
4471   case Instruction::Add:
4472   case Instruction::Sub:
4473     // For add/sub, we can use the opposite instruction for INT32_MIN.
4474     if (Idx == 1 && Imm.getBitWidth() == 64 && Imm.getZExtValue() == 0x80000000)
4475       return TTI::TCC_Free;
4476     ImmIdx = 1;
4477     break;
4478   case Instruction::UDiv:
4479   case Instruction::SDiv:
4480   case Instruction::URem:
4481   case Instruction::SRem:
4482     // Division by constant is typically expanded later into a different
4483     // instruction sequence. This completely changes the constants.
4484     // Report them as "free" to stop ConstantHoist from marking them as opaque.
4485     return TTI::TCC_Free;
4486   case Instruction::Mul:
4487   case Instruction::Or:
4488   case Instruction::Xor:
4489     ImmIdx = 1;
4490     break;
4491   // Always return TCC_Free for the shift value of a shift instruction.
4492   case Instruction::Shl:
4493   case Instruction::LShr:
4494   case Instruction::AShr:
4495     if (Idx == 1)
4496       return TTI::TCC_Free;
4497     break;
4498   case Instruction::Trunc:
4499   case Instruction::ZExt:
4500   case Instruction::SExt:
4501   case Instruction::IntToPtr:
4502   case Instruction::PtrToInt:
4503   case Instruction::BitCast:
4504   case Instruction::PHI:
4505   case Instruction::Call:
4506   case Instruction::Select:
4507   case Instruction::Ret:
4508   case Instruction::Load:
4509     break;
4510   }
4511 
4512   if (Idx == ImmIdx) {
4513     int NumConstants = divideCeil(BitSize, 64);
4514     InstructionCost Cost = X86TTIImpl::getIntImmCost(Imm, Ty, CostKind);
4515     return (Cost <= NumConstants * TTI::TCC_Basic)
4516                ? static_cast<int>(TTI::TCC_Free)
4517                : Cost;
4518   }
4519 
4520   return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind);
4521 }
4522 
4523 InstructionCost X86TTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
4524                                                 const APInt &Imm, Type *Ty,
4525                                                 TTI::TargetCostKind CostKind) {
4526   assert(Ty->isIntegerTy());
4527 
4528   unsigned BitSize = Ty->getPrimitiveSizeInBits();
4529   // There is no cost model for constants with a bit size of 0. Return TCC_Free
4530   // here, so that constant hoisting will ignore this constant.
4531   if (BitSize == 0)
4532     return TTI::TCC_Free;
4533 
4534   switch (IID) {
4535   default:
4536     return TTI::TCC_Free;
4537   case Intrinsic::sadd_with_overflow:
4538   case Intrinsic::uadd_with_overflow:
4539   case Intrinsic::ssub_with_overflow:
4540   case Intrinsic::usub_with_overflow:
4541   case Intrinsic::smul_with_overflow:
4542   case Intrinsic::umul_with_overflow:
4543     if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
4544       return TTI::TCC_Free;
4545     break;
4546   case Intrinsic::experimental_stackmap:
4547     if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
4548       return TTI::TCC_Free;
4549     break;
4550   case Intrinsic::experimental_patchpoint_void:
4551   case Intrinsic::experimental_patchpoint_i64:
4552     if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
4553       return TTI::TCC_Free;
4554     break;
4555   }
4556   return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind);
4557 }
4558 
4559 InstructionCost X86TTIImpl::getCFInstrCost(unsigned Opcode,
4560                                            TTI::TargetCostKind CostKind,
4561                                            const Instruction *I) {
4562   if (CostKind != TTI::TCK_RecipThroughput)
4563     return Opcode == Instruction::PHI ? 0 : 1;
4564   // Branches are assumed to be predicted.
4565   return 0;
4566 }
4567 
4568 int X86TTIImpl::getGatherOverhead() const {
4569   // Some CPUs have more overhead for gather. The specified overhead is relative
4570   // to the Load operation. "2" is the number provided by Intel architects. This
4571   // parameter is used for cost estimation of Gather Op and comparison with
4572   // other alternatives.
4573   // TODO: Remove the explicit hasAVX512()?, That would mean we would only
4574   // enable gather with a -march.
4575   if (ST->hasAVX512() || (ST->hasAVX2() && ST->hasFastGather()))
4576     return 2;
4577 
4578   return 1024;
4579 }
4580 
4581 int X86TTIImpl::getScatterOverhead() const {
4582   if (ST->hasAVX512())
4583     return 2;
4584 
4585   return 1024;
4586 }
4587 
4588 // Return an average cost of Gather / Scatter instruction, maybe improved later.
4589 // FIXME: Add TargetCostKind support.
4590 InstructionCost X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy,
4591                                             const Value *Ptr, Align Alignment,
4592                                             unsigned AddressSpace) {
4593 
4594   assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost");
4595   unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements();
4596 
4597   // Try to reduce index size from 64 bit (default for GEP)
4598   // to 32. It is essential for VF 16. If the index can't be reduced to 32, the
4599   // operation will use 16 x 64 indices which do not fit in a zmm and needs
4600   // to split. Also check that the base pointer is the same for all lanes,
4601   // and that there's at most one variable index.
4602   auto getIndexSizeInBits = [](const Value *Ptr, const DataLayout &DL) {
4603     unsigned IndexSize = DL.getPointerSizeInBits();
4604     const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
4605     if (IndexSize < 64 || !GEP)
4606       return IndexSize;
4607 
4608     unsigned NumOfVarIndices = 0;
4609     const Value *Ptrs = GEP->getPointerOperand();
4610     if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs))
4611       return IndexSize;
4612     for (unsigned i = 1; i < GEP->getNumOperands(); ++i) {
4613       if (isa<Constant>(GEP->getOperand(i)))
4614         continue;
4615       Type *IndxTy = GEP->getOperand(i)->getType();
4616       if (auto *IndexVTy = dyn_cast<VectorType>(IndxTy))
4617         IndxTy = IndexVTy->getElementType();
4618       if ((IndxTy->getPrimitiveSizeInBits() == 64 &&
4619           !isa<SExtInst>(GEP->getOperand(i))) ||
4620          ++NumOfVarIndices > 1)
4621         return IndexSize; // 64
4622     }
4623     return (unsigned)32;
4624   };
4625 
4626   // Trying to reduce IndexSize to 32 bits for vector 16.
4627   // By default the IndexSize is equal to pointer size.
4628   unsigned IndexSize = (ST->hasAVX512() && VF >= 16)
4629                            ? getIndexSizeInBits(Ptr, DL)
4630                            : DL.getPointerSizeInBits();
4631 
4632   auto *IndexVTy = FixedVectorType::get(
4633       IntegerType::get(SrcVTy->getContext(), IndexSize), VF);
4634   std::pair<InstructionCost, MVT> IdxsLT =
4635       TLI->getTypeLegalizationCost(DL, IndexVTy);
4636   std::pair<InstructionCost, MVT> SrcLT =
4637       TLI->getTypeLegalizationCost(DL, SrcVTy);
4638   InstructionCost::CostType SplitFactor =
4639       *std::max(IdxsLT.first, SrcLT.first).getValue();
4640   if (SplitFactor > 1) {
4641     // Handle splitting of vector of pointers
4642     auto *SplitSrcTy =
4643         FixedVectorType::get(SrcVTy->getScalarType(), VF / SplitFactor);
4644     return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment,
4645                                          AddressSpace);
4646   }
4647 
4648   // The gather / scatter cost is given by Intel architects. It is a rough
4649   // number since we are looking at one instruction in a time.
4650   const int GSOverhead = (Opcode == Instruction::Load)
4651                              ? getGatherOverhead()
4652                              : getScatterOverhead();
4653   return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
4654                                            MaybeAlign(Alignment), AddressSpace,
4655                                            TTI::TCK_RecipThroughput);
4656 }
4657 
4658 /// Return the cost of full scalarization of gather / scatter operation.
4659 ///
4660 /// Opcode - Load or Store instruction.
4661 /// SrcVTy - The type of the data vector that should be gathered or scattered.
4662 /// VariableMask - The mask is non-constant at compile time.
4663 /// Alignment - Alignment for one element.
4664 /// AddressSpace - pointer[s] address space.
4665 ///
4666 /// FIXME: Add TargetCostKind support.
4667 InstructionCost X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy,
4668                                             bool VariableMask, Align Alignment,
4669                                             unsigned AddressSpace) {
4670   unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements();
4671   APInt DemandedElts = APInt::getAllOnes(VF);
4672   TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
4673 
4674   InstructionCost MaskUnpackCost = 0;
4675   if (VariableMask) {
4676     auto *MaskTy =
4677         FixedVectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF);
4678     MaskUnpackCost =
4679         getScalarizationOverhead(MaskTy, DemandedElts, false, true);
4680     InstructionCost ScalarCompareCost = getCmpSelInstrCost(
4681         Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()), nullptr,
4682         CmpInst::BAD_ICMP_PREDICATE, CostKind);
4683     InstructionCost BranchCost = getCFInstrCost(Instruction::Br, CostKind);
4684     MaskUnpackCost += VF * (BranchCost + ScalarCompareCost);
4685   }
4686 
4687   // The cost of the scalar loads/stores.
4688   InstructionCost MemoryOpCost =
4689       VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
4690                            MaybeAlign(Alignment), AddressSpace, CostKind);
4691 
4692   InstructionCost InsertExtractCost = 0;
4693   if (Opcode == Instruction::Load)
4694     for (unsigned i = 0; i < VF; ++i)
4695       // Add the cost of inserting each scalar load into the vector
4696       InsertExtractCost +=
4697         getVectorInstrCost(Instruction::InsertElement, SrcVTy, i);
4698   else
4699     for (unsigned i = 0; i < VF; ++i)
4700       // Add the cost of extracting each element out of the data vector
4701       InsertExtractCost +=
4702         getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i);
4703 
4704   return MemoryOpCost + MaskUnpackCost + InsertExtractCost;
4705 }
4706 
4707 /// Calculate the cost of Gather / Scatter operation
4708 InstructionCost X86TTIImpl::getGatherScatterOpCost(
4709     unsigned Opcode, Type *SrcVTy, const Value *Ptr, bool VariableMask,
4710     Align Alignment, TTI::TargetCostKind CostKind,
4711     const Instruction *I = nullptr) {
4712   if (CostKind != TTI::TCK_RecipThroughput) {
4713     if ((Opcode == Instruction::Load &&
4714          isLegalMaskedGather(SrcVTy, Align(Alignment))) ||
4715         (Opcode == Instruction::Store &&
4716          isLegalMaskedScatter(SrcVTy, Align(Alignment))))
4717       return 1;
4718     return BaseT::getGatherScatterOpCost(Opcode, SrcVTy, Ptr, VariableMask,
4719                                          Alignment, CostKind, I);
4720   }
4721 
4722   assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter");
4723   PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4724   if (!PtrTy && Ptr->getType()->isVectorTy())
4725     PtrTy = dyn_cast<PointerType>(
4726         cast<VectorType>(Ptr->getType())->getElementType());
4727   assert(PtrTy && "Unexpected type for Ptr argument");
4728   unsigned AddressSpace = PtrTy->getAddressSpace();
4729 
4730   if ((Opcode == Instruction::Load &&
4731        !isLegalMaskedGather(SrcVTy, Align(Alignment))) ||
4732       (Opcode == Instruction::Store &&
4733        !isLegalMaskedScatter(SrcVTy, Align(Alignment))))
4734     return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment,
4735                            AddressSpace);
4736 
4737   return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace);
4738 }
4739 
4740 bool X86TTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1,
4741                                TargetTransformInfo::LSRCost &C2) {
4742     // X86 specific here are "instruction number 1st priority".
4743     return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost,
4744                     C1.NumIVMuls, C1.NumBaseAdds,
4745                     C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
4746            std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost,
4747                     C2.NumIVMuls, C2.NumBaseAdds,
4748                     C2.ScaleCost, C2.ImmCost, C2.SetupCost);
4749 }
4750 
4751 bool X86TTIImpl::canMacroFuseCmp() {
4752   return ST->hasMacroFusion() || ST->hasBranchFusion();
4753 }
4754 
4755 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy, Align Alignment) {
4756   if (!ST->hasAVX())
4757     return false;
4758 
4759   // The backend can't handle a single element vector.
4760   if (isa<VectorType>(DataTy) &&
4761       cast<FixedVectorType>(DataTy)->getNumElements() == 1)
4762     return false;
4763   Type *ScalarTy = DataTy->getScalarType();
4764 
4765   if (ScalarTy->isPointerTy())
4766     return true;
4767 
4768   if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
4769     return true;
4770 
4771   if (ScalarTy->isHalfTy() && ST->hasBWI() && ST->hasFP16())
4772     return true;
4773 
4774   if (!ScalarTy->isIntegerTy())
4775     return false;
4776 
4777   unsigned IntWidth = ScalarTy->getIntegerBitWidth();
4778   return IntWidth == 32 || IntWidth == 64 ||
4779          ((IntWidth == 8 || IntWidth == 16) && ST->hasBWI());
4780 }
4781 
4782 bool X86TTIImpl::isLegalMaskedStore(Type *DataType, Align Alignment) {
4783   return isLegalMaskedLoad(DataType, Alignment);
4784 }
4785 
4786 bool X86TTIImpl::isLegalNTLoad(Type *DataType, Align Alignment) {
4787   unsigned DataSize = DL.getTypeStoreSize(DataType);
4788   // The only supported nontemporal loads are for aligned vectors of 16 or 32
4789   // bytes.  Note that 32-byte nontemporal vector loads are supported by AVX2
4790   // (the equivalent stores only require AVX).
4791   if (Alignment >= DataSize && (DataSize == 16 || DataSize == 32))
4792     return DataSize == 16 ?  ST->hasSSE1() : ST->hasAVX2();
4793 
4794   return false;
4795 }
4796 
4797 bool X86TTIImpl::isLegalNTStore(Type *DataType, Align Alignment) {
4798   unsigned DataSize = DL.getTypeStoreSize(DataType);
4799 
4800   // SSE4A supports nontemporal stores of float and double at arbitrary
4801   // alignment.
4802   if (ST->hasSSE4A() && (DataType->isFloatTy() || DataType->isDoubleTy()))
4803     return true;
4804 
4805   // Besides the SSE4A subtarget exception above, only aligned stores are
4806   // available nontemporaly on any other subtarget.  And only stores with a size
4807   // of 4..32 bytes (powers of 2, only) are permitted.
4808   if (Alignment < DataSize || DataSize < 4 || DataSize > 32 ||
4809       !isPowerOf2_32(DataSize))
4810     return false;
4811 
4812   // 32-byte vector nontemporal stores are supported by AVX (the equivalent
4813   // loads require AVX2).
4814   if (DataSize == 32)
4815     return ST->hasAVX();
4816   if (DataSize == 16)
4817     return ST->hasSSE1();
4818   return true;
4819 }
4820 
4821 bool X86TTIImpl::isLegalMaskedExpandLoad(Type *DataTy) {
4822   if (!isa<VectorType>(DataTy))
4823     return false;
4824 
4825   if (!ST->hasAVX512())
4826     return false;
4827 
4828   // The backend can't handle a single element vector.
4829   if (cast<FixedVectorType>(DataTy)->getNumElements() == 1)
4830     return false;
4831 
4832   Type *ScalarTy = cast<VectorType>(DataTy)->getElementType();
4833 
4834   if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
4835     return true;
4836 
4837   if (!ScalarTy->isIntegerTy())
4838     return false;
4839 
4840   unsigned IntWidth = ScalarTy->getIntegerBitWidth();
4841   return IntWidth == 32 || IntWidth == 64 ||
4842          ((IntWidth == 8 || IntWidth == 16) && ST->hasVBMI2());
4843 }
4844 
4845 bool X86TTIImpl::isLegalMaskedCompressStore(Type *DataTy) {
4846   return isLegalMaskedExpandLoad(DataTy);
4847 }
4848 
4849 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy, Align Alignment) {
4850   // Some CPUs have better gather performance than others.
4851   // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only
4852   // enable gather with a -march.
4853   if (!(ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2())))
4854     return false;
4855 
4856   // This function is called now in two cases: from the Loop Vectorizer
4857   // and from the Scalarizer.
4858   // When the Loop Vectorizer asks about legality of the feature,
4859   // the vectorization factor is not calculated yet. The Loop Vectorizer
4860   // sends a scalar type and the decision is based on the width of the
4861   // scalar element.
4862   // Later on, the cost model will estimate usage this intrinsic based on
4863   // the vector type.
4864   // The Scalarizer asks again about legality. It sends a vector type.
4865   // In this case we can reject non-power-of-2 vectors.
4866   // We also reject single element vectors as the type legalizer can't
4867   // scalarize it.
4868   if (auto *DataVTy = dyn_cast<FixedVectorType>(DataTy)) {
4869     unsigned NumElts = DataVTy->getNumElements();
4870     if (NumElts == 1)
4871       return false;
4872     // Gather / Scatter for vector 2 is not profitable on KNL / SKX
4873     // Vector-4 of gather/scatter instruction does not exist on KNL.
4874     // We can extend it to 8 elements, but zeroing upper bits of
4875     // the mask vector will add more instructions. Right now we give the scalar
4876     // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter
4877     // instruction is better in the VariableMask case.
4878     if (ST->hasAVX512() && (NumElts == 2 || (NumElts == 4 && !ST->hasVLX())))
4879       return false;
4880   }
4881   Type *ScalarTy = DataTy->getScalarType();
4882   if (ScalarTy->isPointerTy())
4883     return true;
4884 
4885   if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
4886     return true;
4887 
4888   if (!ScalarTy->isIntegerTy())
4889     return false;
4890 
4891   unsigned IntWidth = ScalarTy->getIntegerBitWidth();
4892   return IntWidth == 32 || IntWidth == 64;
4893 }
4894 
4895 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType, Align Alignment) {
4896   // AVX2 doesn't support scatter
4897   if (!ST->hasAVX512())
4898     return false;
4899   return isLegalMaskedGather(DataType, Alignment);
4900 }
4901 
4902 bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) {
4903   EVT VT = TLI->getValueType(DL, DataType);
4904   return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT);
4905 }
4906 
4907 bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
4908   return false;
4909 }
4910 
4911 bool X86TTIImpl::areInlineCompatible(const Function *Caller,
4912                                      const Function *Callee) const {
4913   const TargetMachine &TM = getTLI()->getTargetMachine();
4914 
4915   // Work this as a subsetting of subtarget features.
4916   const FeatureBitset &CallerBits =
4917       TM.getSubtargetImpl(*Caller)->getFeatureBits();
4918   const FeatureBitset &CalleeBits =
4919       TM.getSubtargetImpl(*Callee)->getFeatureBits();
4920 
4921   FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList;
4922   FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList;
4923   return (RealCallerBits & RealCalleeBits) == RealCalleeBits;
4924 }
4925 
4926 bool X86TTIImpl::areFunctionArgsABICompatible(
4927     const Function *Caller, const Function *Callee,
4928     SmallPtrSetImpl<Argument *> &Args) const {
4929   if (!BaseT::areFunctionArgsABICompatible(Caller, Callee, Args))
4930     return false;
4931 
4932   // If we get here, we know the target features match. If one function
4933   // considers 512-bit vectors legal and the other does not, consider them
4934   // incompatible.
4935   const TargetMachine &TM = getTLI()->getTargetMachine();
4936 
4937   if (TM.getSubtarget<X86Subtarget>(*Caller).useAVX512Regs() ==
4938       TM.getSubtarget<X86Subtarget>(*Callee).useAVX512Regs())
4939     return true;
4940 
4941   // Consider the arguments compatible if they aren't vectors or aggregates.
4942   // FIXME: Look at the size of vectors.
4943   // FIXME: Look at the element types of aggregates to see if there are vectors.
4944   // FIXME: The API of this function seems intended to allow arguments
4945   // to be removed from the set, but the caller doesn't check if the set
4946   // becomes empty so that may not work in practice.
4947   return llvm::none_of(Args, [](Argument *A) {
4948     auto *EltTy = cast<PointerType>(A->getType())->getElementType();
4949     return EltTy->isVectorTy() || EltTy->isAggregateType();
4950   });
4951 }
4952 
4953 X86TTIImpl::TTI::MemCmpExpansionOptions
4954 X86TTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
4955   TTI::MemCmpExpansionOptions Options;
4956   Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize);
4957   Options.NumLoadsPerBlock = 2;
4958   // All GPR and vector loads can be unaligned.
4959   Options.AllowOverlappingLoads = true;
4960   if (IsZeroCmp) {
4961     // Only enable vector loads for equality comparison. Right now the vector
4962     // version is not as fast for three way compare (see #33329).
4963     const unsigned PreferredWidth = ST->getPreferVectorWidth();
4964     if (PreferredWidth >= 512 && ST->hasAVX512()) Options.LoadSizes.push_back(64);
4965     if (PreferredWidth >= 256 && ST->hasAVX()) Options.LoadSizes.push_back(32);
4966     if (PreferredWidth >= 128 && ST->hasSSE2()) Options.LoadSizes.push_back(16);
4967   }
4968   if (ST->is64Bit()) {
4969     Options.LoadSizes.push_back(8);
4970   }
4971   Options.LoadSizes.push_back(4);
4972   Options.LoadSizes.push_back(2);
4973   Options.LoadSizes.push_back(1);
4974   return Options;
4975 }
4976 
4977 bool X86TTIImpl::enableInterleavedAccessVectorization() {
4978   // TODO: We expect this to be beneficial regardless of arch,
4979   // but there are currently some unexplained performance artifacts on Atom.
4980   // As a temporary solution, disable on Atom.
4981   return !(ST->isAtom());
4982 }
4983 
4984 // Get estimation for interleaved load/store operations for AVX2.
4985 // \p Factor is the interleaved-access factor (stride) - number of
4986 // (interleaved) elements in the group.
4987 // \p Indices contains the indices for a strided load: when the
4988 // interleaved load has gaps they indicate which elements are used.
4989 // If Indices is empty (or if the number of indices is equal to the size
4990 // of the interleaved-access as given in \p Factor) the access has no gaps.
4991 //
4992 // As opposed to AVX-512, AVX2 does not have generic shuffles that allow
4993 // computing the cost using a generic formula as a function of generic
4994 // shuffles. We therefore use a lookup table instead, filled according to
4995 // the instruction sequences that codegen currently generates.
4996 InstructionCost X86TTIImpl::getInterleavedMemoryOpCostAVX2(
4997     unsigned Opcode, FixedVectorType *VecTy, unsigned Factor,
4998     ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace,
4999     TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) {
5000 
5001   if (UseMaskForCond || UseMaskForGaps)
5002     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5003                                              Alignment, AddressSpace, CostKind,
5004                                              UseMaskForCond, UseMaskForGaps);
5005 
5006   // We currently Support only fully-interleaved groups, with no gaps.
5007   // TODO: Support also strided loads (interleaved-groups with gaps).
5008   if (Indices.size() && Indices.size() != Factor)
5009     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5010                                              Alignment, AddressSpace, CostKind);
5011 
5012   // VecTy for interleave memop is <VF*Factor x Elt>.
5013   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
5014   // VecTy = <12 x i32>.
5015   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
5016 
5017   // This function can be called with VecTy=<6xi128>, Factor=3, in which case
5018   // the VF=2, while v2i128 is an unsupported MVT vector type
5019   // (see MachineValueType.h::getVectorVT()).
5020   if (!LegalVT.isVector())
5021     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5022                                              Alignment, AddressSpace, CostKind);
5023 
5024   unsigned VF = VecTy->getNumElements() / Factor;
5025   Type *ScalarTy = VecTy->getElementType();
5026   // Deduplicate entries, model floats/pointers as appropriately-sized integers.
5027   if (!ScalarTy->isIntegerTy())
5028     ScalarTy =
5029         Type::getIntNTy(ScalarTy->getContext(), DL.getTypeSizeInBits(ScalarTy));
5030 
5031   // Get the cost of all the memory operations.
5032   InstructionCost MemOpCosts = getMemoryOpCost(
5033       Opcode, VecTy, MaybeAlign(Alignment), AddressSpace, CostKind);
5034 
5035   auto *VT = FixedVectorType::get(ScalarTy, VF);
5036   EVT ETy = TLI->getValueType(DL, VT);
5037   if (!ETy.isSimple())
5038     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5039                                              Alignment, AddressSpace, CostKind);
5040 
5041   // TODO: Complete for other data-types and strides.
5042   // Each combination of Stride, element bit width and VF results in a different
5043   // sequence; The cost tables are therefore accessed with:
5044   // Factor (stride) and VectorType=VFxiN.
5045   // The Cost accounts only for the shuffle sequence;
5046   // The cost of the loads/stores is accounted for separately.
5047   //
5048   static const CostTblEntry AVX2InterleavedLoadTbl[] = {
5049       {2, MVT::v4i64, 6}, // (load 8i64 and) deinterleave into 2 x 4i64
5050 
5051       {3, MVT::v2i8, 10},  // (load 6i8 and) deinterleave into 3 x 2i8
5052       {3, MVT::v4i8, 4},   // (load 12i8 and) deinterleave into 3 x 4i8
5053       {3, MVT::v8i8, 9},   // (load 24i8 and) deinterleave into 3 x 8i8
5054       {3, MVT::v16i8, 11}, // (load 48i8 and) deinterleave into 3 x 16i8
5055       {3, MVT::v32i8, 13}, // (load 96i8 and) deinterleave into 3 x 32i8
5056 
5057       {3, MVT::v8i32, 17}, // (load 24i32 and) deinterleave into 3 x 8i32
5058 
5059       {4, MVT::v2i8, 12},  // (load 8i8 and) deinterleave into 4 x 2i8
5060       {4, MVT::v4i8, 4},   // (load 16i8 and) deinterleave into 4 x 4i8
5061       {4, MVT::v8i8, 20},  // (load 32i8 and) deinterleave into 4 x 8i8
5062       {4, MVT::v16i8, 39}, // (load 64i8 and) deinterleave into 4 x 16i8
5063       {4, MVT::v32i8, 80}, // (load 128i8 and) deinterleave into 4 x 32i8
5064 
5065       {8, MVT::v8i32, 40} // (load 64i32 and) deinterleave into 8 x 8i32
5066   };
5067 
5068   static const CostTblEntry AVX2InterleavedStoreTbl[] = {
5069       {2, MVT::v4i64, 6}, // interleave 2 x 4i64 into 8i64 (and store)
5070 
5071       {3, MVT::v2i8, 7},   // interleave 3 x 2i8 into 6i8 (and store)
5072       {3, MVT::v4i8, 8},   // interleave 3 x 4i8 into 12i8 (and store)
5073       {3, MVT::v8i8, 11},  // interleave 3 x 8i8 into 24i8 (and store)
5074       {3, MVT::v16i8, 11}, // interleave 3 x 16i8 into 48i8 (and store)
5075       {3, MVT::v32i8, 13}, // interleave 3 x 32i8 into 96i8 (and store)
5076 
5077       {4, MVT::v2i8, 12},  // interleave 4 x 2i8 into 8i8 (and store)
5078       {4, MVT::v4i8, 9},   // interleave 4 x 4i8 into 16i8 (and store)
5079       {4, MVT::v8i8, 10},  // interleave 4 x 8i8 into 32i8 (and store)
5080       {4, MVT::v16i8, 10}, // interleave 4 x 16i8 into 64i8 (and store)
5081       {4, MVT::v32i8, 12}  // interleave 4 x 32i8 into 128i8 (and store)
5082   };
5083 
5084   if (Opcode == Instruction::Load) {
5085     if (const auto *Entry =
5086             CostTableLookup(AVX2InterleavedLoadTbl, Factor, ETy.getSimpleVT()))
5087       return MemOpCosts + Entry->Cost;
5088   } else {
5089     assert(Opcode == Instruction::Store &&
5090            "Expected Store Instruction at this  point");
5091     if (const auto *Entry =
5092             CostTableLookup(AVX2InterleavedStoreTbl, Factor, ETy.getSimpleVT()))
5093       return MemOpCosts + Entry->Cost;
5094   }
5095 
5096   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5097                                            Alignment, AddressSpace, CostKind);
5098 }
5099 
5100 // Get estimation for interleaved load/store operations and strided load.
5101 // \p Indices contains indices for strided load.
5102 // \p Factor - the factor of interleaving.
5103 // AVX-512 provides 3-src shuffles that significantly reduces the cost.
5104 InstructionCost X86TTIImpl::getInterleavedMemoryOpCostAVX512(
5105     unsigned Opcode, FixedVectorType *VecTy, unsigned Factor,
5106     ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace,
5107     TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) {
5108 
5109   if (UseMaskForCond || UseMaskForGaps)
5110     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5111                                              Alignment, AddressSpace, CostKind,
5112                                              UseMaskForCond, UseMaskForGaps);
5113 
5114   // VecTy for interleave memop is <VF*Factor x Elt>.
5115   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
5116   // VecTy = <12 x i32>.
5117 
5118   // Calculate the number of memory operations (NumOfMemOps), required
5119   // for load/store the VecTy.
5120   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
5121   unsigned VecTySize = DL.getTypeStoreSize(VecTy);
5122   unsigned LegalVTSize = LegalVT.getStoreSize();
5123   unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize;
5124 
5125   // Get the cost of one memory operation.
5126   auto *SingleMemOpTy = FixedVectorType::get(VecTy->getElementType(),
5127                                              LegalVT.getVectorNumElements());
5128   InstructionCost MemOpCost = getMemoryOpCost(
5129       Opcode, SingleMemOpTy, MaybeAlign(Alignment), AddressSpace, CostKind);
5130 
5131   unsigned VF = VecTy->getNumElements() / Factor;
5132   MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF);
5133 
5134   if (Opcode == Instruction::Load) {
5135     // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl)
5136     // contain the cost of the optimized shuffle sequence that the
5137     // X86InterleavedAccess pass will generate.
5138     // The cost of loads and stores are computed separately from the table.
5139 
5140     // X86InterleavedAccess support only the following interleaved-access group.
5141     static const CostTblEntry AVX512InterleavedLoadTbl[] = {
5142         {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8
5143         {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8
5144         {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8
5145     };
5146 
5147     if (const auto *Entry =
5148             CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT))
5149       return NumOfMemOps * MemOpCost + Entry->Cost;
5150     //If an entry does not exist, fallback to the default implementation.
5151 
5152     // Kind of shuffle depends on number of loaded values.
5153     // If we load the entire data in one register, we can use a 1-src shuffle.
5154     // Otherwise, we'll merge 2 sources in each operation.
5155     TTI::ShuffleKind ShuffleKind =
5156         (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc;
5157 
5158     InstructionCost ShuffleCost =
5159         getShuffleCost(ShuffleKind, SingleMemOpTy, None, 0, nullptr);
5160 
5161     unsigned NumOfLoadsInInterleaveGrp =
5162         Indices.size() ? Indices.size() : Factor;
5163     auto *ResultTy = FixedVectorType::get(VecTy->getElementType(),
5164                                           VecTy->getNumElements() / Factor);
5165     InstructionCost NumOfResults =
5166         getTLI()->getTypeLegalizationCost(DL, ResultTy).first *
5167         NumOfLoadsInInterleaveGrp;
5168 
5169     // About a half of the loads may be folded in shuffles when we have only
5170     // one result. If we have more than one result, we do not fold loads at all.
5171     unsigned NumOfUnfoldedLoads =
5172         NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2;
5173 
5174     // Get a number of shuffle operations per result.
5175     unsigned NumOfShufflesPerResult =
5176         std::max((unsigned)1, (unsigned)(NumOfMemOps - 1));
5177 
5178     // The SK_MergeTwoSrc shuffle clobbers one of src operands.
5179     // When we have more than one destination, we need additional instructions
5180     // to keep sources.
5181     InstructionCost NumOfMoves = 0;
5182     if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc)
5183       NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2;
5184 
5185     InstructionCost Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost +
5186                            NumOfUnfoldedLoads * MemOpCost + NumOfMoves;
5187 
5188     return Cost;
5189   }
5190 
5191   // Store.
5192   assert(Opcode == Instruction::Store &&
5193          "Expected Store Instruction at this  point");
5194   // X86InterleavedAccess support only the following interleaved-access group.
5195   static const CostTblEntry AVX512InterleavedStoreTbl[] = {
5196       {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store)
5197       {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store)
5198       {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store)
5199 
5200       {4, MVT::v8i8, 10},  // interleave 4 x 8i8  into 32i8  (and store)
5201       {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8  (and store)
5202       {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store)
5203       {4, MVT::v64i8, 24}  // interleave 4 x 32i8 into 256i8 (and store)
5204   };
5205 
5206   if (const auto *Entry =
5207           CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT))
5208     return NumOfMemOps * MemOpCost + Entry->Cost;
5209   //If an entry does not exist, fallback to the default implementation.
5210 
5211   // There is no strided stores meanwhile. And store can't be folded in
5212   // shuffle.
5213   unsigned NumOfSources = Factor; // The number of values to be merged.
5214   InstructionCost ShuffleCost =
5215       getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, None, 0, nullptr);
5216   unsigned NumOfShufflesPerStore = NumOfSources - 1;
5217 
5218   // The SK_MergeTwoSrc shuffle clobbers one of src operands.
5219   // We need additional instructions to keep sources.
5220   unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2;
5221   InstructionCost Cost =
5222       NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) +
5223       NumOfMoves;
5224   return Cost;
5225 }
5226 
5227 InstructionCost X86TTIImpl::getInterleavedMemoryOpCost(
5228     unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
5229     Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
5230     bool UseMaskForCond, bool UseMaskForGaps) {
5231   auto isSupportedOnAVX512 = [&](Type *VecTy, bool HasBW) {
5232     Type *EltTy = cast<VectorType>(VecTy)->getElementType();
5233     if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) ||
5234         EltTy->isIntegerTy(32) || EltTy->isPointerTy())
5235       return true;
5236     if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8) ||
5237         (!ST->useSoftFloat() && ST->hasFP16() && EltTy->isHalfTy()))
5238       return HasBW;
5239     return false;
5240   };
5241   if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI()))
5242     return getInterleavedMemoryOpCostAVX512(
5243         Opcode, cast<FixedVectorType>(VecTy), Factor, Indices, Alignment,
5244         AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps);
5245   if (ST->hasAVX2())
5246     return getInterleavedMemoryOpCostAVX2(
5247         Opcode, cast<FixedVectorType>(VecTy), Factor, Indices, Alignment,
5248         AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps);
5249 
5250   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5251                                            Alignment, AddressSpace, CostKind,
5252                                            UseMaskForCond, UseMaskForGaps);
5253 }
5254