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