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