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/InstIterator.h"
47 #include "llvm/IR/IntrinsicInst.h"
48 #include "llvm/Support/Debug.h"
49 
50 using namespace llvm;
51 
52 #define DEBUG_TYPE "x86tti"
53 
54 //===----------------------------------------------------------------------===//
55 //
56 // X86 cost model.
57 //
58 //===----------------------------------------------------------------------===//
59 
60 TargetTransformInfo::PopcntSupportKind
61 X86TTIImpl::getPopcntSupport(unsigned TyWidth) {
62   assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
63   // TODO: Currently the __builtin_popcount() implementation using SSE3
64   //   instructions is inefficient. Once the problem is fixed, we should
65   //   call ST->hasSSE3() instead of ST->hasPOPCNT().
66   return ST->hasPOPCNT() ? TTI::PSK_FastHardware : TTI::PSK_Software;
67 }
68 
69 llvm::Optional<unsigned> X86TTIImpl::getCacheSize(
70   TargetTransformInfo::CacheLevel Level) const {
71   switch (Level) {
72   case TargetTransformInfo::CacheLevel::L1D:
73     //   - Penryn
74     //   - Nehalem
75     //   - Westmere
76     //   - Sandy Bridge
77     //   - Ivy Bridge
78     //   - Haswell
79     //   - Broadwell
80     //   - Skylake
81     //   - Kabylake
82     return 32 * 1024;  //  32 KByte
83   case TargetTransformInfo::CacheLevel::L2D:
84     //   - Penryn
85     //   - Nehalem
86     //   - Westmere
87     //   - Sandy Bridge
88     //   - Ivy Bridge
89     //   - Haswell
90     //   - Broadwell
91     //   - Skylake
92     //   - Kabylake
93     return 256 * 1024; // 256 KByte
94   }
95 
96   llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
97 }
98 
99 llvm::Optional<unsigned> X86TTIImpl::getCacheAssociativity(
100   TargetTransformInfo::CacheLevel Level) const {
101   //   - Penryn
102   //   - Nehalem
103   //   - Westmere
104   //   - Sandy Bridge
105   //   - Ivy Bridge
106   //   - Haswell
107   //   - Broadwell
108   //   - Skylake
109   //   - Kabylake
110   switch (Level) {
111   case TargetTransformInfo::CacheLevel::L1D:
112     LLVM_FALLTHROUGH;
113   case TargetTransformInfo::CacheLevel::L2D:
114     return 8;
115   }
116 
117   llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
118 }
119 
120 unsigned X86TTIImpl::getNumberOfRegisters(unsigned ClassID) const {
121   bool Vector = (ClassID == 1);
122   if (Vector && !ST->hasSSE1())
123     return 0;
124 
125   if (ST->is64Bit()) {
126     if (Vector && ST->hasAVX512())
127       return 32;
128     return 16;
129   }
130   return 8;
131 }
132 
133 TypeSize
134 X86TTIImpl::getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const {
135   unsigned PreferVectorWidth = ST->getPreferVectorWidth();
136   switch (K) {
137   case TargetTransformInfo::RGK_Scalar:
138     return TypeSize::getFixed(ST->is64Bit() ? 64 : 32);
139   case TargetTransformInfo::RGK_FixedWidthVector:
140     if (ST->hasAVX512() && PreferVectorWidth >= 512)
141       return TypeSize::getFixed(512);
142     if (ST->hasAVX() && PreferVectorWidth >= 256)
143       return TypeSize::getFixed(256);
144     if (ST->hasSSE1() && PreferVectorWidth >= 128)
145       return TypeSize::getFixed(128);
146     return TypeSize::getFixed(0);
147   case TargetTransformInfo::RGK_ScalableVector:
148     return TypeSize::getScalable(0);
149   }
150 
151   llvm_unreachable("Unsupported register kind");
152 }
153 
154 unsigned X86TTIImpl::getLoadStoreVecRegBitWidth(unsigned) const {
155   return getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
156       .getFixedSize();
157 }
158 
159 unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) {
160   // If the loop will not be vectorized, don't interleave the loop.
161   // Let regular unroll to unroll the loop, which saves the overflow
162   // check and memory check cost.
163   if (VF == 1)
164     return 1;
165 
166   if (ST->isAtom())
167     return 1;
168 
169   // Sandybridge and Haswell have multiple execution ports and pipelined
170   // vector units.
171   if (ST->hasAVX())
172     return 4;
173 
174   return 2;
175 }
176 
177 InstructionCost X86TTIImpl::getArithmeticInstrCost(
178     unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
179     TTI::OperandValueKind Op1Info, TTI::OperandValueKind Op2Info,
180     TTI::OperandValueProperties Opd1PropInfo,
181     TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args,
182     const Instruction *CxtI) {
183   // TODO: Handle more cost kinds.
184   if (CostKind != TTI::TCK_RecipThroughput)
185     return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
186                                          Op2Info, Opd1PropInfo,
187                                          Opd2PropInfo, Args, CxtI);
188 
189   // vXi8 multiplications are always promoted to vXi16.
190   if (Opcode == Instruction::Mul && Ty->isVectorTy() &&
191       Ty->getScalarSizeInBits() == 8) {
192     Type *WideVecTy =
193         VectorType::getExtendedElementVectorType(cast<VectorType>(Ty));
194     return getCastInstrCost(Instruction::ZExt, WideVecTy, Ty,
195                             TargetTransformInfo::CastContextHint::None,
196                             CostKind) +
197            getCastInstrCost(Instruction::Trunc, Ty, WideVecTy,
198                             TargetTransformInfo::CastContextHint::None,
199                             CostKind) +
200            getArithmeticInstrCost(Opcode, WideVecTy, CostKind, Op1Info, Op2Info,
201                                   Opd1PropInfo, Opd2PropInfo);
202   }
203 
204   // Legalize the type.
205   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
206 
207   int ISD = TLI->InstructionOpcodeToISD(Opcode);
208   assert(ISD && "Invalid opcode");
209 
210   if (ISD == ISD::MUL && Args.size() == 2 && LT.second.isVector() &&
211       LT.second.getScalarType() == MVT::i32) {
212     // Check if the operands can be represented as a smaller datatype.
213     bool Op1Signed = false, Op2Signed = false;
214     unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed);
215     unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed);
216     unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize);
217 
218     // If both are representable as i15 and at least one is constant,
219     // zero-extended, or sign-extended from vXi16 (or less pre-SSE41) then we
220     // can treat this as PMADDWD which has the same costs as a vXi16 multiply.
221     if (OpMinSize <= 15 && !ST->isPMADDWDSlow()) {
222       bool Op1Constant =
223           isa<ConstantDataVector>(Args[0]) || isa<ConstantVector>(Args[0]);
224       bool Op2Constant =
225           isa<ConstantDataVector>(Args[1]) || isa<ConstantVector>(Args[1]);
226       bool Op1Sext = isa<SExtInst>(Args[0]) &&
227                      (Op1MinSize == 15 || (Op1MinSize < 15 && !ST->hasSSE41()));
228       bool Op2Sext = isa<SExtInst>(Args[1]) &&
229                      (Op2MinSize == 15 || (Op2MinSize < 15 && !ST->hasSSE41()));
230 
231       bool IsZeroExtended = !Op1Signed || !Op2Signed;
232       bool IsConstant = Op1Constant || Op2Constant;
233       bool IsSext = Op1Sext || Op2Sext;
234       if (IsConstant || IsZeroExtended || IsSext)
235         LT.second =
236             MVT::getVectorVT(MVT::i16, 2 * LT.second.getVectorNumElements());
237     }
238   }
239 
240   // Vector multiply by pow2 will be simplified to shifts.
241   if (ISD == ISD::MUL &&
242       (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
243        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
244       Opd2PropInfo == TargetTransformInfo::OP_PowerOf2)
245     return getArithmeticInstrCost(Instruction::Shl, Ty, CostKind, Op1Info,
246                                   Op2Info, TargetTransformInfo::OP_None,
247                                   TargetTransformInfo::OP_None);
248 
249   // On X86, vector signed division by constants power-of-two are
250   // normally expanded to the sequence SRA + SRL + ADD + SRA.
251   // The OperandValue properties may not be the same as that of the previous
252   // operation; conservatively assume OP_None.
253   if ((ISD == ISD::SDIV || ISD == ISD::SREM) &&
254       (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
255        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
256       Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
257     InstructionCost Cost =
258         2 * getArithmeticInstrCost(Instruction::AShr, Ty, CostKind, Op1Info,
259                                    Op2Info, TargetTransformInfo::OP_None,
260                                    TargetTransformInfo::OP_None);
261     Cost += getArithmeticInstrCost(Instruction::LShr, Ty, CostKind, Op1Info,
262                                    Op2Info, TargetTransformInfo::OP_None,
263                                    TargetTransformInfo::OP_None);
264     Cost += getArithmeticInstrCost(Instruction::Add, Ty, CostKind, Op1Info,
265                                    Op2Info, TargetTransformInfo::OP_None,
266                                    TargetTransformInfo::OP_None);
267 
268     if (ISD == ISD::SREM) {
269       // For SREM: (X % C) is the equivalent of (X - (X/C)*C)
270       Cost += getArithmeticInstrCost(Instruction::Mul, Ty, CostKind, Op1Info,
271                                      Op2Info);
272       Cost += getArithmeticInstrCost(Instruction::Sub, Ty, CostKind, Op1Info,
273                                      Op2Info);
274     }
275 
276     return Cost;
277   }
278 
279   // Vector unsigned division/remainder will be simplified to shifts/masks.
280   if ((ISD == ISD::UDIV || ISD == ISD::UREM) &&
281       (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
282        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
283       Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
284     if (ISD == ISD::UDIV)
285       return getArithmeticInstrCost(Instruction::LShr, Ty, CostKind, Op1Info,
286                                     Op2Info, TargetTransformInfo::OP_None,
287                                     TargetTransformInfo::OP_None);
288     // UREM
289     return getArithmeticInstrCost(Instruction::And, Ty, CostKind, Op1Info,
290                                   Op2Info, TargetTransformInfo::OP_None,
291                                   TargetTransformInfo::OP_None);
292   }
293 
294   static const CostTblEntry GLMCostTable[] = {
295     { ISD::FDIV,  MVT::f32,   18 }, // divss
296     { ISD::FDIV,  MVT::v4f32, 35 }, // divps
297     { ISD::FDIV,  MVT::f64,   33 }, // divsd
298     { ISD::FDIV,  MVT::v2f64, 65 }, // divpd
299   };
300 
301   if (ST->useGLMDivSqrtCosts())
302     if (const auto *Entry = CostTableLookup(GLMCostTable, ISD,
303                                             LT.second))
304       return LT.first * Entry->Cost;
305 
306   static const CostTblEntry SLMCostTable[] = {
307     { ISD::MUL,   MVT::v4i32, 11 }, // pmulld
308     { ISD::MUL,   MVT::v8i16, 2  }, // pmullw
309     { ISD::FMUL,  MVT::f64,   2  }, // mulsd
310     { ISD::FMUL,  MVT::v2f64, 4  }, // mulpd
311     { ISD::FMUL,  MVT::v4f32, 2  }, // mulps
312     { ISD::FDIV,  MVT::f32,   17 }, // divss
313     { ISD::FDIV,  MVT::v4f32, 39 }, // divps
314     { ISD::FDIV,  MVT::f64,   32 }, // divsd
315     { ISD::FDIV,  MVT::v2f64, 69 }, // divpd
316     { ISD::FADD,  MVT::v2f64, 2  }, // addpd
317     { ISD::FSUB,  MVT::v2f64, 2  }, // subpd
318     // v2i64/v4i64 mul is custom lowered as a series of long:
319     // multiplies(3), shifts(3) and adds(2)
320     // slm muldq version throughput is 2 and addq throughput 4
321     // thus: 3X2 (muldq throughput) + 3X1 (shift throughput) +
322     //       3X4 (addq throughput) = 17
323     { ISD::MUL,   MVT::v2i64, 17 },
324     // slm addq\subq throughput is 4
325     { ISD::ADD,   MVT::v2i64, 4  },
326     { ISD::SUB,   MVT::v2i64, 4  },
327   };
328 
329   if (ST->useSLMArithCosts()) {
330     if (Args.size() == 2 && ISD == ISD::MUL && LT.second == MVT::v4i32) {
331       // Check if the operands can be shrinked into a smaller datatype.
332       // TODO: Merge this into generiic vXi32 MUL patterns above.
333       bool Op1Signed = false;
334       unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed);
335       bool Op2Signed = false;
336       unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed);
337 
338       bool SignedMode = Op1Signed || Op2Signed;
339       unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize);
340 
341       if (OpMinSize <= 7)
342         return LT.first * 3; // pmullw/sext
343       if (!SignedMode && OpMinSize <= 8)
344         return LT.first * 3; // pmullw/zext
345       if (OpMinSize <= 15)
346         return LT.first * 5; // pmullw/pmulhw/pshuf
347       if (!SignedMode && OpMinSize <= 16)
348         return LT.first * 5; // pmullw/pmulhw/pshuf
349     }
350 
351     if (const auto *Entry = CostTableLookup(SLMCostTable, ISD,
352                                             LT.second)) {
353       return LT.first * Entry->Cost;
354     }
355   }
356 
357   static const CostTblEntry AVX512BWUniformConstCostTable[] = {
358     { ISD::SHL,  MVT::v64i8,   2 }, // psllw + pand.
359     { ISD::SRL,  MVT::v64i8,   2 }, // psrlw + pand.
360     { ISD::SRA,  MVT::v64i8,   4 }, // psrlw, pand, pxor, psubb.
361   };
362 
363   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
364       ST->hasBWI()) {
365     if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD,
366                                             LT.second))
367       return LT.first * Entry->Cost;
368   }
369 
370   static const CostTblEntry AVX512UniformConstCostTable[] = {
371     { ISD::SRA,  MVT::v2i64,   1 },
372     { ISD::SRA,  MVT::v4i64,   1 },
373     { ISD::SRA,  MVT::v8i64,   1 },
374 
375     { ISD::SHL,  MVT::v64i8,   4 }, // psllw + pand.
376     { ISD::SRL,  MVT::v64i8,   4 }, // psrlw + pand.
377     { ISD::SRA,  MVT::v64i8,   8 }, // psrlw, pand, pxor, psubb.
378 
379     { ISD::SDIV, MVT::v16i32,  6 }, // pmuludq sequence
380     { ISD::SREM, MVT::v16i32,  8 }, // pmuludq+mul+sub sequence
381     { ISD::UDIV, MVT::v16i32,  5 }, // pmuludq sequence
382     { ISD::UREM, MVT::v16i32,  7 }, // pmuludq+mul+sub sequence
383   };
384 
385   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
386       ST->hasAVX512()) {
387     if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD,
388                                             LT.second))
389       return LT.first * Entry->Cost;
390   }
391 
392   static const CostTblEntry AVX2UniformConstCostTable[] = {
393     { ISD::SHL,  MVT::v32i8,   2 }, // psllw + pand.
394     { ISD::SRL,  MVT::v32i8,   2 }, // psrlw + pand.
395     { ISD::SRA,  MVT::v32i8,   4 }, // psrlw, pand, pxor, psubb.
396 
397     { ISD::SRA,  MVT::v4i64,   4 }, // 2 x psrad + shuffle.
398 
399     { ISD::SDIV, MVT::v8i32,   6 }, // pmuludq sequence
400     { ISD::SREM, MVT::v8i32,   8 }, // pmuludq+mul+sub sequence
401     { ISD::UDIV, MVT::v8i32,   5 }, // pmuludq sequence
402     { ISD::UREM, MVT::v8i32,   7 }, // pmuludq+mul+sub sequence
403   };
404 
405   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
406       ST->hasAVX2()) {
407     if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD,
408                                             LT.second))
409       return LT.first * Entry->Cost;
410   }
411 
412   static const CostTblEntry SSE2UniformConstCostTable[] = {
413     { ISD::SHL,  MVT::v16i8,     2 }, // psllw + pand.
414     { ISD::SRL,  MVT::v16i8,     2 }, // psrlw + pand.
415     { ISD::SRA,  MVT::v16i8,     4 }, // psrlw, pand, pxor, psubb.
416 
417     { ISD::SHL,  MVT::v32i8,   4+2 }, // 2*(psllw + pand) + split.
418     { ISD::SRL,  MVT::v32i8,   4+2 }, // 2*(psrlw + pand) + split.
419     { ISD::SRA,  MVT::v32i8,   8+2 }, // 2*(psrlw, pand, pxor, psubb) + split.
420 
421     { ISD::SDIV, MVT::v8i32,  12+2 }, // 2*pmuludq sequence + split.
422     { ISD::SREM, MVT::v8i32,  16+2 }, // 2*pmuludq+mul+sub sequence + split.
423     { ISD::SDIV, MVT::v4i32,     6 }, // pmuludq sequence
424     { ISD::SREM, MVT::v4i32,     8 }, // pmuludq+mul+sub sequence
425     { ISD::UDIV, MVT::v8i32,  10+2 }, // 2*pmuludq sequence + split.
426     { ISD::UREM, MVT::v8i32,  14+2 }, // 2*pmuludq+mul+sub sequence + split.
427     { ISD::UDIV, MVT::v4i32,     5 }, // pmuludq sequence
428     { ISD::UREM, MVT::v4i32,     7 }, // pmuludq+mul+sub sequence
429   };
430 
431   // XOP has faster vXi8 shifts.
432   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
433       ST->hasSSE2() && !ST->hasXOP()) {
434     if (const auto *Entry =
435             CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second))
436       return LT.first * Entry->Cost;
437   }
438 
439   static const CostTblEntry AVX512BWConstCostTable[] = {
440     { ISD::SDIV, MVT::v64i8,  14 }, // 2*ext+2*pmulhw sequence
441     { ISD::SREM, MVT::v64i8,  16 }, // 2*ext+2*pmulhw+mul+sub sequence
442     { ISD::UDIV, MVT::v64i8,  14 }, // 2*ext+2*pmulhw sequence
443     { ISD::UREM, MVT::v64i8,  16 }, // 2*ext+2*pmulhw+mul+sub sequence
444     { ISD::SDIV, MVT::v32i16,  6 }, // vpmulhw sequence
445     { ISD::SREM, MVT::v32i16,  8 }, // vpmulhw+mul+sub sequence
446     { ISD::UDIV, MVT::v32i16,  6 }, // vpmulhuw sequence
447     { ISD::UREM, MVT::v32i16,  8 }, // vpmulhuw+mul+sub sequence
448   };
449 
450   if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
451        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
452       ST->hasBWI()) {
453     if (const auto *Entry =
454             CostTableLookup(AVX512BWConstCostTable, ISD, LT.second))
455       return LT.first * Entry->Cost;
456   }
457 
458   static const CostTblEntry AVX512ConstCostTable[] = {
459     { ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence
460     { ISD::SREM, MVT::v16i32, 17 }, // vpmuldq+mul+sub sequence
461     { ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence
462     { ISD::UREM, MVT::v16i32, 17 }, // vpmuludq+mul+sub sequence
463     { ISD::SDIV, MVT::v64i8,  28 }, // 4*ext+4*pmulhw sequence
464     { ISD::SREM, MVT::v64i8,  32 }, // 4*ext+4*pmulhw+mul+sub sequence
465     { ISD::UDIV, MVT::v64i8,  28 }, // 4*ext+4*pmulhw sequence
466     { ISD::UREM, MVT::v64i8,  32 }, // 4*ext+4*pmulhw+mul+sub sequence
467     { ISD::SDIV, MVT::v32i16, 12 }, // 2*vpmulhw sequence
468     { ISD::SREM, MVT::v32i16, 16 }, // 2*vpmulhw+mul+sub sequence
469     { ISD::UDIV, MVT::v32i16, 12 }, // 2*vpmulhuw sequence
470     { ISD::UREM, MVT::v32i16, 16 }, // 2*vpmulhuw+mul+sub sequence
471   };
472 
473   if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
474        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
475       ST->hasAVX512()) {
476     if (const auto *Entry =
477             CostTableLookup(AVX512ConstCostTable, ISD, LT.second))
478       return LT.first * Entry->Cost;
479   }
480 
481   static const CostTblEntry AVX2ConstCostTable[] = {
482     { ISD::SDIV, MVT::v32i8,  14 }, // 2*ext+2*pmulhw sequence
483     { ISD::SREM, MVT::v32i8,  16 }, // 2*ext+2*pmulhw+mul+sub sequence
484     { ISD::UDIV, MVT::v32i8,  14 }, // 2*ext+2*pmulhw sequence
485     { ISD::UREM, MVT::v32i8,  16 }, // 2*ext+2*pmulhw+mul+sub sequence
486     { ISD::SDIV, MVT::v16i16,  6 }, // vpmulhw sequence
487     { ISD::SREM, MVT::v16i16,  8 }, // vpmulhw+mul+sub sequence
488     { ISD::UDIV, MVT::v16i16,  6 }, // vpmulhuw sequence
489     { ISD::UREM, MVT::v16i16,  8 }, // vpmulhuw+mul+sub sequence
490     { ISD::SDIV, MVT::v8i32,  15 }, // vpmuldq sequence
491     { ISD::SREM, MVT::v8i32,  19 }, // vpmuldq+mul+sub sequence
492     { ISD::UDIV, MVT::v8i32,  15 }, // vpmuludq sequence
493     { ISD::UREM, MVT::v8i32,  19 }, // vpmuludq+mul+sub sequence
494   };
495 
496   if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
497        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
498       ST->hasAVX2()) {
499     if (const auto *Entry = CostTableLookup(AVX2ConstCostTable, ISD, LT.second))
500       return LT.first * Entry->Cost;
501   }
502 
503   static const CostTblEntry SSE2ConstCostTable[] = {
504     { ISD::SDIV, MVT::v32i8,  28+2 }, // 4*ext+4*pmulhw sequence + split.
505     { ISD::SREM, MVT::v32i8,  32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split.
506     { ISD::SDIV, MVT::v16i8,    14 }, // 2*ext+2*pmulhw sequence
507     { ISD::SREM, MVT::v16i8,    16 }, // 2*ext+2*pmulhw+mul+sub sequence
508     { ISD::UDIV, MVT::v32i8,  28+2 }, // 4*ext+4*pmulhw sequence + split.
509     { ISD::UREM, MVT::v32i8,  32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split.
510     { ISD::UDIV, MVT::v16i8,    14 }, // 2*ext+2*pmulhw sequence
511     { ISD::UREM, MVT::v16i8,    16 }, // 2*ext+2*pmulhw+mul+sub sequence
512     { ISD::SDIV, MVT::v16i16, 12+2 }, // 2*pmulhw sequence + split.
513     { ISD::SREM, MVT::v16i16, 16+2 }, // 2*pmulhw+mul+sub sequence + split.
514     { ISD::SDIV, MVT::v8i16,     6 }, // pmulhw sequence
515     { ISD::SREM, MVT::v8i16,     8 }, // pmulhw+mul+sub sequence
516     { ISD::UDIV, MVT::v16i16, 12+2 }, // 2*pmulhuw sequence + split.
517     { ISD::UREM, MVT::v16i16, 16+2 }, // 2*pmulhuw+mul+sub sequence + split.
518     { ISD::UDIV, MVT::v8i16,     6 }, // pmulhuw sequence
519     { ISD::UREM, MVT::v8i16,     8 }, // pmulhuw+mul+sub sequence
520     { ISD::SDIV, MVT::v8i32,  38+2 }, // 2*pmuludq sequence + split.
521     { ISD::SREM, MVT::v8i32,  48+2 }, // 2*pmuludq+mul+sub sequence + split.
522     { ISD::SDIV, MVT::v4i32,    19 }, // pmuludq sequence
523     { ISD::SREM, MVT::v4i32,    24 }, // pmuludq+mul+sub sequence
524     { ISD::UDIV, MVT::v8i32,  30+2 }, // 2*pmuludq sequence + split.
525     { ISD::UREM, MVT::v8i32,  40+2 }, // 2*pmuludq+mul+sub sequence + split.
526     { ISD::UDIV, MVT::v4i32,    15 }, // pmuludq sequence
527     { ISD::UREM, MVT::v4i32,    20 }, // pmuludq+mul+sub sequence
528   };
529 
530   if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
531        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
532       ST->hasSSE2()) {
533     // pmuldq sequence.
534     if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX())
535       return LT.first * 32;
536     if (ISD == ISD::SREM && LT.second == MVT::v8i32 && ST->hasAVX())
537       return LT.first * 38;
538     if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
539       return LT.first * 15;
540     if (ISD == ISD::SREM && LT.second == MVT::v4i32 && ST->hasSSE41())
541       return LT.first * 20;
542 
543     if (const auto *Entry = CostTableLookup(SSE2ConstCostTable, ISD, LT.second))
544       return LT.first * Entry->Cost;
545   }
546 
547   static const CostTblEntry AVX512BWShiftCostTable[] = {
548     { ISD::SHL,   MVT::v16i8,      4 }, // extend/vpsllvw/pack sequence.
549     { ISD::SRL,   MVT::v16i8,      4 }, // extend/vpsrlvw/pack sequence.
550     { ISD::SRA,   MVT::v16i8,      4 }, // extend/vpsravw/pack sequence.
551     { ISD::SHL,   MVT::v32i8,      4 }, // extend/vpsllvw/pack sequence.
552     { ISD::SRL,   MVT::v32i8,      4 }, // extend/vpsrlvw/pack sequence.
553     { ISD::SRA,   MVT::v32i8,      6 }, // extend/vpsravw/pack sequence.
554     { ISD::SHL,   MVT::v64i8,      6 }, // extend/vpsllvw/pack sequence.
555     { ISD::SRL,   MVT::v64i8,      7 }, // extend/vpsrlvw/pack sequence.
556     { ISD::SRA,   MVT::v64i8,     15 }, // extend/vpsravw/pack sequence.
557 
558     { ISD::SHL,   MVT::v8i16,      1 }, // vpsllvw
559     { ISD::SRL,   MVT::v8i16,      1 }, // vpsrlvw
560     { ISD::SRA,   MVT::v8i16,      1 }, // vpsravw
561     { ISD::SHL,   MVT::v16i16,     1 }, // vpsllvw
562     { ISD::SRL,   MVT::v16i16,     1 }, // vpsrlvw
563     { ISD::SRA,   MVT::v16i16,     1 }, // vpsravw
564     { ISD::SHL,   MVT::v32i16,     1 }, // vpsllvw
565     { ISD::SRL,   MVT::v32i16,     1 }, // vpsrlvw
566     { ISD::SRA,   MVT::v32i16,     1 }, // vpsravw
567   };
568 
569   if (ST->hasBWI())
570     if (const auto *Entry = CostTableLookup(AVX512BWShiftCostTable, ISD, LT.second))
571       return LT.first * Entry->Cost;
572 
573   static const CostTblEntry AVX2UniformCostTable[] = {
574     // Uniform splats are cheaper for the following instructions.
575     { ISD::SHL,  MVT::v16i16, 1 }, // psllw.
576     { ISD::SRL,  MVT::v16i16, 1 }, // psrlw.
577     { ISD::SRA,  MVT::v16i16, 1 }, // psraw.
578     { ISD::SHL,  MVT::v32i16, 2 }, // 2*psllw.
579     { ISD::SRL,  MVT::v32i16, 2 }, // 2*psrlw.
580     { ISD::SRA,  MVT::v32i16, 2 }, // 2*psraw.
581 
582     { ISD::SHL,  MVT::v8i32,  1 }, // pslld
583     { ISD::SRL,  MVT::v8i32,  1 }, // psrld
584     { ISD::SRA,  MVT::v8i32,  1 }, // psrad
585     { ISD::SHL,  MVT::v4i64,  1 }, // psllq
586     { ISD::SRL,  MVT::v4i64,  1 }, // psrlq
587   };
588 
589   if (ST->hasAVX2() &&
590       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
591        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
592     if (const auto *Entry =
593             CostTableLookup(AVX2UniformCostTable, ISD, LT.second))
594       return LT.first * Entry->Cost;
595   }
596 
597   static const CostTblEntry SSE2UniformCostTable[] = {
598     // Uniform splats are cheaper for the following instructions.
599     { ISD::SHL,  MVT::v8i16,  1 }, // psllw.
600     { ISD::SHL,  MVT::v4i32,  1 }, // pslld
601     { ISD::SHL,  MVT::v2i64,  1 }, // psllq.
602 
603     { ISD::SRL,  MVT::v8i16,  1 }, // psrlw.
604     { ISD::SRL,  MVT::v4i32,  1 }, // psrld.
605     { ISD::SRL,  MVT::v2i64,  1 }, // psrlq.
606 
607     { ISD::SRA,  MVT::v8i16,  1 }, // psraw.
608     { ISD::SRA,  MVT::v4i32,  1 }, // psrad.
609   };
610 
611   if (ST->hasSSE2() &&
612       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
613        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
614     if (const auto *Entry =
615             CostTableLookup(SSE2UniformCostTable, ISD, LT.second))
616       return LT.first * Entry->Cost;
617   }
618 
619   static const CostTblEntry AVX512DQCostTable[] = {
620     { ISD::MUL,  MVT::v2i64, 2 }, // pmullq
621     { ISD::MUL,  MVT::v4i64, 2 }, // pmullq
622     { ISD::MUL,  MVT::v8i64, 2 }  // pmullq
623   };
624 
625   // Look for AVX512DQ lowering tricks for custom cases.
626   if (ST->hasDQI())
627     if (const auto *Entry = CostTableLookup(AVX512DQCostTable, ISD, LT.second))
628       return LT.first * Entry->Cost;
629 
630   static const CostTblEntry AVX512BWCostTable[] = {
631     { ISD::SHL,   MVT::v64i8,     11 }, // vpblendvb sequence.
632     { ISD::SRL,   MVT::v64i8,     11 }, // vpblendvb sequence.
633     { ISD::SRA,   MVT::v64i8,     24 }, // vpblendvb sequence.
634   };
635 
636   // Look for AVX512BW lowering tricks for custom cases.
637   if (ST->hasBWI())
638     if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD, LT.second))
639       return LT.first * Entry->Cost;
640 
641   static const CostTblEntry AVX512CostTable[] = {
642     { ISD::SHL,     MVT::v4i32,      1 },
643     { ISD::SRL,     MVT::v4i32,      1 },
644     { ISD::SRA,     MVT::v4i32,      1 },
645     { ISD::SHL,     MVT::v8i32,      1 },
646     { ISD::SRL,     MVT::v8i32,      1 },
647     { ISD::SRA,     MVT::v8i32,      1 },
648     { ISD::SHL,     MVT::v16i32,     1 },
649     { ISD::SRL,     MVT::v16i32,     1 },
650     { ISD::SRA,     MVT::v16i32,     1 },
651 
652     { ISD::SHL,     MVT::v2i64,      1 },
653     { ISD::SRL,     MVT::v2i64,      1 },
654     { ISD::SHL,     MVT::v4i64,      1 },
655     { ISD::SRL,     MVT::v4i64,      1 },
656     { ISD::SHL,     MVT::v8i64,      1 },
657     { ISD::SRL,     MVT::v8i64,      1 },
658 
659     { ISD::SRA,     MVT::v2i64,      1 },
660     { ISD::SRA,     MVT::v4i64,      1 },
661     { ISD::SRA,     MVT::v8i64,      1 },
662 
663     { ISD::MUL,     MVT::v16i32,     1 }, // pmulld (Skylake from agner.org)
664     { ISD::MUL,     MVT::v8i32,      1 }, // pmulld (Skylake from agner.org)
665     { ISD::MUL,     MVT::v4i32,      1 }, // pmulld (Skylake from agner.org)
666     { ISD::MUL,     MVT::v8i64,      6 }, // 3*pmuludq/3*shift/2*add
667     { ISD::MUL,     MVT::i64,        1 }, // Skylake from http://www.agner.org/
668 
669     { ISD::FNEG,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
670     { ISD::FADD,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
671     { ISD::FSUB,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
672     { ISD::FMUL,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
673     { ISD::FDIV,    MVT::f64,        4 }, // Skylake from http://www.agner.org/
674     { ISD::FDIV,    MVT::v2f64,      4 }, // Skylake from http://www.agner.org/
675     { ISD::FDIV,    MVT::v4f64,      8 }, // Skylake from http://www.agner.org/
676     { ISD::FDIV,    MVT::v8f64,     16 }, // Skylake from http://www.agner.org/
677 
678     { ISD::FNEG,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
679     { ISD::FADD,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
680     { ISD::FSUB,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
681     { ISD::FMUL,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
682     { ISD::FDIV,    MVT::f32,        3 }, // Skylake from http://www.agner.org/
683     { ISD::FDIV,    MVT::v4f32,      3 }, // Skylake from http://www.agner.org/
684     { ISD::FDIV,    MVT::v8f32,      5 }, // Skylake from http://www.agner.org/
685     { ISD::FDIV,    MVT::v16f32,    10 }, // Skylake from http://www.agner.org/
686   };
687 
688   if (ST->hasAVX512())
689     if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second))
690       return LT.first * Entry->Cost;
691 
692   static const CostTblEntry AVX2ShiftCostTable[] = {
693     // Shifts on vXi64/vXi32 on AVX2 is legal even though we declare to
694     // customize them to detect the cases where shift amount is a scalar one.
695     { ISD::SHL,     MVT::v4i32,    2 }, // vpsllvd (Haswell from agner.org)
696     { ISD::SRL,     MVT::v4i32,    2 }, // vpsrlvd (Haswell from agner.org)
697     { ISD::SRA,     MVT::v4i32,    2 }, // vpsravd (Haswell from agner.org)
698     { ISD::SHL,     MVT::v8i32,    2 }, // vpsllvd (Haswell from agner.org)
699     { ISD::SRL,     MVT::v8i32,    2 }, // vpsrlvd (Haswell from agner.org)
700     { ISD::SRA,     MVT::v8i32,    2 }, // vpsravd (Haswell from agner.org)
701     { ISD::SHL,     MVT::v2i64,    1 }, // vpsllvq (Haswell from agner.org)
702     { ISD::SRL,     MVT::v2i64,    1 }, // vpsrlvq (Haswell from agner.org)
703     { ISD::SHL,     MVT::v4i64,    1 }, // vpsllvq (Haswell from agner.org)
704     { ISD::SRL,     MVT::v4i64,    1 }, // vpsrlvq (Haswell from agner.org)
705   };
706 
707   if (ST->hasAVX512()) {
708     if (ISD == ISD::SHL && LT.second == MVT::v32i16 &&
709         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
710          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
711       // On AVX512, a packed v32i16 shift left by a constant build_vector
712       // is lowered into a vector multiply (vpmullw).
713       return getArithmeticInstrCost(Instruction::Mul, Ty, CostKind,
714                                     Op1Info, Op2Info,
715                                     TargetTransformInfo::OP_None,
716                                     TargetTransformInfo::OP_None);
717   }
718 
719   // Look for AVX2 lowering tricks (XOP is always better at v4i32 shifts).
720   if (ST->hasAVX2() && !(ST->hasXOP() && LT.second == MVT::v4i32)) {
721     if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
722         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
723          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
724       // On AVX2, a packed v16i16 shift left by a constant build_vector
725       // is lowered into a vector multiply (vpmullw).
726       return getArithmeticInstrCost(Instruction::Mul, Ty, CostKind,
727                                     Op1Info, Op2Info,
728                                     TargetTransformInfo::OP_None,
729                                     TargetTransformInfo::OP_None);
730 
731     if (const auto *Entry = CostTableLookup(AVX2ShiftCostTable, ISD, LT.second))
732       return LT.first * Entry->Cost;
733   }
734 
735   static const CostTblEntry XOPShiftCostTable[] = {
736     // 128bit shifts take 1cy, but right shifts require negation beforehand.
737     { ISD::SHL,     MVT::v16i8,    1 },
738     { ISD::SRL,     MVT::v16i8,    2 },
739     { ISD::SRA,     MVT::v16i8,    2 },
740     { ISD::SHL,     MVT::v8i16,    1 },
741     { ISD::SRL,     MVT::v8i16,    2 },
742     { ISD::SRA,     MVT::v8i16,    2 },
743     { ISD::SHL,     MVT::v4i32,    1 },
744     { ISD::SRL,     MVT::v4i32,    2 },
745     { ISD::SRA,     MVT::v4i32,    2 },
746     { ISD::SHL,     MVT::v2i64,    1 },
747     { ISD::SRL,     MVT::v2i64,    2 },
748     { ISD::SRA,     MVT::v2i64,    2 },
749     // 256bit shifts require splitting if AVX2 didn't catch them above.
750     { ISD::SHL,     MVT::v32i8,  2+2 },
751     { ISD::SRL,     MVT::v32i8,  4+2 },
752     { ISD::SRA,     MVT::v32i8,  4+2 },
753     { ISD::SHL,     MVT::v16i16, 2+2 },
754     { ISD::SRL,     MVT::v16i16, 4+2 },
755     { ISD::SRA,     MVT::v16i16, 4+2 },
756     { ISD::SHL,     MVT::v8i32,  2+2 },
757     { ISD::SRL,     MVT::v8i32,  4+2 },
758     { ISD::SRA,     MVT::v8i32,  4+2 },
759     { ISD::SHL,     MVT::v4i64,  2+2 },
760     { ISD::SRL,     MVT::v4i64,  4+2 },
761     { ISD::SRA,     MVT::v4i64,  4+2 },
762   };
763 
764   // Look for XOP lowering tricks.
765   if (ST->hasXOP()) {
766     // If the right shift is constant then we'll fold the negation so
767     // it's as cheap as a left shift.
768     int ShiftISD = ISD;
769     if ((ShiftISD == ISD::SRL || ShiftISD == ISD::SRA) &&
770         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
771          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
772       ShiftISD = ISD::SHL;
773     if (const auto *Entry =
774             CostTableLookup(XOPShiftCostTable, ShiftISD, LT.second))
775       return LT.first * Entry->Cost;
776   }
777 
778   static const CostTblEntry SSE2UniformShiftCostTable[] = {
779     // Uniform splats are cheaper for the following instructions.
780     { ISD::SHL,  MVT::v16i16, 2+2 }, // 2*psllw + split.
781     { ISD::SHL,  MVT::v8i32,  2+2 }, // 2*pslld + split.
782     { ISD::SHL,  MVT::v4i64,  2+2 }, // 2*psllq + split.
783 
784     { ISD::SRL,  MVT::v16i16, 2+2 }, // 2*psrlw + split.
785     { ISD::SRL,  MVT::v8i32,  2+2 }, // 2*psrld + split.
786     { ISD::SRL,  MVT::v4i64,  2+2 }, // 2*psrlq + split.
787 
788     { ISD::SRA,  MVT::v16i16, 2+2 }, // 2*psraw + split.
789     { ISD::SRA,  MVT::v8i32,  2+2 }, // 2*psrad + split.
790     { ISD::SRA,  MVT::v2i64,    4 }, // 2*psrad + shuffle.
791     { ISD::SRA,  MVT::v4i64,  8+2 }, // 2*(2*psrad + shuffle) + split.
792   };
793 
794   if (ST->hasSSE2() &&
795       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
796        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
797 
798     // Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table.
799     if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2())
800       return LT.first * 4; // 2*psrad + shuffle.
801 
802     if (const auto *Entry =
803             CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second))
804       return LT.first * Entry->Cost;
805   }
806 
807   if (ISD == ISD::SHL &&
808       Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
809     MVT VT = LT.second;
810     // Vector shift left by non uniform constant can be lowered
811     // into vector multiply.
812     if (((VT == MVT::v8i16 || VT == MVT::v4i32) && ST->hasSSE2()) ||
813         ((VT == MVT::v16i16 || VT == MVT::v8i32) && ST->hasAVX()))
814       ISD = ISD::MUL;
815   }
816 
817   static const CostTblEntry AVX2CostTable[] = {
818     { ISD::SHL,  MVT::v16i8,      6 }, // vpblendvb sequence.
819     { ISD::SHL,  MVT::v32i8,      6 }, // vpblendvb sequence.
820     { ISD::SHL,  MVT::v64i8,     12 }, // 2*vpblendvb sequence.
821     { ISD::SHL,  MVT::v8i16,      5 }, // extend/vpsrlvd/pack sequence.
822     { ISD::SHL,  MVT::v16i16,     7 }, // extend/vpsrlvd/pack sequence.
823     { ISD::SHL,  MVT::v32i16,    14 }, // 2*extend/vpsrlvd/pack sequence.
824 
825     { ISD::SRL,  MVT::v16i8,      6 }, // vpblendvb sequence.
826     { ISD::SRL,  MVT::v32i8,      6 }, // vpblendvb sequence.
827     { ISD::SRL,  MVT::v64i8,     12 }, // 2*vpblendvb sequence.
828     { ISD::SRL,  MVT::v8i16,      5 }, // extend/vpsrlvd/pack sequence.
829     { ISD::SRL,  MVT::v16i16,     7 }, // extend/vpsrlvd/pack sequence.
830     { ISD::SRL,  MVT::v32i16,    14 }, // 2*extend/vpsrlvd/pack sequence.
831 
832     { ISD::SRA,  MVT::v16i8,     17 }, // vpblendvb sequence.
833     { ISD::SRA,  MVT::v32i8,     17 }, // vpblendvb sequence.
834     { ISD::SRA,  MVT::v64i8,     34 }, // 2*vpblendvb sequence.
835     { ISD::SRA,  MVT::v8i16,      5 }, // extend/vpsravd/pack sequence.
836     { ISD::SRA,  MVT::v16i16,     7 }, // extend/vpsravd/pack sequence.
837     { ISD::SRA,  MVT::v32i16,    14 }, // 2*extend/vpsravd/pack sequence.
838     { ISD::SRA,  MVT::v2i64,      2 }, // srl/xor/sub sequence.
839     { ISD::SRA,  MVT::v4i64,      2 }, // srl/xor/sub sequence.
840 
841     { ISD::SUB,  MVT::v32i8,      1 }, // psubb
842     { ISD::ADD,  MVT::v32i8,      1 }, // paddb
843     { ISD::SUB,  MVT::v16i16,     1 }, // psubw
844     { ISD::ADD,  MVT::v16i16,     1 }, // paddw
845     { ISD::SUB,  MVT::v8i32,      1 }, // psubd
846     { ISD::ADD,  MVT::v8i32,      1 }, // paddd
847     { ISD::SUB,  MVT::v4i64,      1 }, // psubq
848     { ISD::ADD,  MVT::v4i64,      1 }, // paddq
849 
850     { ISD::MUL,  MVT::v16i16,     1 }, // pmullw
851     { ISD::MUL,  MVT::v8i32,      2 }, // pmulld (Haswell from agner.org)
852     { ISD::MUL,  MVT::v4i64,      6 }, // 3*pmuludq/3*shift/2*add
853 
854     { ISD::FNEG, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
855     { ISD::FNEG, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
856     { ISD::FADD, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
857     { ISD::FADD, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
858     { ISD::FSUB, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
859     { ISD::FSUB, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
860     { ISD::FMUL, MVT::f64,        1 }, // Haswell from http://www.agner.org/
861     { ISD::FMUL, MVT::v2f64,      1 }, // Haswell from http://www.agner.org/
862     { ISD::FMUL, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
863     { ISD::FMUL, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
864 
865     { ISD::FDIV, MVT::f32,        7 }, // Haswell from http://www.agner.org/
866     { ISD::FDIV, MVT::v4f32,      7 }, // Haswell from http://www.agner.org/
867     { ISD::FDIV, MVT::v8f32,     14 }, // Haswell from http://www.agner.org/
868     { ISD::FDIV, MVT::f64,       14 }, // Haswell from http://www.agner.org/
869     { ISD::FDIV, MVT::v2f64,     14 }, // Haswell from http://www.agner.org/
870     { ISD::FDIV, MVT::v4f64,     28 }, // Haswell from http://www.agner.org/
871   };
872 
873   // Look for AVX2 lowering tricks for custom cases.
874   if (ST->hasAVX2())
875     if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second))
876       return LT.first * Entry->Cost;
877 
878   static const CostTblEntry AVX1CostTable[] = {
879     // We don't have to scalarize unsupported ops. We can issue two half-sized
880     // operations and we only need to extract the upper YMM half.
881     // Two ops + 1 extract + 1 insert = 4.
882     { ISD::MUL,     MVT::v16i16,     4 },
883     { ISD::MUL,     MVT::v8i32,      5 }, // BTVER2 from http://www.agner.org/
884     { ISD::MUL,     MVT::v4i64,     12 },
885 
886     { ISD::SUB,     MVT::v32i8,      4 },
887     { ISD::ADD,     MVT::v32i8,      4 },
888     { ISD::SUB,     MVT::v16i16,     4 },
889     { ISD::ADD,     MVT::v16i16,     4 },
890     { ISD::SUB,     MVT::v8i32,      4 },
891     { ISD::ADD,     MVT::v8i32,      4 },
892     { ISD::SUB,     MVT::v4i64,      4 },
893     { ISD::ADD,     MVT::v4i64,      4 },
894 
895     { ISD::SHL,     MVT::v32i8,     22 }, // pblendvb sequence + split.
896     { ISD::SHL,     MVT::v8i16,      6 }, // pblendvb sequence.
897     { ISD::SHL,     MVT::v16i16,    13 }, // pblendvb sequence + split.
898     { ISD::SHL,     MVT::v4i32,      3 }, // pslld/paddd/cvttps2dq/pmulld
899     { ISD::SHL,     MVT::v8i32,      9 }, // pslld/paddd/cvttps2dq/pmulld + split
900     { ISD::SHL,     MVT::v2i64,      2 }, // Shift each lane + blend.
901     { ISD::SHL,     MVT::v4i64,      6 }, // Shift each lane + blend + split.
902 
903     { ISD::SRL,     MVT::v32i8,     23 }, // pblendvb sequence + split.
904     { ISD::SRL,     MVT::v16i16,    28 }, // pblendvb sequence + split.
905     { ISD::SRL,     MVT::v4i32,      6 }, // Shift each lane + blend.
906     { ISD::SRL,     MVT::v8i32,     14 }, // Shift each lane + blend + split.
907     { ISD::SRL,     MVT::v2i64,      2 }, // Shift each lane + blend.
908     { ISD::SRL,     MVT::v4i64,      6 }, // Shift each lane + blend + split.
909 
910     { ISD::SRA,     MVT::v32i8,     44 }, // pblendvb sequence + split.
911     { ISD::SRA,     MVT::v16i16,    28 }, // pblendvb sequence + split.
912     { ISD::SRA,     MVT::v4i32,      6 }, // Shift each lane + blend.
913     { ISD::SRA,     MVT::v8i32,     14 }, // Shift each lane + blend + split.
914     { ISD::SRA,     MVT::v2i64,      5 }, // Shift each lane + blend.
915     { ISD::SRA,     MVT::v4i64,     12 }, // Shift each lane + blend + split.
916 
917     { ISD::FNEG,    MVT::v4f64,      2 }, // BTVER2 from http://www.agner.org/
918     { ISD::FNEG,    MVT::v8f32,      2 }, // BTVER2 from http://www.agner.org/
919 
920     { ISD::FMUL,    MVT::f64,        2 }, // BTVER2 from http://www.agner.org/
921     { ISD::FMUL,    MVT::v2f64,      2 }, // BTVER2 from http://www.agner.org/
922     { ISD::FMUL,    MVT::v4f64,      4 }, // BTVER2 from http://www.agner.org/
923 
924     { ISD::FDIV,    MVT::f32,       14 }, // SNB from http://www.agner.org/
925     { ISD::FDIV,    MVT::v4f32,     14 }, // SNB from http://www.agner.org/
926     { ISD::FDIV,    MVT::v8f32,     28 }, // SNB from http://www.agner.org/
927     { ISD::FDIV,    MVT::f64,       22 }, // SNB from http://www.agner.org/
928     { ISD::FDIV,    MVT::v2f64,     22 }, // SNB from http://www.agner.org/
929     { ISD::FDIV,    MVT::v4f64,     44 }, // SNB from http://www.agner.org/
930   };
931 
932   if (ST->hasAVX())
933     if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, LT.second))
934       return LT.first * Entry->Cost;
935 
936   static const CostTblEntry SSE42CostTable[] = {
937     { ISD::FADD, MVT::f64,     1 }, // Nehalem from http://www.agner.org/
938     { ISD::FADD, MVT::f32,     1 }, // Nehalem from http://www.agner.org/
939     { ISD::FADD, MVT::v2f64,   1 }, // Nehalem from http://www.agner.org/
940     { ISD::FADD, MVT::v4f32,   1 }, // Nehalem from http://www.agner.org/
941 
942     { ISD::FSUB, MVT::f64,     1 }, // Nehalem from http://www.agner.org/
943     { ISD::FSUB, MVT::f32 ,    1 }, // Nehalem from http://www.agner.org/
944     { ISD::FSUB, MVT::v2f64,   1 }, // Nehalem from http://www.agner.org/
945     { ISD::FSUB, MVT::v4f32,   1 }, // Nehalem from http://www.agner.org/
946 
947     { ISD::FMUL, MVT::f64,     1 }, // Nehalem from http://www.agner.org/
948     { ISD::FMUL, MVT::f32,     1 }, // Nehalem from http://www.agner.org/
949     { ISD::FMUL, MVT::v2f64,   1 }, // Nehalem from http://www.agner.org/
950     { ISD::FMUL, MVT::v4f32,   1 }, // Nehalem from http://www.agner.org/
951 
952     { ISD::FDIV,  MVT::f32,   14 }, // Nehalem from http://www.agner.org/
953     { ISD::FDIV,  MVT::v4f32, 14 }, // Nehalem from http://www.agner.org/
954     { ISD::FDIV,  MVT::f64,   22 }, // Nehalem from http://www.agner.org/
955     { ISD::FDIV,  MVT::v2f64, 22 }, // Nehalem from http://www.agner.org/
956 
957     { ISD::MUL,   MVT::v2i64,  6 }  // 3*pmuludq/3*shift/2*add
958   };
959 
960   if (ST->hasSSE42())
961     if (const auto *Entry = CostTableLookup(SSE42CostTable, ISD, LT.second))
962       return LT.first * Entry->Cost;
963 
964   static const CostTblEntry SSE41CostTable[] = {
965     { ISD::SHL,  MVT::v16i8,      10 }, // pblendvb sequence.
966     { ISD::SHL,  MVT::v8i16,      11 }, // pblendvb sequence.
967     { ISD::SHL,  MVT::v4i32,       4 }, // pslld/paddd/cvttps2dq/pmulld
968 
969     { ISD::SRL,  MVT::v16i8,      11 }, // pblendvb sequence.
970     { ISD::SRL,  MVT::v8i16,      13 }, // pblendvb sequence.
971     { ISD::SRL,  MVT::v4i32,      16 }, // Shift each lane + blend.
972 
973     { ISD::SRA,  MVT::v16i8,      21 }, // pblendvb sequence.
974     { ISD::SRA,  MVT::v8i16,      13 }, // pblendvb sequence.
975 
976     { ISD::MUL,  MVT::v4i32,       2 }  // pmulld (Nehalem from agner.org)
977   };
978 
979   if (ST->hasSSE41())
980     if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second))
981       return LT.first * Entry->Cost;
982 
983   static const CostTblEntry SSE2CostTable[] = {
984     // We don't correctly identify costs of casts because they are marked as
985     // custom.
986     { ISD::SHL,  MVT::v16i8,      13 }, // cmpgtb sequence.
987     { ISD::SHL,  MVT::v8i16,      25 }, // cmpgtw sequence.
988     { ISD::SHL,  MVT::v4i32,      16 }, // pslld/paddd/cvttps2dq/pmuludq.
989     { ISD::SHL,  MVT::v2i64,       4 }, // splat+shuffle sequence.
990 
991     { ISD::SRL,  MVT::v16i8,      14 }, // cmpgtb sequence.
992     { ISD::SRL,  MVT::v8i16,      16 }, // cmpgtw sequence.
993     { ISD::SRL,  MVT::v4i32,      12 }, // Shift each lane + blend.
994     { ISD::SRL,  MVT::v2i64,       4 }, // splat+shuffle sequence.
995 
996     { ISD::SRA,  MVT::v16i8,      27 }, // unpacked cmpgtb sequence.
997     { ISD::SRA,  MVT::v8i16,      16 }, // cmpgtw sequence.
998     { ISD::SRA,  MVT::v4i32,      12 }, // Shift each lane + blend.
999     { ISD::SRA,  MVT::v2i64,       8 }, // srl/xor/sub splat+shuffle sequence.
1000 
1001     { ISD::MUL,  MVT::v8i16,       1 }, // pmullw
1002     { ISD::MUL,  MVT::v4i32,       6 }, // 3*pmuludq/4*shuffle
1003     { ISD::MUL,  MVT::v2i64,       8 }, // 3*pmuludq/3*shift/2*add
1004 
1005     { ISD::FDIV, MVT::f32,        23 }, // Pentium IV from http://www.agner.org/
1006     { ISD::FDIV, MVT::v4f32,      39 }, // Pentium IV from http://www.agner.org/
1007     { ISD::FDIV, MVT::f64,        38 }, // Pentium IV from http://www.agner.org/
1008     { ISD::FDIV, MVT::v2f64,      69 }, // Pentium IV from http://www.agner.org/
1009 
1010     { ISD::FNEG, MVT::f32,         1 }, // Pentium IV from http://www.agner.org/
1011     { ISD::FNEG, MVT::f64,         1 }, // Pentium IV from http://www.agner.org/
1012     { ISD::FNEG, MVT::v4f32,       1 }, // Pentium IV from http://www.agner.org/
1013     { ISD::FNEG, MVT::v2f64,       1 }, // Pentium IV from http://www.agner.org/
1014 
1015     { ISD::FADD, MVT::f32,         2 }, // Pentium IV from http://www.agner.org/
1016     { ISD::FADD, MVT::f64,         2 }, // Pentium IV from http://www.agner.org/
1017 
1018     { ISD::FSUB, MVT::f32,         2 }, // Pentium IV from http://www.agner.org/
1019     { ISD::FSUB, MVT::f64,         2 }, // Pentium IV from http://www.agner.org/
1020   };
1021 
1022   if (ST->hasSSE2())
1023     if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second))
1024       return LT.first * Entry->Cost;
1025 
1026   static const CostTblEntry SSE1CostTable[] = {
1027     { ISD::FDIV, MVT::f32,   17 }, // Pentium III from http://www.agner.org/
1028     { ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/
1029 
1030     { ISD::FNEG, MVT::f32,    2 }, // Pentium III from http://www.agner.org/
1031     { ISD::FNEG, MVT::v4f32,  2 }, // Pentium III from http://www.agner.org/
1032 
1033     { ISD::FADD, MVT::f32,    1 }, // Pentium III from http://www.agner.org/
1034     { ISD::FADD, MVT::v4f32,  2 }, // Pentium III from http://www.agner.org/
1035 
1036     { ISD::FSUB, MVT::f32,    1 }, // Pentium III from http://www.agner.org/
1037     { ISD::FSUB, MVT::v4f32,  2 }, // Pentium III from http://www.agner.org/
1038   };
1039 
1040   if (ST->hasSSE1())
1041     if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second))
1042       return LT.first * Entry->Cost;
1043 
1044   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
1045     { ISD::ADD,  MVT::i64,    1 }, // Core (Merom) from http://www.agner.org/
1046     { ISD::SUB,  MVT::i64,    1 }, // Core (Merom) from http://www.agner.org/
1047     { ISD::MUL,  MVT::i64,    2 }, // Nehalem from http://www.agner.org/
1048   };
1049 
1050   if (ST->is64Bit())
1051     if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, LT.second))
1052       return LT.first * Entry->Cost;
1053 
1054   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
1055     { ISD::ADD,  MVT::i8,    1 }, // Pentium III from http://www.agner.org/
1056     { ISD::ADD,  MVT::i16,   1 }, // Pentium III from http://www.agner.org/
1057     { ISD::ADD,  MVT::i32,   1 }, // Pentium III from http://www.agner.org/
1058 
1059     { ISD::SUB,  MVT::i8,    1 }, // Pentium III from http://www.agner.org/
1060     { ISD::SUB,  MVT::i16,   1 }, // Pentium III from http://www.agner.org/
1061     { ISD::SUB,  MVT::i32,   1 }, // Pentium III from http://www.agner.org/
1062   };
1063 
1064   if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, LT.second))
1065     return LT.first * Entry->Cost;
1066 
1067   // It is not a good idea to vectorize division. We have to scalarize it and
1068   // in the process we will often end up having to spilling regular
1069   // registers. The overhead of division is going to dominate most kernels
1070   // anyways so try hard to prevent vectorization of division - it is
1071   // generally a bad idea. Assume somewhat arbitrarily that we have to be able
1072   // to hide "20 cycles" for each lane.
1073   if (LT.second.isVector() && (ISD == ISD::SDIV || ISD == ISD::SREM ||
1074                                ISD == ISD::UDIV || ISD == ISD::UREM)) {
1075     InstructionCost ScalarCost = getArithmeticInstrCost(
1076         Opcode, Ty->getScalarType(), CostKind, Op1Info, Op2Info,
1077         TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
1078     return 20 * LT.first * LT.second.getVectorNumElements() * ScalarCost;
1079   }
1080 
1081   // Fallback to the default implementation.
1082   return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info);
1083 }
1084 
1085 InstructionCost X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
1086                                            VectorType *BaseTp,
1087                                            ArrayRef<int> Mask, int Index,
1088                                            VectorType *SubTp,
1089                                            ArrayRef<const Value *> Args) {
1090   // 64-bit packed float vectors (v2f32) are widened to type v4f32.
1091   // 64-bit packed integer vectors (v2i32) are widened to type v4i32.
1092   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, BaseTp);
1093 
1094   Kind = improveShuffleKindFromMask(Kind, Mask);
1095   // Treat Transpose as 2-op shuffles - there's no difference in lowering.
1096   if (Kind == TTI::SK_Transpose)
1097     Kind = TTI::SK_PermuteTwoSrc;
1098 
1099   // For Broadcasts we are splatting the first element from the first input
1100   // register, so only need to reference that input and all the output
1101   // registers are the same.
1102   if (Kind == TTI::SK_Broadcast)
1103     LT.first = 1;
1104 
1105   // Subvector extractions are free if they start at the beginning of a
1106   // vector and cheap if the subvectors are aligned.
1107   if (Kind == TTI::SK_ExtractSubvector && LT.second.isVector()) {
1108     int NumElts = LT.second.getVectorNumElements();
1109     if ((Index % NumElts) == 0)
1110       return 0;
1111     std::pair<InstructionCost, MVT> SubLT =
1112         TLI->getTypeLegalizationCost(DL, SubTp);
1113     if (SubLT.second.isVector()) {
1114       int NumSubElts = SubLT.second.getVectorNumElements();
1115       if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0)
1116         return SubLT.first;
1117       // Handle some cases for widening legalization. For now we only handle
1118       // cases where the original subvector was naturally aligned and evenly
1119       // fit in its legalized subvector type.
1120       // FIXME: Remove some of the alignment restrictions.
1121       // FIXME: We can use permq for 64-bit or larger extracts from 256-bit
1122       // vectors.
1123       int OrigSubElts = cast<FixedVectorType>(SubTp)->getNumElements();
1124       if (NumSubElts > OrigSubElts && (Index % OrigSubElts) == 0 &&
1125           (NumSubElts % OrigSubElts) == 0 &&
1126           LT.second.getVectorElementType() ==
1127               SubLT.second.getVectorElementType() &&
1128           LT.second.getVectorElementType().getSizeInBits() ==
1129               BaseTp->getElementType()->getPrimitiveSizeInBits()) {
1130         assert(NumElts >= NumSubElts && NumElts > OrigSubElts &&
1131                "Unexpected number of elements!");
1132         auto *VecTy = FixedVectorType::get(BaseTp->getElementType(),
1133                                            LT.second.getVectorNumElements());
1134         auto *SubTy = FixedVectorType::get(BaseTp->getElementType(),
1135                                            SubLT.second.getVectorNumElements());
1136         int ExtractIndex = alignDown((Index % NumElts), NumSubElts);
1137         InstructionCost ExtractCost = getShuffleCost(
1138             TTI::SK_ExtractSubvector, VecTy, None, ExtractIndex, SubTy);
1139 
1140         // If the original size is 32-bits or more, we can use pshufd. Otherwise
1141         // if we have SSSE3 we can use pshufb.
1142         if (SubTp->getPrimitiveSizeInBits() >= 32 || ST->hasSSSE3())
1143           return ExtractCost + 1; // pshufd or pshufb
1144 
1145         assert(SubTp->getPrimitiveSizeInBits() == 16 &&
1146                "Unexpected vector size");
1147 
1148         return ExtractCost + 2; // worst case pshufhw + pshufd
1149       }
1150     }
1151   }
1152 
1153   // Subvector insertions are cheap if the subvectors are aligned.
1154   // Note that in general, the insertion starting at the beginning of a vector
1155   // isn't free, because we need to preserve the rest of the wide vector.
1156   if (Kind == TTI::SK_InsertSubvector && LT.second.isVector()) {
1157     int NumElts = LT.second.getVectorNumElements();
1158     std::pair<InstructionCost, MVT> SubLT =
1159         TLI->getTypeLegalizationCost(DL, SubTp);
1160     if (SubLT.second.isVector()) {
1161       int NumSubElts = SubLT.second.getVectorNumElements();
1162       if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0)
1163         return SubLT.first;
1164     }
1165 
1166     // If the insertion isn't aligned, treat it like a 2-op shuffle.
1167     Kind = TTI::SK_PermuteTwoSrc;
1168   }
1169 
1170   // Handle some common (illegal) sub-vector types as they are often very cheap
1171   // to shuffle even on targets without PSHUFB.
1172   EVT VT = TLI->getValueType(DL, BaseTp);
1173   if (VT.isSimple() && VT.isVector() && VT.getSizeInBits() < 128 &&
1174       !ST->hasSSSE3()) {
1175      static const CostTblEntry SSE2SubVectorShuffleTbl[] = {
1176       {TTI::SK_Broadcast,        MVT::v4i16, 1}, // pshuflw
1177       {TTI::SK_Broadcast,        MVT::v2i16, 1}, // pshuflw
1178       {TTI::SK_Broadcast,        MVT::v8i8,  2}, // punpck/pshuflw
1179       {TTI::SK_Broadcast,        MVT::v4i8,  2}, // punpck/pshuflw
1180       {TTI::SK_Broadcast,        MVT::v2i8,  1}, // punpck
1181 
1182       {TTI::SK_Reverse,          MVT::v4i16, 1}, // pshuflw
1183       {TTI::SK_Reverse,          MVT::v2i16, 1}, // pshuflw
1184       {TTI::SK_Reverse,          MVT::v4i8,  3}, // punpck/pshuflw/packus
1185       {TTI::SK_Reverse,          MVT::v2i8,  1}, // punpck
1186 
1187       {TTI::SK_PermuteTwoSrc,    MVT::v4i16, 2}, // punpck/pshuflw
1188       {TTI::SK_PermuteTwoSrc,    MVT::v2i16, 2}, // punpck/pshuflw
1189       {TTI::SK_PermuteTwoSrc,    MVT::v8i8,  7}, // punpck/pshuflw
1190       {TTI::SK_PermuteTwoSrc,    MVT::v4i8,  4}, // punpck/pshuflw
1191       {TTI::SK_PermuteTwoSrc,    MVT::v2i8,  2}, // punpck
1192 
1193       {TTI::SK_PermuteSingleSrc, MVT::v4i16, 1}, // pshuflw
1194       {TTI::SK_PermuteSingleSrc, MVT::v2i16, 1}, // pshuflw
1195       {TTI::SK_PermuteSingleSrc, MVT::v8i8,  5}, // punpck/pshuflw
1196       {TTI::SK_PermuteSingleSrc, MVT::v4i8,  3}, // punpck/pshuflw
1197       {TTI::SK_PermuteSingleSrc, MVT::v2i8,  1}, // punpck
1198     };
1199 
1200     if (ST->hasSSE2())
1201       if (const auto *Entry =
1202               CostTableLookup(SSE2SubVectorShuffleTbl, Kind, VT.getSimpleVT()))
1203         return Entry->Cost;
1204   }
1205 
1206   // We are going to permute multiple sources and the result will be in multiple
1207   // destinations. Providing an accurate cost only for splits where the element
1208   // type remains the same.
1209   if (Kind == TTI::SK_PermuteSingleSrc && LT.first != 1) {
1210     MVT LegalVT = LT.second;
1211     if (LegalVT.isVector() &&
1212         LegalVT.getVectorElementType().getSizeInBits() ==
1213             BaseTp->getElementType()->getPrimitiveSizeInBits() &&
1214         LegalVT.getVectorNumElements() <
1215             cast<FixedVectorType>(BaseTp)->getNumElements()) {
1216 
1217       unsigned VecTySize = DL.getTypeStoreSize(BaseTp);
1218       unsigned LegalVTSize = LegalVT.getStoreSize();
1219       // Number of source vectors after legalization:
1220       unsigned NumOfSrcs = (VecTySize + LegalVTSize - 1) / LegalVTSize;
1221       // Number of destination vectors after legalization:
1222       InstructionCost NumOfDests = LT.first;
1223 
1224       auto *SingleOpTy = FixedVectorType::get(BaseTp->getElementType(),
1225                                               LegalVT.getVectorNumElements());
1226 
1227       if (!Mask.empty() && NumOfDests.isValid()) {
1228         // Try to perform better estimation of the permutation.
1229         // 1. Split the source/destination vectors into real registers.
1230         // 2. Do the mask analysis to identify which real registers are
1231         // permuted. If more than 1 source registers are used for the
1232         // destination register building, the cost for this destination register
1233         // is (Number_of_source_register - 1) * Cost_PermuteTwoSrc. If only one
1234         // source register is used, build mask and calculate the cost as a cost
1235         // of PermuteSingleSrc.
1236         // Also, for the single register permute we try to identify if the
1237         // destination register is just a copy of the source register or the
1238         // copy of the previous destination register (the cost is
1239         // TTI::TCC_Basic). If the source register is just reused, the cost for
1240         // this operation is 0.
1241         unsigned E = *NumOfDests.getValue();
1242         unsigned NormalizedVF =
1243             LegalVT.getVectorNumElements() * std::max(NumOfSrcs, E);
1244         unsigned NumOfSrcRegs = NormalizedVF / LegalVT.getVectorNumElements();
1245         unsigned NumOfDestRegs = NormalizedVF / LegalVT.getVectorNumElements();
1246         SmallVector<int> NormalizedMask(NormalizedVF, UndefMaskElem);
1247         copy(Mask, NormalizedMask.begin());
1248         unsigned PrevSrcReg = 0;
1249         ArrayRef<int> PrevRegMask;
1250         InstructionCost Cost = 0;
1251         processShuffleMasks(
1252             NormalizedMask, NumOfSrcRegs, NumOfDestRegs, NumOfDestRegs, []() {},
1253             [this, SingleOpTy, &PrevSrcReg, &PrevRegMask,
1254              &Cost](ArrayRef<int> RegMask, unsigned SrcReg, unsigned DestReg) {
1255               if (!ShuffleVectorInst::isIdentityMask(RegMask)) {
1256                 // Check if the previous register can be just copied to the next
1257                 // one.
1258                 if (PrevRegMask.empty() || PrevSrcReg != SrcReg ||
1259                     PrevRegMask != RegMask)
1260                   Cost += getShuffleCost(TTI::SK_PermuteSingleSrc, SingleOpTy,
1261                                          RegMask, 0, nullptr);
1262                 else
1263                   // Just a copy of previous destination register.
1264                   Cost += TTI::TCC_Basic;
1265                 return;
1266               }
1267               if (SrcReg != DestReg &&
1268                   any_of(RegMask, [](int I) { return I != UndefMaskElem; })) {
1269                 // Just a copy of the source register.
1270                 Cost += TTI::TCC_Basic;
1271               }
1272               PrevSrcReg = SrcReg;
1273               PrevRegMask = RegMask;
1274             },
1275             [this, SingleOpTy, &Cost](ArrayRef<int> RegMask,
1276                                       unsigned /*Unused*/,
1277                                       unsigned /*Unused*/) {
1278               Cost += getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy, RegMask,
1279                                      0, nullptr);
1280             });
1281         return Cost;
1282       }
1283 
1284       InstructionCost NumOfShuffles = (NumOfSrcs - 1) * NumOfDests;
1285       return NumOfShuffles * getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy,
1286                                             None, 0, nullptr);
1287     }
1288 
1289     return BaseT::getShuffleCost(Kind, BaseTp, Mask, Index, SubTp);
1290   }
1291 
1292   // For 2-input shuffles, we must account for splitting the 2 inputs into many.
1293   if (Kind == TTI::SK_PermuteTwoSrc && LT.first != 1) {
1294     // We assume that source and destination have the same vector type.
1295     InstructionCost NumOfDests = LT.first;
1296     InstructionCost NumOfShufflesPerDest = LT.first * 2 - 1;
1297     LT.first = NumOfDests * NumOfShufflesPerDest;
1298   }
1299 
1300   static const CostTblEntry AVX512FP16ShuffleTbl[] = {
1301       {TTI::SK_Broadcast, MVT::v32f16, 1}, // vpbroadcastw
1302       {TTI::SK_Broadcast, MVT::v16f16, 1}, // vpbroadcastw
1303       {TTI::SK_Broadcast, MVT::v8f16, 1},  // vpbroadcastw
1304 
1305       {TTI::SK_Reverse, MVT::v32f16, 2}, // vpermw
1306       {TTI::SK_Reverse, MVT::v16f16, 2}, // vpermw
1307       {TTI::SK_Reverse, MVT::v8f16, 1},  // vpshufb
1308 
1309       {TTI::SK_PermuteSingleSrc, MVT::v32f16, 2}, // vpermw
1310       {TTI::SK_PermuteSingleSrc, MVT::v16f16, 2}, // vpermw
1311       {TTI::SK_PermuteSingleSrc, MVT::v8f16, 1},  // vpshufb
1312 
1313       {TTI::SK_PermuteTwoSrc, MVT::v32f16, 2}, // vpermt2w
1314       {TTI::SK_PermuteTwoSrc, MVT::v16f16, 2}, // vpermt2w
1315       {TTI::SK_PermuteTwoSrc, MVT::v8f16, 2}   // vpermt2w
1316   };
1317 
1318   if (!ST->useSoftFloat() && ST->hasFP16())
1319     if (const auto *Entry =
1320             CostTableLookup(AVX512FP16ShuffleTbl, Kind, LT.second))
1321       return LT.first * Entry->Cost;
1322 
1323   static const CostTblEntry AVX512VBMIShuffleTbl[] = {
1324       {TTI::SK_Reverse, MVT::v64i8, 1}, // vpermb
1325       {TTI::SK_Reverse, MVT::v32i8, 1}, // vpermb
1326 
1327       {TTI::SK_PermuteSingleSrc, MVT::v64i8, 1}, // vpermb
1328       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 1}, // vpermb
1329 
1330       {TTI::SK_PermuteTwoSrc, MVT::v64i8, 2}, // vpermt2b
1331       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 2}, // vpermt2b
1332       {TTI::SK_PermuteTwoSrc, MVT::v16i8, 2}  // vpermt2b
1333   };
1334 
1335   if (ST->hasVBMI())
1336     if (const auto *Entry =
1337             CostTableLookup(AVX512VBMIShuffleTbl, Kind, LT.second))
1338       return LT.first * Entry->Cost;
1339 
1340   static const CostTblEntry AVX512BWShuffleTbl[] = {
1341       {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw
1342       {TTI::SK_Broadcast, MVT::v64i8, 1},  // vpbroadcastb
1343 
1344       {TTI::SK_Reverse, MVT::v32i16, 2}, // vpermw
1345       {TTI::SK_Reverse, MVT::v16i16, 2}, // vpermw
1346       {TTI::SK_Reverse, MVT::v64i8, 2},  // pshufb + vshufi64x2
1347 
1348       {TTI::SK_PermuteSingleSrc, MVT::v32i16, 2}, // vpermw
1349       {TTI::SK_PermuteSingleSrc, MVT::v16i16, 2}, // vpermw
1350       {TTI::SK_PermuteSingleSrc, MVT::v64i8, 8},  // extend to v32i16
1351 
1352       {TTI::SK_PermuteTwoSrc, MVT::v32i16, 2}, // vpermt2w
1353       {TTI::SK_PermuteTwoSrc, MVT::v16i16, 2}, // vpermt2w
1354       {TTI::SK_PermuteTwoSrc, MVT::v8i16, 2},  // vpermt2w
1355       {TTI::SK_PermuteTwoSrc, MVT::v64i8, 19}, // 6 * v32i8 + 1
1356 
1357       {TTI::SK_Select, MVT::v32i16, 1}, // vblendmw
1358       {TTI::SK_Select, MVT::v64i8,  1}, // vblendmb
1359   };
1360 
1361   if (ST->hasBWI())
1362     if (const auto *Entry =
1363             CostTableLookup(AVX512BWShuffleTbl, Kind, LT.second))
1364       return LT.first * Entry->Cost;
1365 
1366   static const CostTblEntry AVX512ShuffleTbl[] = {
1367       {TTI::SK_Broadcast, MVT::v8f64, 1},  // vbroadcastpd
1368       {TTI::SK_Broadcast, MVT::v16f32, 1}, // vbroadcastps
1369       {TTI::SK_Broadcast, MVT::v8i64, 1},  // vpbroadcastq
1370       {TTI::SK_Broadcast, MVT::v16i32, 1}, // vpbroadcastd
1371       {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw
1372       {TTI::SK_Broadcast, MVT::v64i8, 1},  // vpbroadcastb
1373 
1374       {TTI::SK_Reverse, MVT::v8f64, 1},  // vpermpd
1375       {TTI::SK_Reverse, MVT::v16f32, 1}, // vpermps
1376       {TTI::SK_Reverse, MVT::v8i64, 1},  // vpermq
1377       {TTI::SK_Reverse, MVT::v16i32, 1}, // vpermd
1378       {TTI::SK_Reverse, MVT::v32i16, 7}, // per mca
1379       {TTI::SK_Reverse, MVT::v64i8,  7}, // per mca
1380 
1381       {TTI::SK_PermuteSingleSrc, MVT::v8f64, 1},  // vpermpd
1382       {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1},  // vpermpd
1383       {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1},  // vpermpd
1384       {TTI::SK_PermuteSingleSrc, MVT::v16f32, 1}, // vpermps
1385       {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1},  // vpermps
1386       {TTI::SK_PermuteSingleSrc, MVT::v4f32, 1},  // vpermps
1387       {TTI::SK_PermuteSingleSrc, MVT::v8i64, 1},  // vpermq
1388       {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1},  // vpermq
1389       {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1},  // vpermq
1390       {TTI::SK_PermuteSingleSrc, MVT::v16i32, 1}, // vpermd
1391       {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1},  // vpermd
1392       {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1},  // vpermd
1393       {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1},  // pshufb
1394 
1395       {TTI::SK_PermuteTwoSrc, MVT::v8f64, 1},  // vpermt2pd
1396       {TTI::SK_PermuteTwoSrc, MVT::v16f32, 1}, // vpermt2ps
1397       {TTI::SK_PermuteTwoSrc, MVT::v8i64, 1},  // vpermt2q
1398       {TTI::SK_PermuteTwoSrc, MVT::v16i32, 1}, // vpermt2d
1399       {TTI::SK_PermuteTwoSrc, MVT::v4f64, 1},  // vpermt2pd
1400       {TTI::SK_PermuteTwoSrc, MVT::v8f32, 1},  // vpermt2ps
1401       {TTI::SK_PermuteTwoSrc, MVT::v4i64, 1},  // vpermt2q
1402       {TTI::SK_PermuteTwoSrc, MVT::v8i32, 1},  // vpermt2d
1403       {TTI::SK_PermuteTwoSrc, MVT::v2f64, 1},  // vpermt2pd
1404       {TTI::SK_PermuteTwoSrc, MVT::v4f32, 1},  // vpermt2ps
1405       {TTI::SK_PermuteTwoSrc, MVT::v2i64, 1},  // vpermt2q
1406       {TTI::SK_PermuteTwoSrc, MVT::v4i32, 1},  // vpermt2d
1407 
1408       // FIXME: This just applies the type legalization cost rules above
1409       // assuming these completely split.
1410       {TTI::SK_PermuteSingleSrc, MVT::v32i16, 14},
1411       {TTI::SK_PermuteSingleSrc, MVT::v64i8,  14},
1412       {TTI::SK_PermuteTwoSrc,    MVT::v32i16, 42},
1413       {TTI::SK_PermuteTwoSrc,    MVT::v64i8,  42},
1414 
1415       {TTI::SK_Select, MVT::v32i16, 1}, // vpternlogq
1416       {TTI::SK_Select, MVT::v64i8,  1}, // vpternlogq
1417       {TTI::SK_Select, MVT::v8f64,  1}, // vblendmpd
1418       {TTI::SK_Select, MVT::v16f32, 1}, // vblendmps
1419       {TTI::SK_Select, MVT::v8i64,  1}, // vblendmq
1420       {TTI::SK_Select, MVT::v16i32, 1}, // vblendmd
1421   };
1422 
1423   if (ST->hasAVX512())
1424     if (const auto *Entry = CostTableLookup(AVX512ShuffleTbl, Kind, LT.second))
1425       return LT.first * Entry->Cost;
1426 
1427   static const CostTblEntry AVX2ShuffleTbl[] = {
1428       {TTI::SK_Broadcast, MVT::v4f64, 1},  // vbroadcastpd
1429       {TTI::SK_Broadcast, MVT::v8f32, 1},  // vbroadcastps
1430       {TTI::SK_Broadcast, MVT::v4i64, 1},  // vpbroadcastq
1431       {TTI::SK_Broadcast, MVT::v8i32, 1},  // vpbroadcastd
1432       {TTI::SK_Broadcast, MVT::v16i16, 1}, // vpbroadcastw
1433       {TTI::SK_Broadcast, MVT::v32i8, 1},  // vpbroadcastb
1434 
1435       {TTI::SK_Reverse, MVT::v4f64, 1},  // vpermpd
1436       {TTI::SK_Reverse, MVT::v8f32, 1},  // vpermps
1437       {TTI::SK_Reverse, MVT::v4i64, 1},  // vpermq
1438       {TTI::SK_Reverse, MVT::v8i32, 1},  // vpermd
1439       {TTI::SK_Reverse, MVT::v16i16, 2}, // vperm2i128 + pshufb
1440       {TTI::SK_Reverse, MVT::v32i8, 2},  // vperm2i128 + pshufb
1441 
1442       {TTI::SK_Select, MVT::v16i16, 1}, // vpblendvb
1443       {TTI::SK_Select, MVT::v32i8, 1},  // vpblendvb
1444 
1445       {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1},  // vpermpd
1446       {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1},  // vpermps
1447       {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1},  // vpermq
1448       {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1},  // vpermd
1449       {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vperm2i128 + 2*vpshufb
1450                                                   // + vpblendvb
1451       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4},  // vperm2i128 + 2*vpshufb
1452                                                   // + vpblendvb
1453 
1454       {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3},  // 2*vpermpd + vblendpd
1455       {TTI::SK_PermuteTwoSrc, MVT::v8f32, 3},  // 2*vpermps + vblendps
1456       {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3},  // 2*vpermq + vpblendd
1457       {TTI::SK_PermuteTwoSrc, MVT::v8i32, 3},  // 2*vpermd + vpblendd
1458       {TTI::SK_PermuteTwoSrc, MVT::v16i16, 7}, // 2*vperm2i128 + 4*vpshufb
1459                                                // + vpblendvb
1460       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 7},  // 2*vperm2i128 + 4*vpshufb
1461                                                // + vpblendvb
1462   };
1463 
1464   if (ST->hasAVX2())
1465     if (const auto *Entry = CostTableLookup(AVX2ShuffleTbl, Kind, LT.second))
1466       return LT.first * Entry->Cost;
1467 
1468   static const CostTblEntry XOPShuffleTbl[] = {
1469       {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2},  // vperm2f128 + vpermil2pd
1470       {TTI::SK_PermuteSingleSrc, MVT::v8f32, 2},  // vperm2f128 + vpermil2ps
1471       {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2},  // vperm2f128 + vpermil2pd
1472       {TTI::SK_PermuteSingleSrc, MVT::v8i32, 2},  // vperm2f128 + vpermil2ps
1473       {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vextractf128 + 2*vpperm
1474                                                   // + vinsertf128
1475       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4},  // vextractf128 + 2*vpperm
1476                                                   // + vinsertf128
1477 
1478       {TTI::SK_PermuteTwoSrc, MVT::v16i16, 9}, // 2*vextractf128 + 6*vpperm
1479                                                // + vinsertf128
1480       {TTI::SK_PermuteTwoSrc, MVT::v8i16, 1},  // vpperm
1481       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 9},  // 2*vextractf128 + 6*vpperm
1482                                                // + vinsertf128
1483       {TTI::SK_PermuteTwoSrc, MVT::v16i8, 1},  // vpperm
1484   };
1485 
1486   if (ST->hasXOP())
1487     if (const auto *Entry = CostTableLookup(XOPShuffleTbl, Kind, LT.second))
1488       return LT.first * Entry->Cost;
1489 
1490   static const CostTblEntry AVX1ShuffleTbl[] = {
1491       {TTI::SK_Broadcast, MVT::v4f64, 2},  // vperm2f128 + vpermilpd
1492       {TTI::SK_Broadcast, MVT::v8f32, 2},  // vperm2f128 + vpermilps
1493       {TTI::SK_Broadcast, MVT::v4i64, 2},  // vperm2f128 + vpermilpd
1494       {TTI::SK_Broadcast, MVT::v8i32, 2},  // vperm2f128 + vpermilps
1495       {TTI::SK_Broadcast, MVT::v16i16, 3}, // vpshuflw + vpshufd + vinsertf128
1496       {TTI::SK_Broadcast, MVT::v32i8, 2},  // vpshufb + vinsertf128
1497 
1498       {TTI::SK_Reverse, MVT::v4f64, 2},  // vperm2f128 + vpermilpd
1499       {TTI::SK_Reverse, MVT::v8f32, 2},  // vperm2f128 + vpermilps
1500       {TTI::SK_Reverse, MVT::v4i64, 2},  // vperm2f128 + vpermilpd
1501       {TTI::SK_Reverse, MVT::v8i32, 2},  // vperm2f128 + vpermilps
1502       {TTI::SK_Reverse, MVT::v16i16, 4}, // vextractf128 + 2*pshufb
1503                                          // + vinsertf128
1504       {TTI::SK_Reverse, MVT::v32i8, 4},  // vextractf128 + 2*pshufb
1505                                          // + vinsertf128
1506 
1507       {TTI::SK_Select, MVT::v4i64, 1},  // vblendpd
1508       {TTI::SK_Select, MVT::v4f64, 1},  // vblendpd
1509       {TTI::SK_Select, MVT::v8i32, 1},  // vblendps
1510       {TTI::SK_Select, MVT::v8f32, 1},  // vblendps
1511       {TTI::SK_Select, MVT::v16i16, 3}, // vpand + vpandn + vpor
1512       {TTI::SK_Select, MVT::v32i8, 3},  // vpand + vpandn + vpor
1513 
1514       {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2},  // vperm2f128 + vshufpd
1515       {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2},  // vperm2f128 + vshufpd
1516       {TTI::SK_PermuteSingleSrc, MVT::v8f32, 4},  // 2*vperm2f128 + 2*vshufps
1517       {TTI::SK_PermuteSingleSrc, MVT::v8i32, 4},  // 2*vperm2f128 + 2*vshufps
1518       {TTI::SK_PermuteSingleSrc, MVT::v16i16, 8}, // vextractf128 + 4*pshufb
1519                                                   // + 2*por + vinsertf128
1520       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 8},  // vextractf128 + 4*pshufb
1521                                                   // + 2*por + vinsertf128
1522 
1523       {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3},   // 2*vperm2f128 + vshufpd
1524       {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3},   // 2*vperm2f128 + vshufpd
1525       {TTI::SK_PermuteTwoSrc, MVT::v8f32, 4},   // 2*vperm2f128 + 2*vshufps
1526       {TTI::SK_PermuteTwoSrc, MVT::v8i32, 4},   // 2*vperm2f128 + 2*vshufps
1527       {TTI::SK_PermuteTwoSrc, MVT::v16i16, 15}, // 2*vextractf128 + 8*pshufb
1528                                                 // + 4*por + vinsertf128
1529       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 15},  // 2*vextractf128 + 8*pshufb
1530                                                 // + 4*por + vinsertf128
1531   };
1532 
1533   if (ST->hasAVX())
1534     if (const auto *Entry = CostTableLookup(AVX1ShuffleTbl, Kind, LT.second))
1535       return LT.first * Entry->Cost;
1536 
1537   static const CostTblEntry SSE41ShuffleTbl[] = {
1538       {TTI::SK_Select, MVT::v2i64, 1}, // pblendw
1539       {TTI::SK_Select, MVT::v2f64, 1}, // movsd
1540       {TTI::SK_Select, MVT::v4i32, 1}, // pblendw
1541       {TTI::SK_Select, MVT::v4f32, 1}, // blendps
1542       {TTI::SK_Select, MVT::v8i16, 1}, // pblendw
1543       {TTI::SK_Select, MVT::v16i8, 1}  // pblendvb
1544   };
1545 
1546   if (ST->hasSSE41())
1547     if (const auto *Entry = CostTableLookup(SSE41ShuffleTbl, Kind, LT.second))
1548       return LT.first * Entry->Cost;
1549 
1550   static const CostTblEntry SSSE3ShuffleTbl[] = {
1551       {TTI::SK_Broadcast, MVT::v8i16, 1}, // pshufb
1552       {TTI::SK_Broadcast, MVT::v16i8, 1}, // pshufb
1553 
1554       {TTI::SK_Reverse, MVT::v8i16, 1}, // pshufb
1555       {TTI::SK_Reverse, MVT::v16i8, 1}, // pshufb
1556 
1557       {TTI::SK_Select, MVT::v8i16, 3}, // 2*pshufb + por
1558       {TTI::SK_Select, MVT::v16i8, 3}, // 2*pshufb + por
1559 
1560       {TTI::SK_PermuteSingleSrc, MVT::v8i16, 1}, // pshufb
1561       {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb
1562 
1563       {TTI::SK_PermuteTwoSrc, MVT::v8i16, 3}, // 2*pshufb + por
1564       {TTI::SK_PermuteTwoSrc, MVT::v16i8, 3}, // 2*pshufb + por
1565   };
1566 
1567   if (ST->hasSSSE3())
1568     if (const auto *Entry = CostTableLookup(SSSE3ShuffleTbl, Kind, LT.second))
1569       return LT.first * Entry->Cost;
1570 
1571   static const CostTblEntry SSE2ShuffleTbl[] = {
1572       {TTI::SK_Broadcast, MVT::v2f64, 1}, // shufpd
1573       {TTI::SK_Broadcast, MVT::v2i64, 1}, // pshufd
1574       {TTI::SK_Broadcast, MVT::v4i32, 1}, // pshufd
1575       {TTI::SK_Broadcast, MVT::v8i16, 2}, // pshuflw + pshufd
1576       {TTI::SK_Broadcast, MVT::v16i8, 3}, // unpck + pshuflw + pshufd
1577 
1578       {TTI::SK_Reverse, MVT::v2f64, 1}, // shufpd
1579       {TTI::SK_Reverse, MVT::v2i64, 1}, // pshufd
1580       {TTI::SK_Reverse, MVT::v4i32, 1}, // pshufd
1581       {TTI::SK_Reverse, MVT::v8i16, 3}, // pshuflw + pshufhw + pshufd
1582       {TTI::SK_Reverse, MVT::v16i8, 9}, // 2*pshuflw + 2*pshufhw
1583                                         // + 2*pshufd + 2*unpck + packus
1584 
1585       {TTI::SK_Select, MVT::v2i64, 1}, // movsd
1586       {TTI::SK_Select, MVT::v2f64, 1}, // movsd
1587       {TTI::SK_Select, MVT::v4i32, 2}, // 2*shufps
1588       {TTI::SK_Select, MVT::v8i16, 3}, // pand + pandn + por
1589       {TTI::SK_Select, MVT::v16i8, 3}, // pand + pandn + por
1590 
1591       {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // shufpd
1592       {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // pshufd
1593       {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // pshufd
1594       {TTI::SK_PermuteSingleSrc, MVT::v8i16, 5}, // 2*pshuflw + 2*pshufhw
1595                                                   // + pshufd/unpck
1596     { TTI::SK_PermuteSingleSrc, MVT::v16i8, 10 }, // 2*pshuflw + 2*pshufhw
1597                                                   // + 2*pshufd + 2*unpck + 2*packus
1598 
1599     { TTI::SK_PermuteTwoSrc,    MVT::v2f64,  1 }, // shufpd
1600     { TTI::SK_PermuteTwoSrc,    MVT::v2i64,  1 }, // shufpd
1601     { TTI::SK_PermuteTwoSrc,    MVT::v4i32,  2 }, // 2*{unpck,movsd,pshufd}
1602     { TTI::SK_PermuteTwoSrc,    MVT::v8i16,  8 }, // blend+permute
1603     { TTI::SK_PermuteTwoSrc,    MVT::v16i8, 13 }, // blend+permute
1604   };
1605 
1606   static const CostTblEntry SSE3BroadcastLoadTbl[] = {
1607       {TTI::SK_Broadcast, MVT::v2f64, 0}, // broadcast handled by movddup
1608   };
1609 
1610   if (ST->hasSSE2()) {
1611     bool IsLoad =
1612         llvm::any_of(Args, [](const auto &V) { return isa<LoadInst>(V); });
1613     if (ST->hasSSE3() && IsLoad)
1614       if (const auto *Entry =
1615               CostTableLookup(SSE3BroadcastLoadTbl, Kind, LT.second)) {
1616         assert(isLegalBroadcastLoad(BaseTp->getElementType(),
1617                                     LT.second.getVectorElementCount()) &&
1618                "Table entry missing from isLegalBroadcastLoad()");
1619         return LT.first * Entry->Cost;
1620       }
1621 
1622     if (const auto *Entry = CostTableLookup(SSE2ShuffleTbl, Kind, LT.second))
1623       return LT.first * Entry->Cost;
1624   }
1625 
1626   static const CostTblEntry SSE1ShuffleTbl[] = {
1627     { TTI::SK_Broadcast,        MVT::v4f32, 1 }, // shufps
1628     { TTI::SK_Reverse,          MVT::v4f32, 1 }, // shufps
1629     { TTI::SK_Select,           MVT::v4f32, 2 }, // 2*shufps
1630     { TTI::SK_PermuteSingleSrc, MVT::v4f32, 1 }, // shufps
1631     { TTI::SK_PermuteTwoSrc,    MVT::v4f32, 2 }, // 2*shufps
1632   };
1633 
1634   if (ST->hasSSE1())
1635     if (const auto *Entry = CostTableLookup(SSE1ShuffleTbl, Kind, LT.second))
1636       return LT.first * Entry->Cost;
1637 
1638   return BaseT::getShuffleCost(Kind, BaseTp, Mask, Index, SubTp);
1639 }
1640 
1641 InstructionCost X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
1642                                              Type *Src,
1643                                              TTI::CastContextHint CCH,
1644                                              TTI::TargetCostKind CostKind,
1645                                              const Instruction *I) {
1646   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1647   assert(ISD && "Invalid opcode");
1648 
1649   // TODO: Allow non-throughput costs that aren't binary.
1650   auto AdjustCost = [&CostKind](InstructionCost Cost) -> InstructionCost {
1651     if (CostKind != TTI::TCK_RecipThroughput)
1652       return Cost == 0 ? 0 : 1;
1653     return Cost;
1654   };
1655 
1656   // The cost tables include both specific, custom (non-legal) src/dst type
1657   // conversions and generic, legalized types. We test for customs first, before
1658   // falling back to legalization.
1659   // FIXME: Need a better design of the cost table to handle non-simple types of
1660   // potential massive combinations (elem_num x src_type x dst_type).
1661   static const TypeConversionCostTblEntry AVX512BWConversionTbl[] {
1662     { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8, 1 },
1663     { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8, 1 },
1664 
1665     // Mask sign extend has an instruction.
1666     { ISD::SIGN_EXTEND, MVT::v2i8,   MVT::v2i1,   1 },
1667     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v2i1,   1 },
1668     { ISD::SIGN_EXTEND, MVT::v2i16,  MVT::v2i1,   1 },
1669     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v2i1,   1 },
1670     { ISD::SIGN_EXTEND, MVT::v4i8,   MVT::v4i1,   1 },
1671     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v4i1,   1 },
1672     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i1,   1 },
1673     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v4i1,   1 },
1674     { ISD::SIGN_EXTEND, MVT::v8i8,   MVT::v8i1,   1 },
1675     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v8i1,   1 },
1676     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i1,   1 },
1677     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v16i1,  1 },
1678     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1,  1 },
1679     { ISD::SIGN_EXTEND, MVT::v32i8,  MVT::v32i1,  1 },
1680     { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i1,  1 },
1681     { ISD::SIGN_EXTEND, MVT::v64i8,  MVT::v64i1,  1 },
1682     { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v64i1,  1 },
1683 
1684     // Mask zero extend is a sext + shift.
1685     { ISD::ZERO_EXTEND, MVT::v2i8,   MVT::v2i1,   2 },
1686     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v2i1,   2 },
1687     { ISD::ZERO_EXTEND, MVT::v2i16,  MVT::v2i1,   2 },
1688     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v2i1,   2 },
1689     { ISD::ZERO_EXTEND, MVT::v4i8,   MVT::v4i1,   2 },
1690     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v4i1,   2 },
1691     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i1,   2 },
1692     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v4i1,   2 },
1693     { ISD::ZERO_EXTEND, MVT::v8i8,   MVT::v8i1,   2 },
1694     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v8i1,   2 },
1695     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i1,   2 },
1696     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v16i1,  2 },
1697     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1,  2 },
1698     { ISD::ZERO_EXTEND, MVT::v32i8,  MVT::v32i1,  2 },
1699     { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i1,  2 },
1700     { ISD::ZERO_EXTEND, MVT::v64i8,  MVT::v64i1,  2 },
1701     { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v64i1,  2 },
1702 
1703     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i8,   2 },
1704     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v16i8,  2 },
1705     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i16,  2 },
1706     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v8i16,  2 },
1707     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i8,   2 },
1708     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v16i8,  2 },
1709     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i16,  2 },
1710     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v8i16,  2 },
1711     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i8,   2 },
1712     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v16i8,  2 },
1713     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i16,  2 },
1714     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i8,  2 },
1715     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i16, 2 },
1716     { ISD::TRUNCATE,    MVT::v32i1,  MVT::v32i8,  2 },
1717     { ISD::TRUNCATE,    MVT::v32i1,  MVT::v32i16, 2 },
1718     { ISD::TRUNCATE,    MVT::v64i1,  MVT::v64i8,  2 },
1719     { ISD::TRUNCATE,    MVT::v64i1,  MVT::v32i16, 2 },
1720 
1721     { ISD::TRUNCATE,    MVT::v32i8,  MVT::v32i16, 2 },
1722     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 2 }, // widen to zmm
1723     { ISD::TRUNCATE,    MVT::v2i8,   MVT::v2i16,  2 }, // vpmovwb
1724     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  2 }, // vpmovwb
1725     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  2 }, // vpmovwb
1726   };
1727 
1728   static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = {
1729     // Mask sign extend has an instruction.
1730     { ISD::SIGN_EXTEND, MVT::v2i64,  MVT::v2i1,   1 },
1731     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v2i1,   1 },
1732     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i1,   1 },
1733     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   1 },
1734     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   1 },
1735     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v16i1,  1 },
1736     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i1,   1 },
1737     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1,  1 },
1738 
1739     // Mask zero extend is a sext + shift.
1740     { ISD::ZERO_EXTEND, MVT::v2i64,  MVT::v2i1,   2 },
1741     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v2i1,   2 },
1742     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i1,   2 },
1743     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   2 },
1744     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   2 },
1745     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v16i1,  2 },
1746     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i1,   2 },
1747     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
1748 
1749     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i64,  2 },
1750     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v4i32,  2 },
1751     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i32,  2 },
1752     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i64,  2 },
1753     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i32,  2 },
1754     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i64,  2 },
1755     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i32, 2 },
1756     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v8i64,  2 },
1757 
1758     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
1759     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
1760 
1761     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
1762     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
1763 
1764     { ISD::FP_TO_SINT,  MVT::v8i64,  MVT::v8f32,  1 },
1765     { ISD::FP_TO_SINT,  MVT::v8i64,  MVT::v8f64,  1 },
1766 
1767     { ISD::FP_TO_UINT,  MVT::v8i64,  MVT::v8f32,  1 },
1768     { ISD::FP_TO_UINT,  MVT::v8i64,  MVT::v8f64,  1 },
1769   };
1770 
1771   // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and
1772   // 256-bit wide vectors.
1773 
1774   static const TypeConversionCostTblEntry AVX512FConversionTbl[] = {
1775     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v8f32,  1 },
1776     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v16f32, 3 },
1777     { ISD::FP_ROUND,  MVT::v8f32,   MVT::v8f64,  1 },
1778 
1779     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i8,   3 }, // sext+vpslld+vptestmd
1780     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i8,   3 }, // sext+vpslld+vptestmd
1781     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i8,   3 }, // sext+vpslld+vptestmd
1782     { ISD::TRUNCATE,  MVT::v16i1,   MVT::v16i8,  3 }, // sext+vpslld+vptestmd
1783     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i16,  3 }, // sext+vpsllq+vptestmq
1784     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i16,  3 }, // sext+vpsllq+vptestmq
1785     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i16,  3 }, // sext+vpsllq+vptestmq
1786     { ISD::TRUNCATE,  MVT::v16i1,   MVT::v16i16, 3 }, // sext+vpslld+vptestmd
1787     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i32,  2 }, // zmm vpslld+vptestmd
1788     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i32,  2 }, // zmm vpslld+vptestmd
1789     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i32,  2 }, // zmm vpslld+vptestmd
1790     { ISD::TRUNCATE,  MVT::v16i1,   MVT::v16i32, 2 }, // vpslld+vptestmd
1791     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i64,  2 }, // zmm vpsllq+vptestmq
1792     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i64,  2 }, // zmm vpsllq+vptestmq
1793     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i64,  2 }, // vpsllq+vptestmq
1794     { ISD::TRUNCATE,  MVT::v2i8,    MVT::v2i32,  2 }, // vpmovdb
1795     { ISD::TRUNCATE,  MVT::v4i8,    MVT::v4i32,  2 }, // vpmovdb
1796     { ISD::TRUNCATE,  MVT::v16i8,   MVT::v16i32, 2 }, // vpmovdb
1797     { ISD::TRUNCATE,  MVT::v32i8,   MVT::v16i32, 2 }, // vpmovdb
1798     { ISD::TRUNCATE,  MVT::v64i8,   MVT::v16i32, 2 }, // vpmovdb
1799     { ISD::TRUNCATE,  MVT::v16i16,  MVT::v16i32, 2 }, // vpmovdw
1800     { ISD::TRUNCATE,  MVT::v32i16,  MVT::v16i32, 2 }, // vpmovdw
1801     { ISD::TRUNCATE,  MVT::v2i8,    MVT::v2i64,  2 }, // vpmovqb
1802     { ISD::TRUNCATE,  MVT::v2i16,   MVT::v2i64,  1 }, // vpshufb
1803     { ISD::TRUNCATE,  MVT::v8i8,    MVT::v8i64,  2 }, // vpmovqb
1804     { ISD::TRUNCATE,  MVT::v16i8,   MVT::v8i64,  2 }, // vpmovqb
1805     { ISD::TRUNCATE,  MVT::v32i8,   MVT::v8i64,  2 }, // vpmovqb
1806     { ISD::TRUNCATE,  MVT::v64i8,   MVT::v8i64,  2 }, // vpmovqb
1807     { ISD::TRUNCATE,  MVT::v8i16,   MVT::v8i64,  2 }, // vpmovqw
1808     { ISD::TRUNCATE,  MVT::v16i16,  MVT::v8i64,  2 }, // vpmovqw
1809     { ISD::TRUNCATE,  MVT::v32i16,  MVT::v8i64,  2 }, // vpmovqw
1810     { ISD::TRUNCATE,  MVT::v8i32,   MVT::v8i64,  1 }, // vpmovqd
1811     { ISD::TRUNCATE,  MVT::v4i32,   MVT::v4i64,  1 }, // zmm vpmovqd
1812     { ISD::TRUNCATE,  MVT::v16i8,   MVT::v16i64, 5 },// 2*vpmovqd+concat+vpmovdb
1813 
1814     { ISD::TRUNCATE,  MVT::v16i8,  MVT::v16i16,  3 }, // extend to v16i32
1815     { ISD::TRUNCATE,  MVT::v32i8,  MVT::v32i16,  8 },
1816     { ISD::TRUNCATE,  MVT::v64i8,  MVT::v32i16,  8 },
1817 
1818     // Sign extend is zmm vpternlogd+vptruncdb.
1819     // Zero extend is zmm broadcast load+vptruncdw.
1820     { ISD::SIGN_EXTEND, MVT::v2i8,   MVT::v2i1,   3 },
1821     { ISD::ZERO_EXTEND, MVT::v2i8,   MVT::v2i1,   4 },
1822     { ISD::SIGN_EXTEND, MVT::v4i8,   MVT::v4i1,   3 },
1823     { ISD::ZERO_EXTEND, MVT::v4i8,   MVT::v4i1,   4 },
1824     { ISD::SIGN_EXTEND, MVT::v8i8,   MVT::v8i1,   3 },
1825     { ISD::ZERO_EXTEND, MVT::v8i8,   MVT::v8i1,   4 },
1826     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v16i1,  3 },
1827     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v16i1,  4 },
1828 
1829     // Sign extend is zmm vpternlogd+vptruncdw.
1830     // Zero extend is zmm vpternlogd+vptruncdw+vpsrlw.
1831     { ISD::SIGN_EXTEND, MVT::v2i16,  MVT::v2i1,   3 },
1832     { ISD::ZERO_EXTEND, MVT::v2i16,  MVT::v2i1,   4 },
1833     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i1,   3 },
1834     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i1,   4 },
1835     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i1,   3 },
1836     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i1,   4 },
1837     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1,  3 },
1838     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1,  4 },
1839 
1840     { ISD::SIGN_EXTEND, MVT::v2i32,  MVT::v2i1,   1 }, // zmm vpternlogd
1841     { ISD::ZERO_EXTEND, MVT::v2i32,  MVT::v2i1,   2 }, // zmm vpternlogd+psrld
1842     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i1,   1 }, // zmm vpternlogd
1843     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i1,   2 }, // zmm vpternlogd+psrld
1844     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   1 }, // zmm vpternlogd
1845     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   2 }, // zmm vpternlogd+psrld
1846     { ISD::SIGN_EXTEND, MVT::v2i64,  MVT::v2i1,   1 }, // zmm vpternlogq
1847     { ISD::ZERO_EXTEND, MVT::v2i64,  MVT::v2i1,   2 }, // zmm vpternlogq+psrlq
1848     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   1 }, // zmm vpternlogq
1849     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   2 }, // zmm vpternlogq+psrlq
1850 
1851     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1,  1 }, // vpternlogd
1852     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1,  2 }, // vpternlogd+psrld
1853     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i1,   1 }, // vpternlogq
1854     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i1,   2 }, // vpternlogq+psrlq
1855 
1856     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
1857     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
1858     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1859     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1860     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i8,   1 },
1861     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i8,   1 },
1862     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
1863     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
1864     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
1865     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
1866 
1867     { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8,  3 }, // FIXME: May not be right
1868     { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8,  3 }, // FIXME: May not be right
1869 
1870     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
1871     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
1872     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v16i8,  2 },
1873     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i8,  1 },
1874     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
1875     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i16, 1 },
1876     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
1877     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
1878 
1879     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
1880     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
1881     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v16i8,  2 },
1882     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i8,  1 },
1883     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
1884     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i16, 1 },
1885     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
1886     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
1887     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64, 26 },
1888     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  5 },
1889 
1890     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v16f32, 2 },
1891     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v16f64, 7 },
1892     { ISD::FP_TO_SINT,  MVT::v32i8,  MVT::v32f64,15 },
1893     { ISD::FP_TO_SINT,  MVT::v64i8,  MVT::v64f32,11 },
1894     { ISD::FP_TO_SINT,  MVT::v64i8,  MVT::v64f64,31 },
1895     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v8f64,  3 },
1896     { ISD::FP_TO_SINT,  MVT::v16i16, MVT::v16f64, 7 },
1897     { ISD::FP_TO_SINT,  MVT::v32i16, MVT::v32f32, 5 },
1898     { ISD::FP_TO_SINT,  MVT::v32i16, MVT::v32f64,15 },
1899     { ISD::FP_TO_SINT,  MVT::v8i32,  MVT::v8f64,  1 },
1900     { ISD::FP_TO_SINT,  MVT::v16i32, MVT::v16f64, 3 },
1901 
1902     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f64,  1 },
1903     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v8f64,  3 },
1904     { ISD::FP_TO_UINT,  MVT::v8i8,   MVT::v8f64,  3 },
1905     { ISD::FP_TO_UINT,  MVT::v16i32, MVT::v16f32, 1 },
1906     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v16f32, 3 },
1907     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v16f32, 3 },
1908   };
1909 
1910   static const TypeConversionCostTblEntry AVX512BWVLConversionTbl[] {
1911     // Mask sign extend has an instruction.
1912     { ISD::SIGN_EXTEND, MVT::v2i8,   MVT::v2i1,   1 },
1913     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v2i1,   1 },
1914     { ISD::SIGN_EXTEND, MVT::v2i16,  MVT::v2i1,   1 },
1915     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v2i1,   1 },
1916     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i1,   1 },
1917     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v4i1,   1 },
1918     { ISD::SIGN_EXTEND, MVT::v4i8,   MVT::v4i1,   1 },
1919     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v4i1,   1 },
1920     { ISD::SIGN_EXTEND, MVT::v8i8,   MVT::v8i1,   1 },
1921     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v8i1,   1 },
1922     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i1,   1 },
1923     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v16i1,  1 },
1924     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1,  1 },
1925     { ISD::SIGN_EXTEND, MVT::v32i8,  MVT::v32i1,  1 },
1926     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v32i1,  1 },
1927     { ISD::SIGN_EXTEND, MVT::v32i8,  MVT::v64i1,  1 },
1928     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v64i1,  1 },
1929 
1930     // Mask zero extend is a sext + shift.
1931     { ISD::ZERO_EXTEND, MVT::v2i8,   MVT::v2i1,   2 },
1932     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v2i1,   2 },
1933     { ISD::ZERO_EXTEND, MVT::v2i16,  MVT::v2i1,   2 },
1934     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v2i1,   2 },
1935     { ISD::ZERO_EXTEND, MVT::v4i8,   MVT::v4i1,   2 },
1936     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v4i1,   2 },
1937     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i1,   2 },
1938     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v4i1,   2 },
1939     { ISD::ZERO_EXTEND, MVT::v8i8,   MVT::v8i1,   2 },
1940     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v8i1,   2 },
1941     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i1,   2 },
1942     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v16i1,  2 },
1943     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1,  2 },
1944     { ISD::ZERO_EXTEND, MVT::v32i8,  MVT::v32i1,  2 },
1945     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v32i1,  2 },
1946     { ISD::ZERO_EXTEND, MVT::v32i8,  MVT::v64i1,  2 },
1947     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v64i1,  2 },
1948 
1949     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i8,   2 },
1950     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v16i8,  2 },
1951     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i16,  2 },
1952     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v8i16,  2 },
1953     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i8,   2 },
1954     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v16i8,  2 },
1955     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i16,  2 },
1956     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v8i16,  2 },
1957     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i8,   2 },
1958     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v16i8,  2 },
1959     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i16,  2 },
1960     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i8,  2 },
1961     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i16, 2 },
1962     { ISD::TRUNCATE,    MVT::v32i1,  MVT::v32i8,  2 },
1963     { ISD::TRUNCATE,    MVT::v32i1,  MVT::v16i16, 2 },
1964     { ISD::TRUNCATE,    MVT::v64i1,  MVT::v32i8,  2 },
1965     { ISD::TRUNCATE,    MVT::v64i1,  MVT::v16i16, 2 },
1966 
1967     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 2 },
1968   };
1969 
1970   static const TypeConversionCostTblEntry AVX512DQVLConversionTbl[] = {
1971     // Mask sign extend has an instruction.
1972     { ISD::SIGN_EXTEND, MVT::v2i64,  MVT::v2i1,   1 },
1973     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v2i1,   1 },
1974     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i1,   1 },
1975     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v16i1,  1 },
1976     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   1 },
1977     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v8i1,   1 },
1978     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v16i1,  1 },
1979     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   1 },
1980 
1981     // Mask zero extend is a sext + shift.
1982     { ISD::ZERO_EXTEND, MVT::v2i64,  MVT::v2i1,   2 },
1983     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v2i1,   2 },
1984     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i1,   2 },
1985     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v16i1,  2 },
1986     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   2 },
1987     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v8i1,   2 },
1988     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v16i1,  2 },
1989     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   2 },
1990 
1991     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v4i64,  2 },
1992     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v8i32,  2 },
1993     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i64,  2 },
1994     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v4i32,  2 },
1995     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i32,  2 },
1996     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i64,  2 },
1997     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v4i64,  2 },
1998     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i32,  2 },
1999 
2000     { ISD::SINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
2001     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
2002     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
2003     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
2004 
2005     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
2006     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
2007     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
2008     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
2009 
2010     { ISD::FP_TO_SINT,  MVT::v2i64,  MVT::v4f32,  1 },
2011     { ISD::FP_TO_SINT,  MVT::v4i64,  MVT::v4f32,  1 },
2012     { ISD::FP_TO_SINT,  MVT::v2i64,  MVT::v2f64,  1 },
2013     { ISD::FP_TO_SINT,  MVT::v4i64,  MVT::v4f64,  1 },
2014 
2015     { ISD::FP_TO_UINT,  MVT::v2i64,  MVT::v4f32,  1 },
2016     { ISD::FP_TO_UINT,  MVT::v4i64,  MVT::v4f32,  1 },
2017     { ISD::FP_TO_UINT,  MVT::v2i64,  MVT::v2f64,  1 },
2018     { ISD::FP_TO_UINT,  MVT::v4i64,  MVT::v4f64,  1 },
2019   };
2020 
2021   static const TypeConversionCostTblEntry AVX512VLConversionTbl[] = {
2022     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i8,   3 }, // sext+vpslld+vptestmd
2023     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i8,   3 }, // sext+vpslld+vptestmd
2024     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i8,   3 }, // sext+vpslld+vptestmd
2025     { ISD::TRUNCATE,  MVT::v16i1,   MVT::v16i8,  8 }, // split+2*v8i8
2026     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i16,  3 }, // sext+vpsllq+vptestmq
2027     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i16,  3 }, // sext+vpsllq+vptestmq
2028     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i16,  3 }, // sext+vpsllq+vptestmq
2029     { ISD::TRUNCATE,  MVT::v16i1,   MVT::v16i16, 8 }, // split+2*v8i16
2030     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i32,  2 }, // vpslld+vptestmd
2031     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i32,  2 }, // vpslld+vptestmd
2032     { ISD::TRUNCATE,  MVT::v8i1,    MVT::v8i32,  2 }, // vpslld+vptestmd
2033     { ISD::TRUNCATE,  MVT::v2i1,    MVT::v2i64,  2 }, // vpsllq+vptestmq
2034     { ISD::TRUNCATE,  MVT::v4i1,    MVT::v4i64,  2 }, // vpsllq+vptestmq
2035     { ISD::TRUNCATE,  MVT::v4i32,   MVT::v4i64,  1 }, // vpmovqd
2036     { ISD::TRUNCATE,  MVT::v4i8,    MVT::v4i64,  2 }, // vpmovqb
2037     { ISD::TRUNCATE,  MVT::v4i16,   MVT::v4i64,  2 }, // vpmovqw
2038     { ISD::TRUNCATE,  MVT::v8i8,    MVT::v8i32,  2 }, // vpmovwb
2039 
2040     // sign extend is vpcmpeq+maskedmove+vpmovdw+vpacksswb
2041     // zero extend is vpcmpeq+maskedmove+vpmovdw+vpsrlw+vpackuswb
2042     { ISD::SIGN_EXTEND, MVT::v2i8,   MVT::v2i1,   5 },
2043     { ISD::ZERO_EXTEND, MVT::v2i8,   MVT::v2i1,   6 },
2044     { ISD::SIGN_EXTEND, MVT::v4i8,   MVT::v4i1,   5 },
2045     { ISD::ZERO_EXTEND, MVT::v4i8,   MVT::v4i1,   6 },
2046     { ISD::SIGN_EXTEND, MVT::v8i8,   MVT::v8i1,   5 },
2047     { ISD::ZERO_EXTEND, MVT::v8i8,   MVT::v8i1,   6 },
2048     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v16i1, 10 },
2049     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v16i1, 12 },
2050 
2051     // sign extend is vpcmpeq+maskedmove+vpmovdw
2052     // zero extend is vpcmpeq+maskedmove+vpmovdw+vpsrlw
2053     { ISD::SIGN_EXTEND, MVT::v2i16,  MVT::v2i1,   4 },
2054     { ISD::ZERO_EXTEND, MVT::v2i16,  MVT::v2i1,   5 },
2055     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i1,   4 },
2056     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i1,   5 },
2057     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i1,   4 },
2058     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i1,   5 },
2059     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 10 },
2060     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 12 },
2061 
2062     { ISD::SIGN_EXTEND, MVT::v2i32,  MVT::v2i1,   1 }, // vpternlogd
2063     { ISD::ZERO_EXTEND, MVT::v2i32,  MVT::v2i1,   2 }, // vpternlogd+psrld
2064     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i1,   1 }, // vpternlogd
2065     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i1,   2 }, // vpternlogd+psrld
2066     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   1 }, // vpternlogd
2067     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   2 }, // vpternlogd+psrld
2068     { ISD::SIGN_EXTEND, MVT::v2i64,  MVT::v2i1,   1 }, // vpternlogq
2069     { ISD::ZERO_EXTEND, MVT::v2i64,  MVT::v2i1,   2 }, // vpternlogq+psrlq
2070     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   1 }, // vpternlogq
2071     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   2 }, // vpternlogq+psrlq
2072 
2073     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v16i8,  1 },
2074     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v16i8,  1 },
2075     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v16i8,  1 },
2076     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v16i8,  1 },
2077     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
2078     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
2079     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v8i16,  1 },
2080     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v8i16,  1 },
2081     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
2082     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
2083     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
2084     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
2085 
2086     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  1 },
2087     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v16i8,  1 },
2088     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  1 },
2089     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  1 },
2090 
2091     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i64,    1 },
2092     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i64,    1 },
2093     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  1 },
2094     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v16i8,  1 },
2095     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  1 },
2096     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  1 },
2097     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  1 },
2098     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  1 },
2099     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  1 },
2100     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  1 },
2101     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  5 },
2102     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  5 },
2103     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  5 },
2104 
2105     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v8f32,  2 },
2106     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v16f32, 2 },
2107     { ISD::FP_TO_SINT,  MVT::v32i8,  MVT::v32f32, 5 },
2108 
2109     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f32,    1 },
2110     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f64,    1 },
2111     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  1 },
2112     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v2f64,  1 },
2113     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f64,  1 },
2114     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f32,  1 },
2115     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f64,  1 },
2116   };
2117 
2118   static const TypeConversionCostTblEntry AVX2ConversionTbl[] = {
2119     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
2120     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
2121     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
2122     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
2123     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1,  1 },
2124     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1,  1 },
2125 
2126     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v16i8,  2 },
2127     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v16i8,  2 },
2128     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v16i8,  2 },
2129     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v16i8,  2 },
2130     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
2131     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
2132     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v8i16,  2 },
2133     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v8i16,  2 },
2134     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
2135     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
2136     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 3 },
2137     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 3 },
2138     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  2 },
2139     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  2 },
2140 
2141     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i32,  2 },
2142 
2143     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 4 },
2144     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i32, 4 },
2145     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v8i16,  1 },
2146     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v4i32,  1 },
2147     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v2i64,  1 },
2148     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v8i32,  4 },
2149     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v4i64,  4 },
2150     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v4i32,  1 },
2151     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v2i64,  1 },
2152     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v4i64,  5 },
2153     { ISD::TRUNCATE,    MVT::v4i32,  MVT::v4i64,  1 },
2154     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  2 },
2155 
2156     { ISD::FP_EXTEND,   MVT::v8f64,  MVT::v8f32,  3 },
2157     { ISD::FP_ROUND,    MVT::v8f32,  MVT::v8f64,  3 },
2158 
2159     { ISD::FP_TO_SINT,  MVT::v16i16, MVT::v8f32,  1 },
2160     { ISD::FP_TO_SINT,  MVT::v4i32,  MVT::v4f64,  1 },
2161     { ISD::FP_TO_SINT,  MVT::v8i32,  MVT::v8f32,  1 },
2162     { ISD::FP_TO_SINT,  MVT::v8i32,  MVT::v8f64,  3 },
2163 
2164     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f32,    3 },
2165     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f64,    3 },
2166     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v8f32,  1 },
2167     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  3 },
2168     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v2f64,  4 },
2169     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f64,  4 },
2170     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f32,  3 },
2171     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v4f64,  4 },
2172 
2173     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  2 },
2174     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v16i8,  2 },
2175     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  2 },
2176     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  2 },
2177     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  1 },
2178     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  1 },
2179     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  3 },
2180 
2181     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  2 },
2182     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v16i8,  2 },
2183     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  2 },
2184     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  2 },
2185     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  2 },
2186     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i32,  1 },
2187     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  2 },
2188     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  2 },
2189     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  2 },
2190     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  4 },
2191   };
2192 
2193   static const TypeConversionCostTblEntry AVXConversionTbl[] = {
2194     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   6 },
2195     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   4 },
2196     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   7 },
2197     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   4 },
2198     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1,  4 },
2199     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1,  4 },
2200 
2201     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v16i8,  3 },
2202     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v16i8,  3 },
2203     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v16i8,  3 },
2204     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v16i8,  3 },
2205     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  3 },
2206     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  3 },
2207     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v8i16,  3 },
2208     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v8i16,  3 },
2209     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  3 },
2210     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  3 },
2211     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  3 },
2212     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  3 },
2213 
2214     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i64,  4 },
2215     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i32,  5 },
2216     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i16, 4 },
2217     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i64,  9 },
2218     { ISD::TRUNCATE,    MVT::v16i1,  MVT::v16i64, 11 },
2219 
2220     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 6 },
2221     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i32, 6 },
2222     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 2 }, // and+extract+packuswb
2223     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v8i32,  5 },
2224     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  5 },
2225     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v4i64,  5 },
2226     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v4i64,  3 }, // and+extract+2*packusdw
2227     { ISD::TRUNCATE,    MVT::v4i32,  MVT::v4i64,  2 },
2228 
2229     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i1,   3 },
2230     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i1,   3 },
2231     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i1,   8 },
2232     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v16i8,  4 },
2233     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v16i8,  2 },
2234     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  4 },
2235     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v8i16,  2 },
2236     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  2 },
2237     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  2 },
2238     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  4 },
2239     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v2i64,  5 },
2240     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  8 },
2241 
2242     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i1,   7 },
2243     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i1,   7 },
2244     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i1,   6 },
2245     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v16i8,  4 },
2246     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v16i8,  2 },
2247     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  4 },
2248     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v8i16,  2 },
2249     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  4 },
2250     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i32,  4 },
2251     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  5 },
2252     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  6 },
2253     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  8 },
2254     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i32, 10 },
2255     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64, 10 },
2256     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i64, 18 },
2257     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  5 },
2258     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64, 10 },
2259 
2260     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v8f32,  2 },
2261     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v4f64,  2 },
2262     { ISD::FP_TO_SINT,  MVT::v32i8,  MVT::v8f32,  2 },
2263     { ISD::FP_TO_SINT,  MVT::v32i8,  MVT::v4f64,  2 },
2264     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v8f32,  2 },
2265     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v4f64,  2 },
2266     { ISD::FP_TO_SINT,  MVT::v16i16, MVT::v8f32,  2 },
2267     { ISD::FP_TO_SINT,  MVT::v16i16, MVT::v4f64,  2 },
2268     { ISD::FP_TO_SINT,  MVT::v4i32,  MVT::v4f64,  2 },
2269     { ISD::FP_TO_SINT,  MVT::v8i32,  MVT::v8f32,  2 },
2270     { ISD::FP_TO_SINT,  MVT::v8i32,  MVT::v8f64,  5 },
2271 
2272     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v8f32,  2 },
2273     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v4f64,  2 },
2274     { ISD::FP_TO_UINT,  MVT::v32i8,  MVT::v8f32,  2 },
2275     { ISD::FP_TO_UINT,  MVT::v32i8,  MVT::v4f64,  2 },
2276     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v8f32,  2 },
2277     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v4f64,  2 },
2278     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v8f32,  2 },
2279     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v4f64,  2 },
2280     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  3 },
2281     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v2f64,  4 },
2282     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f64,  6 },
2283     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f32,  7 },
2284     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v4f64,  7 },
2285 
2286     { ISD::FP_EXTEND,   MVT::v4f64,  MVT::v4f32,  1 },
2287     { ISD::FP_ROUND,    MVT::v4f32,  MVT::v4f64,  1 },
2288   };
2289 
2290   static const TypeConversionCostTblEntry SSE41ConversionTbl[] = {
2291     { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v16i8,   1 },
2292     { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v16i8,   1 },
2293     { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v16i8,   1 },
2294     { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v16i8,   1 },
2295     { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v16i8,   1 },
2296     { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v16i8,   1 },
2297     { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v8i16,   1 },
2298     { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v8i16,   1 },
2299     { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v8i16,   1 },
2300     { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v8i16,   1 },
2301     { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v4i32,   1 },
2302     { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v4i32,   1 },
2303 
2304     // These truncates end up widening elements.
2305     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i8,   1 }, // PMOVXZBQ
2306     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i16,  1 }, // PMOVXZWQ
2307     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i8,   1 }, // PMOVXZBD
2308 
2309     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v4i32,  2 },
2310     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v4i32,  2 },
2311     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v2i64,  2 },
2312 
2313     { ISD::SINT_TO_FP,  MVT::f32,    MVT::i32,    1 },
2314     { ISD::SINT_TO_FP,  MVT::f64,    MVT::i32,    1 },
2315     { ISD::SINT_TO_FP,  MVT::f32,    MVT::i64,    1 },
2316     { ISD::SINT_TO_FP,  MVT::f64,    MVT::i64,    1 },
2317     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v16i8,  1 },
2318     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  1 },
2319     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v8i16,  1 },
2320     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  1 },
2321     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  1 },
2322     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v4i32,  1 },
2323     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  2 },
2324 
2325     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i32,    1 },
2326     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i32,    1 },
2327     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i64,    4 },
2328     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i64,    4 },
2329     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v16i8,  1 },
2330     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  1 },
2331     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v8i16,  1 },
2332     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  1 },
2333     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  3 },
2334     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  3 },
2335     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v4i32,  2 },
2336     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v2i64, 12 },
2337     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i64, 22 },
2338     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  4 },
2339 
2340     { ISD::FP_TO_SINT,  MVT::i32,    MVT::f32,    1 },
2341     { ISD::FP_TO_SINT,  MVT::i64,    MVT::f32,    1 },
2342     { ISD::FP_TO_SINT,  MVT::i32,    MVT::f64,    1 },
2343     { ISD::FP_TO_SINT,  MVT::i64,    MVT::f64,    1 },
2344     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v4f32,  2 },
2345     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v2f64,  2 },
2346     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v4f32,  1 },
2347     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v2f64,  1 },
2348     { ISD::FP_TO_SINT,  MVT::v4i32,  MVT::v4f32,  1 },
2349     { ISD::FP_TO_SINT,  MVT::v4i32,  MVT::v2f64,  1 },
2350 
2351     { ISD::FP_TO_UINT,  MVT::i32,    MVT::f32,    1 },
2352     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f32,    4 },
2353     { ISD::FP_TO_UINT,  MVT::i32,    MVT::f64,    1 },
2354     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f64,    4 },
2355     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v4f32,  2 },
2356     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v2f64,  2 },
2357     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v4f32,  1 },
2358     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v2f64,  1 },
2359     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  4 },
2360     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v2f64,  4 },
2361   };
2362 
2363   static const TypeConversionCostTblEntry SSE2ConversionTbl[] = {
2364     // These are somewhat magic numbers justified by comparing the
2365     // output of llvm-mca for our various supported scheduler models
2366     // and basing it off the worst case scenario.
2367     { ISD::SINT_TO_FP,  MVT::f32,    MVT::i32,    3 },
2368     { ISD::SINT_TO_FP,  MVT::f64,    MVT::i32,    3 },
2369     { ISD::SINT_TO_FP,  MVT::f32,    MVT::i64,    3 },
2370     { ISD::SINT_TO_FP,  MVT::f64,    MVT::i64,    3 },
2371     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v16i8,  3 },
2372     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  4 },
2373     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v8i16,  3 },
2374     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  4 },
2375     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  3 },
2376     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v4i32,  4 },
2377     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v2i64,  8 },
2378     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  8 },
2379 
2380     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i32,    3 },
2381     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i32,    3 },
2382     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i64,    8 },
2383     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i64,    9 },
2384     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v16i8,  4 },
2385     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v16i8,  4 },
2386     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v8i16,  4 },
2387     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v8i16,  4 },
2388     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  7 },
2389     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v4i32,  7 },
2390     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  5 },
2391     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64, 15 },
2392     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v2i64, 18 },
2393 
2394     { ISD::FP_TO_SINT,  MVT::i32,    MVT::f32,    4 },
2395     { ISD::FP_TO_SINT,  MVT::i64,    MVT::f32,    4 },
2396     { ISD::FP_TO_SINT,  MVT::i32,    MVT::f64,    4 },
2397     { ISD::FP_TO_SINT,  MVT::i64,    MVT::f64,    4 },
2398     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v4f32,  6 },
2399     { ISD::FP_TO_SINT,  MVT::v16i8,  MVT::v2f64,  6 },
2400     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v4f32,  5 },
2401     { ISD::FP_TO_SINT,  MVT::v8i16,  MVT::v2f64,  5 },
2402     { ISD::FP_TO_SINT,  MVT::v4i32,  MVT::v4f32,  4 },
2403     { ISD::FP_TO_SINT,  MVT::v4i32,  MVT::v2f64,  4 },
2404 
2405     { ISD::FP_TO_UINT,  MVT::i32,    MVT::f32,    4 },
2406     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f32,    4 },
2407     { ISD::FP_TO_UINT,  MVT::i32,    MVT::f64,    4 },
2408     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f64,   15 },
2409     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v4f32,  6 },
2410     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v2f64,  6 },
2411     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v4f32,  5 },
2412     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v2f64,  5 },
2413     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  8 },
2414     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v2f64,  8 },
2415 
2416     { ISD::ZERO_EXTEND, MVT::v2i64,  MVT::v16i8,  4 },
2417     { ISD::SIGN_EXTEND, MVT::v2i64,  MVT::v16i8,  4 },
2418     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v16i8,  2 },
2419     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v16i8,  3 },
2420     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v16i8,  1 },
2421     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v16i8,  2 },
2422     { ISD::ZERO_EXTEND, MVT::v2i64,  MVT::v8i16,  2 },
2423     { ISD::SIGN_EXTEND, MVT::v2i64,  MVT::v8i16,  3 },
2424     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v8i16,  1 },
2425     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v8i16,  2 },
2426     { ISD::ZERO_EXTEND, MVT::v2i64,  MVT::v4i32,  1 },
2427     { ISD::SIGN_EXTEND, MVT::v2i64,  MVT::v4i32,  2 },
2428 
2429     // These truncates are really widening elements.
2430     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i32,  1 }, // PSHUFD
2431     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i16,  2 }, // PUNPCKLWD+DQ
2432     { ISD::TRUNCATE,    MVT::v2i1,   MVT::v2i8,   3 }, // PUNPCKLBW+WD+PSHUFD
2433     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i16,  1 }, // PUNPCKLWD
2434     { ISD::TRUNCATE,    MVT::v4i1,   MVT::v4i8,   2 }, // PUNPCKLBW+WD
2435     { ISD::TRUNCATE,    MVT::v8i1,   MVT::v8i8,   1 }, // PUNPCKLBW
2436 
2437     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v8i16,  2 }, // PAND+PACKUSWB
2438     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 3 },
2439     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v4i32,  3 }, // PAND+2*PACKUSWB
2440     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i32, 7 },
2441     { ISD::TRUNCATE,    MVT::v2i16,  MVT::v2i32,  1 },
2442     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v4i32,  3 },
2443     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  5 },
2444     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32,10 },
2445     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v2i64,  4 }, // PAND+3*PACKUSWB
2446     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v2i64,  2 }, // PSHUFD+PSHUFLW
2447     { ISD::TRUNCATE,    MVT::v4i32,  MVT::v2i64,  1 }, // PSHUFD
2448   };
2449 
2450   // Attempt to map directly to (simple) MVT types to let us match custom entries.
2451   EVT SrcTy = TLI->getValueType(DL, Src);
2452   EVT DstTy = TLI->getValueType(DL, Dst);
2453 
2454   // The function getSimpleVT only handles simple value types.
2455   if (SrcTy.isSimple() && DstTy.isSimple()) {
2456     MVT SimpleSrcTy = SrcTy.getSimpleVT();
2457     MVT SimpleDstTy = DstTy.getSimpleVT();
2458 
2459     if (ST->useAVX512Regs()) {
2460       if (ST->hasBWI())
2461         if (const auto *Entry = ConvertCostTableLookup(
2462                 AVX512BWConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2463           return AdjustCost(Entry->Cost);
2464 
2465       if (ST->hasDQI())
2466         if (const auto *Entry = ConvertCostTableLookup(
2467                 AVX512DQConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2468           return AdjustCost(Entry->Cost);
2469 
2470       if (ST->hasAVX512())
2471         if (const auto *Entry = ConvertCostTableLookup(
2472                 AVX512FConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2473           return AdjustCost(Entry->Cost);
2474     }
2475 
2476     if (ST->hasBWI())
2477       if (const auto *Entry = ConvertCostTableLookup(
2478               AVX512BWVLConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2479         return AdjustCost(Entry->Cost);
2480 
2481     if (ST->hasDQI())
2482       if (const auto *Entry = ConvertCostTableLookup(
2483               AVX512DQVLConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2484         return AdjustCost(Entry->Cost);
2485 
2486     if (ST->hasAVX512())
2487       if (const auto *Entry = ConvertCostTableLookup(AVX512VLConversionTbl, ISD,
2488                                                      SimpleDstTy, SimpleSrcTy))
2489         return AdjustCost(Entry->Cost);
2490 
2491     if (ST->hasAVX2()) {
2492       if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
2493                                                      SimpleDstTy, SimpleSrcTy))
2494         return AdjustCost(Entry->Cost);
2495     }
2496 
2497     if (ST->hasAVX()) {
2498       if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
2499                                                      SimpleDstTy, SimpleSrcTy))
2500         return AdjustCost(Entry->Cost);
2501     }
2502 
2503     if (ST->hasSSE41()) {
2504       if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
2505                                                      SimpleDstTy, SimpleSrcTy))
2506         return AdjustCost(Entry->Cost);
2507     }
2508 
2509     if (ST->hasSSE2()) {
2510       if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
2511                                                      SimpleDstTy, SimpleSrcTy))
2512         return AdjustCost(Entry->Cost);
2513     }
2514   }
2515 
2516   // Fall back to legalized types.
2517   std::pair<InstructionCost, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src);
2518   std::pair<InstructionCost, MVT> LTDest =
2519       TLI->getTypeLegalizationCost(DL, Dst);
2520 
2521   if (ST->useAVX512Regs()) {
2522     if (ST->hasBWI())
2523       if (const auto *Entry = ConvertCostTableLookup(
2524               AVX512BWConversionTbl, ISD, LTDest.second, LTSrc.second))
2525         return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2526 
2527     if (ST->hasDQI())
2528       if (const auto *Entry = ConvertCostTableLookup(
2529               AVX512DQConversionTbl, ISD, LTDest.second, LTSrc.second))
2530         return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2531 
2532     if (ST->hasAVX512())
2533       if (const auto *Entry = ConvertCostTableLookup(
2534               AVX512FConversionTbl, ISD, LTDest.second, LTSrc.second))
2535         return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2536   }
2537 
2538   if (ST->hasBWI())
2539     if (const auto *Entry = ConvertCostTableLookup(AVX512BWVLConversionTbl, ISD,
2540                                                    LTDest.second, LTSrc.second))
2541       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2542 
2543   if (ST->hasDQI())
2544     if (const auto *Entry = ConvertCostTableLookup(AVX512DQVLConversionTbl, ISD,
2545                                                    LTDest.second, LTSrc.second))
2546       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2547 
2548   if (ST->hasAVX512())
2549     if (const auto *Entry = ConvertCostTableLookup(AVX512VLConversionTbl, ISD,
2550                                                    LTDest.second, LTSrc.second))
2551       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2552 
2553   if (ST->hasAVX2())
2554     if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
2555                                                    LTDest.second, LTSrc.second))
2556       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2557 
2558   if (ST->hasAVX())
2559     if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
2560                                                    LTDest.second, LTSrc.second))
2561       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2562 
2563   if (ST->hasSSE41())
2564     if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
2565                                                    LTDest.second, LTSrc.second))
2566       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2567 
2568   if (ST->hasSSE2())
2569     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
2570                                                    LTDest.second, LTSrc.second))
2571       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2572 
2573   // Fallback, for i8/i16 sitofp/uitofp cases we need to extend to i32 for
2574   // sitofp.
2575   if ((ISD == ISD::SINT_TO_FP || ISD == ISD::UINT_TO_FP) &&
2576       1 < Src->getScalarSizeInBits() && Src->getScalarSizeInBits() < 32) {
2577     Type *ExtSrc = Src->getWithNewBitWidth(32);
2578     unsigned ExtOpc =
2579         (ISD == ISD::SINT_TO_FP) ? Instruction::SExt : Instruction::ZExt;
2580 
2581     // For scalar loads the extend would be free.
2582     InstructionCost ExtCost = 0;
2583     if (!(Src->isIntegerTy() && I && isa<LoadInst>(I->getOperand(0))))
2584       ExtCost = getCastInstrCost(ExtOpc, ExtSrc, Src, CCH, CostKind);
2585 
2586     return ExtCost + getCastInstrCost(Instruction::SIToFP, Dst, ExtSrc,
2587                                       TTI::CastContextHint::None, CostKind);
2588   }
2589 
2590   // Fallback for fptosi/fptoui i8/i16 cases we need to truncate from fptosi
2591   // i32.
2592   if ((ISD == ISD::FP_TO_SINT || ISD == ISD::FP_TO_UINT) &&
2593       1 < Dst->getScalarSizeInBits() && Dst->getScalarSizeInBits() < 32) {
2594     Type *TruncDst = Dst->getWithNewBitWidth(32);
2595     return getCastInstrCost(Instruction::FPToSI, TruncDst, Src, CCH, CostKind) +
2596            getCastInstrCost(Instruction::Trunc, Dst, TruncDst,
2597                             TTI::CastContextHint::None, CostKind);
2598   }
2599 
2600   return AdjustCost(
2601       BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I));
2602 }
2603 
2604 InstructionCost X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
2605                                                Type *CondTy,
2606                                                CmpInst::Predicate VecPred,
2607                                                TTI::TargetCostKind CostKind,
2608                                                const Instruction *I) {
2609   // TODO: Handle other cost kinds.
2610   if (CostKind != TTI::TCK_RecipThroughput)
2611     return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
2612                                      I);
2613 
2614   // Legalize the type.
2615   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
2616 
2617   MVT MTy = LT.second;
2618 
2619   int ISD = TLI->InstructionOpcodeToISD(Opcode);
2620   assert(ISD && "Invalid opcode");
2621 
2622   unsigned ExtraCost = 0;
2623   if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) {
2624     // Some vector comparison predicates cost extra instructions.
2625     // TODO: Should we invert this and assume worst case cmp costs
2626     // and reduce for particular predicates?
2627     if (MTy.isVector() &&
2628         !((ST->hasXOP() && (!ST->hasAVX2() || MTy.is128BitVector())) ||
2629           (ST->hasAVX512() && 32 <= MTy.getScalarSizeInBits()) ||
2630           ST->hasBWI())) {
2631       // Fallback to I if a specific predicate wasn't specified.
2632       CmpInst::Predicate Pred = VecPred;
2633       if (I && (Pred == CmpInst::BAD_ICMP_PREDICATE ||
2634                 Pred == CmpInst::BAD_FCMP_PREDICATE))
2635         Pred = cast<CmpInst>(I)->getPredicate();
2636 
2637       switch (Pred) {
2638       case CmpInst::Predicate::ICMP_NE:
2639         // xor(cmpeq(x,y),-1)
2640         ExtraCost = 1;
2641         break;
2642       case CmpInst::Predicate::ICMP_SGE:
2643       case CmpInst::Predicate::ICMP_SLE:
2644         // xor(cmpgt(x,y),-1)
2645         ExtraCost = 1;
2646         break;
2647       case CmpInst::Predicate::ICMP_ULT:
2648       case CmpInst::Predicate::ICMP_UGT:
2649         // cmpgt(xor(x,signbit),xor(y,signbit))
2650         // xor(cmpeq(pmaxu(x,y),x),-1)
2651         ExtraCost = 2;
2652         break;
2653       case CmpInst::Predicate::ICMP_ULE:
2654       case CmpInst::Predicate::ICMP_UGE:
2655         if ((ST->hasSSE41() && MTy.getScalarSizeInBits() == 32) ||
2656             (ST->hasSSE2() && MTy.getScalarSizeInBits() < 32)) {
2657           // cmpeq(psubus(x,y),0)
2658           // cmpeq(pminu(x,y),x)
2659           ExtraCost = 1;
2660         } else {
2661           // xor(cmpgt(xor(x,signbit),xor(y,signbit)),-1)
2662           ExtraCost = 3;
2663         }
2664         break;
2665       case CmpInst::Predicate::BAD_ICMP_PREDICATE:
2666       case CmpInst::Predicate::BAD_FCMP_PREDICATE:
2667         // Assume worst case scenario and add the maximum extra cost.
2668         ExtraCost = 3;
2669         break;
2670       default:
2671         break;
2672       }
2673     }
2674   }
2675 
2676   static const CostTblEntry SLMCostTbl[] = {
2677     // slm pcmpeq/pcmpgt throughput is 2
2678     { ISD::SETCC,   MVT::v2i64,   2 },
2679   };
2680 
2681   static const CostTblEntry AVX512BWCostTbl[] = {
2682     { ISD::SETCC,   MVT::v32i16,  1 },
2683     { ISD::SETCC,   MVT::v64i8,   1 },
2684 
2685     { ISD::SELECT,  MVT::v32i16,  1 },
2686     { ISD::SELECT,  MVT::v64i8,   1 },
2687   };
2688 
2689   static const CostTblEntry AVX512CostTbl[] = {
2690     { ISD::SETCC,   MVT::v8i64,   1 },
2691     { ISD::SETCC,   MVT::v16i32,  1 },
2692     { ISD::SETCC,   MVT::v8f64,   1 },
2693     { ISD::SETCC,   MVT::v16f32,  1 },
2694 
2695     { ISD::SELECT,  MVT::v8i64,   1 },
2696     { ISD::SELECT,  MVT::v16i32,  1 },
2697     { ISD::SELECT,  MVT::v8f64,   1 },
2698     { ISD::SELECT,  MVT::v16f32,  1 },
2699 
2700     { ISD::SETCC,   MVT::v32i16,  2 }, // FIXME: should probably be 4
2701     { ISD::SETCC,   MVT::v64i8,   2 }, // FIXME: should probably be 4
2702 
2703     { ISD::SELECT,  MVT::v32i16,  2 }, // FIXME: should be 3
2704     { ISD::SELECT,  MVT::v64i8,   2 }, // FIXME: should be 3
2705   };
2706 
2707   static const CostTblEntry AVX2CostTbl[] = {
2708     { ISD::SETCC,   MVT::v4i64,   1 },
2709     { ISD::SETCC,   MVT::v8i32,   1 },
2710     { ISD::SETCC,   MVT::v16i16,  1 },
2711     { ISD::SETCC,   MVT::v32i8,   1 },
2712 
2713     { ISD::SELECT,  MVT::v4i64,   1 }, // pblendvb
2714     { ISD::SELECT,  MVT::v8i32,   1 }, // pblendvb
2715     { ISD::SELECT,  MVT::v16i16,  1 }, // pblendvb
2716     { ISD::SELECT,  MVT::v32i8,   1 }, // pblendvb
2717   };
2718 
2719   static const CostTblEntry AVX1CostTbl[] = {
2720     { ISD::SETCC,   MVT::v4f64,   1 },
2721     { ISD::SETCC,   MVT::v8f32,   1 },
2722     // AVX1 does not support 8-wide integer compare.
2723     { ISD::SETCC,   MVT::v4i64,   4 },
2724     { ISD::SETCC,   MVT::v8i32,   4 },
2725     { ISD::SETCC,   MVT::v16i16,  4 },
2726     { ISD::SETCC,   MVT::v32i8,   4 },
2727 
2728     { ISD::SELECT,  MVT::v4f64,   1 }, // vblendvpd
2729     { ISD::SELECT,  MVT::v8f32,   1 }, // vblendvps
2730     { ISD::SELECT,  MVT::v4i64,   1 }, // vblendvpd
2731     { ISD::SELECT,  MVT::v8i32,   1 }, // vblendvps
2732     { ISD::SELECT,  MVT::v16i16,  3 }, // vandps + vandnps + vorps
2733     { ISD::SELECT,  MVT::v32i8,   3 }, // vandps + vandnps + vorps
2734   };
2735 
2736   static const CostTblEntry SSE42CostTbl[] = {
2737     { ISD::SETCC,   MVT::v2f64,   1 },
2738     { ISD::SETCC,   MVT::v4f32,   1 },
2739     { ISD::SETCC,   MVT::v2i64,   1 },
2740   };
2741 
2742   static const CostTblEntry SSE41CostTbl[] = {
2743     { ISD::SELECT,  MVT::v2f64,   1 }, // blendvpd
2744     { ISD::SELECT,  MVT::v4f32,   1 }, // blendvps
2745     { ISD::SELECT,  MVT::v2i64,   1 }, // pblendvb
2746     { ISD::SELECT,  MVT::v4i32,   1 }, // pblendvb
2747     { ISD::SELECT,  MVT::v8i16,   1 }, // pblendvb
2748     { ISD::SELECT,  MVT::v16i8,   1 }, // pblendvb
2749   };
2750 
2751   static const CostTblEntry SSE2CostTbl[] = {
2752     { ISD::SETCC,   MVT::v2f64,   2 },
2753     { ISD::SETCC,   MVT::f64,     1 },
2754     { ISD::SETCC,   MVT::v2i64,   5 }, // pcmpeqd/pcmpgtd expansion
2755     { ISD::SETCC,   MVT::v4i32,   1 },
2756     { ISD::SETCC,   MVT::v8i16,   1 },
2757     { ISD::SETCC,   MVT::v16i8,   1 },
2758 
2759     { ISD::SELECT,  MVT::v2f64,   3 }, // andpd + andnpd + orpd
2760     { ISD::SELECT,  MVT::v2i64,   3 }, // pand + pandn + por
2761     { ISD::SELECT,  MVT::v4i32,   3 }, // pand + pandn + por
2762     { ISD::SELECT,  MVT::v8i16,   3 }, // pand + pandn + por
2763     { ISD::SELECT,  MVT::v16i8,   3 }, // pand + pandn + por
2764   };
2765 
2766   static const CostTblEntry SSE1CostTbl[] = {
2767     { ISD::SETCC,   MVT::v4f32,   2 },
2768     { ISD::SETCC,   MVT::f32,     1 },
2769 
2770     { ISD::SELECT,  MVT::v4f32,   3 }, // andps + andnps + orps
2771   };
2772 
2773   if (ST->useSLMArithCosts())
2774     if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy))
2775       return LT.first * (ExtraCost + Entry->Cost);
2776 
2777   if (ST->hasBWI())
2778     if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
2779       return LT.first * (ExtraCost + Entry->Cost);
2780 
2781   if (ST->hasAVX512())
2782     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
2783       return LT.first * (ExtraCost + Entry->Cost);
2784 
2785   if (ST->hasAVX2())
2786     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
2787       return LT.first * (ExtraCost + Entry->Cost);
2788 
2789   if (ST->hasAVX())
2790     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
2791       return LT.first * (ExtraCost + Entry->Cost);
2792 
2793   if (ST->hasSSE42())
2794     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
2795       return LT.first * (ExtraCost + Entry->Cost);
2796 
2797   if (ST->hasSSE41())
2798     if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
2799       return LT.first * (ExtraCost + Entry->Cost);
2800 
2801   if (ST->hasSSE2())
2802     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
2803       return LT.first * (ExtraCost + Entry->Cost);
2804 
2805   if (ST->hasSSE1())
2806     if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
2807       return LT.first * (ExtraCost + Entry->Cost);
2808 
2809   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I);
2810 }
2811 
2812 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; }
2813 
2814 InstructionCost
2815 X86TTIImpl::getTypeBasedIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
2816                                            TTI::TargetCostKind CostKind) {
2817 
2818   // Costs should match the codegen from:
2819   // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll
2820   // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll
2821   // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll
2822   // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll
2823   // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll
2824 
2825   // TODO: Overflow intrinsics (*ADDO, *SUBO, *MULO) with vector types are not
2826   //       specialized in these tables yet.
2827   static const CostTblEntry AVX512BITALGCostTbl[] = {
2828     { ISD::CTPOP,      MVT::v32i16,  1 },
2829     { ISD::CTPOP,      MVT::v64i8,   1 },
2830     { ISD::CTPOP,      MVT::v16i16,  1 },
2831     { ISD::CTPOP,      MVT::v32i8,   1 },
2832     { ISD::CTPOP,      MVT::v8i16,   1 },
2833     { ISD::CTPOP,      MVT::v16i8,   1 },
2834   };
2835   static const CostTblEntry AVX512VPOPCNTDQCostTbl[] = {
2836     { ISD::CTPOP,      MVT::v8i64,   1 },
2837     { ISD::CTPOP,      MVT::v16i32,  1 },
2838     { ISD::CTPOP,      MVT::v4i64,   1 },
2839     { ISD::CTPOP,      MVT::v8i32,   1 },
2840     { ISD::CTPOP,      MVT::v2i64,   1 },
2841     { ISD::CTPOP,      MVT::v4i32,   1 },
2842   };
2843   static const CostTblEntry AVX512CDCostTbl[] = {
2844     { ISD::CTLZ,       MVT::v8i64,   1 },
2845     { ISD::CTLZ,       MVT::v16i32,  1 },
2846     { ISD::CTLZ,       MVT::v32i16,  8 },
2847     { ISD::CTLZ,       MVT::v64i8,  20 },
2848     { ISD::CTLZ,       MVT::v4i64,   1 },
2849     { ISD::CTLZ,       MVT::v8i32,   1 },
2850     { ISD::CTLZ,       MVT::v16i16,  4 },
2851     { ISD::CTLZ,       MVT::v32i8,  10 },
2852     { ISD::CTLZ,       MVT::v2i64,   1 },
2853     { ISD::CTLZ,       MVT::v4i32,   1 },
2854     { ISD::CTLZ,       MVT::v8i16,   4 },
2855     { ISD::CTLZ,       MVT::v16i8,   4 },
2856   };
2857   static const CostTblEntry AVX512BWCostTbl[] = {
2858     { ISD::ABS,        MVT::v32i16,  1 },
2859     { ISD::ABS,        MVT::v64i8,   1 },
2860     { ISD::BITREVERSE, MVT::v8i64,   3 },
2861     { ISD::BITREVERSE, MVT::v16i32,  3 },
2862     { ISD::BITREVERSE, MVT::v32i16,  3 },
2863     { ISD::BITREVERSE, MVT::v64i8,   2 },
2864     { ISD::BSWAP,      MVT::v8i64,   1 },
2865     { ISD::BSWAP,      MVT::v16i32,  1 },
2866     { ISD::BSWAP,      MVT::v32i16,  1 },
2867     { ISD::CTLZ,       MVT::v8i64,  23 },
2868     { ISD::CTLZ,       MVT::v16i32, 22 },
2869     { ISD::CTLZ,       MVT::v32i16, 18 },
2870     { ISD::CTLZ,       MVT::v64i8,  17 },
2871     { ISD::CTPOP,      MVT::v8i64,   7 },
2872     { ISD::CTPOP,      MVT::v16i32, 11 },
2873     { ISD::CTPOP,      MVT::v32i16,  9 },
2874     { ISD::CTPOP,      MVT::v64i8,   6 },
2875     { ISD::CTTZ,       MVT::v8i64,  10 },
2876     { ISD::CTTZ,       MVT::v16i32, 14 },
2877     { ISD::CTTZ,       MVT::v32i16, 12 },
2878     { ISD::CTTZ,       MVT::v64i8,   9 },
2879     { ISD::SADDSAT,    MVT::v32i16,  1 },
2880     { ISD::SADDSAT,    MVT::v64i8,   1 },
2881     { ISD::SMAX,       MVT::v32i16,  1 },
2882     { ISD::SMAX,       MVT::v64i8,   1 },
2883     { ISD::SMIN,       MVT::v32i16,  1 },
2884     { ISD::SMIN,       MVT::v64i8,   1 },
2885     { ISD::SSUBSAT,    MVT::v32i16,  1 },
2886     { ISD::SSUBSAT,    MVT::v64i8,   1 },
2887     { ISD::UADDSAT,    MVT::v32i16,  1 },
2888     { ISD::UADDSAT,    MVT::v64i8,   1 },
2889     { ISD::UMAX,       MVT::v32i16,  1 },
2890     { ISD::UMAX,       MVT::v64i8,   1 },
2891     { ISD::UMIN,       MVT::v32i16,  1 },
2892     { ISD::UMIN,       MVT::v64i8,   1 },
2893     { ISD::USUBSAT,    MVT::v32i16,  1 },
2894     { ISD::USUBSAT,    MVT::v64i8,   1 },
2895   };
2896   static const CostTblEntry AVX512CostTbl[] = {
2897     { ISD::ABS,        MVT::v8i64,   1 },
2898     { ISD::ABS,        MVT::v16i32,  1 },
2899     { ISD::ABS,        MVT::v32i16,  2 },
2900     { ISD::ABS,        MVT::v64i8,   2 },
2901     { ISD::ABS,        MVT::v4i64,   1 },
2902     { ISD::ABS,        MVT::v2i64,   1 },
2903     { ISD::BITREVERSE, MVT::v8i64,  36 },
2904     { ISD::BITREVERSE, MVT::v16i32, 24 },
2905     { ISD::BITREVERSE, MVT::v32i16, 10 },
2906     { ISD::BITREVERSE, MVT::v64i8,  10 },
2907     { ISD::BSWAP,      MVT::v8i64,   4 },
2908     { ISD::BSWAP,      MVT::v16i32,  4 },
2909     { ISD::BSWAP,      MVT::v32i16,  4 },
2910     { ISD::CTLZ,       MVT::v8i64,  29 },
2911     { ISD::CTLZ,       MVT::v16i32, 35 },
2912     { ISD::CTLZ,       MVT::v32i16, 28 },
2913     { ISD::CTLZ,       MVT::v64i8,  18 },
2914     { ISD::CTPOP,      MVT::v8i64,  16 },
2915     { ISD::CTPOP,      MVT::v16i32, 24 },
2916     { ISD::CTPOP,      MVT::v32i16, 18 },
2917     { ISD::CTPOP,      MVT::v64i8,  12 },
2918     { ISD::CTTZ,       MVT::v8i64,  20 },
2919     { ISD::CTTZ,       MVT::v16i32, 28 },
2920     { ISD::CTTZ,       MVT::v32i16, 24 },
2921     { ISD::CTTZ,       MVT::v64i8,  18 },
2922     { ISD::SMAX,       MVT::v8i64,   1 },
2923     { ISD::SMAX,       MVT::v16i32,  1 },
2924     { ISD::SMAX,       MVT::v32i16,  2 },
2925     { ISD::SMAX,       MVT::v64i8,   2 },
2926     { ISD::SMAX,       MVT::v4i64,   1 },
2927     { ISD::SMAX,       MVT::v2i64,   1 },
2928     { ISD::SMIN,       MVT::v8i64,   1 },
2929     { ISD::SMIN,       MVT::v16i32,  1 },
2930     { ISD::SMIN,       MVT::v32i16,  2 },
2931     { ISD::SMIN,       MVT::v64i8,   2 },
2932     { ISD::SMIN,       MVT::v4i64,   1 },
2933     { ISD::SMIN,       MVT::v2i64,   1 },
2934     { ISD::UMAX,       MVT::v8i64,   1 },
2935     { ISD::UMAX,       MVT::v16i32,  1 },
2936     { ISD::UMAX,       MVT::v32i16,  2 },
2937     { ISD::UMAX,       MVT::v64i8,   2 },
2938     { ISD::UMAX,       MVT::v4i64,   1 },
2939     { ISD::UMAX,       MVT::v2i64,   1 },
2940     { ISD::UMIN,       MVT::v8i64,   1 },
2941     { ISD::UMIN,       MVT::v16i32,  1 },
2942     { ISD::UMIN,       MVT::v32i16,  2 },
2943     { ISD::UMIN,       MVT::v64i8,   2 },
2944     { ISD::UMIN,       MVT::v4i64,   1 },
2945     { ISD::UMIN,       MVT::v2i64,   1 },
2946     { ISD::USUBSAT,    MVT::v16i32,  2 }, // pmaxud + psubd
2947     { ISD::USUBSAT,    MVT::v2i64,   2 }, // pmaxuq + psubq
2948     { ISD::USUBSAT,    MVT::v4i64,   2 }, // pmaxuq + psubq
2949     { ISD::USUBSAT,    MVT::v8i64,   2 }, // pmaxuq + psubq
2950     { ISD::UADDSAT,    MVT::v16i32,  3 }, // not + pminud + paddd
2951     { ISD::UADDSAT,    MVT::v2i64,   3 }, // not + pminuq + paddq
2952     { ISD::UADDSAT,    MVT::v4i64,   3 }, // not + pminuq + paddq
2953     { ISD::UADDSAT,    MVT::v8i64,   3 }, // not + pminuq + paddq
2954     { ISD::SADDSAT,    MVT::v32i16,  2 },
2955     { ISD::SADDSAT,    MVT::v64i8,   2 },
2956     { ISD::SSUBSAT,    MVT::v32i16,  2 },
2957     { ISD::SSUBSAT,    MVT::v64i8,   2 },
2958     { ISD::UADDSAT,    MVT::v32i16,  2 },
2959     { ISD::UADDSAT,    MVT::v64i8,   2 },
2960     { ISD::USUBSAT,    MVT::v32i16,  2 },
2961     { ISD::USUBSAT,    MVT::v64i8,   2 },
2962     { ISD::FMAXNUM,    MVT::f32,     2 },
2963     { ISD::FMAXNUM,    MVT::v4f32,   2 },
2964     { ISD::FMAXNUM,    MVT::v8f32,   2 },
2965     { ISD::FMAXNUM,    MVT::v16f32,  2 },
2966     { ISD::FMAXNUM,    MVT::f64,     2 },
2967     { ISD::FMAXNUM,    MVT::v2f64,   2 },
2968     { ISD::FMAXNUM,    MVT::v4f64,   2 },
2969     { ISD::FMAXNUM,    MVT::v8f64,   2 },
2970   };
2971   static const CostTblEntry XOPCostTbl[] = {
2972     { ISD::BITREVERSE, MVT::v4i64,   4 },
2973     { ISD::BITREVERSE, MVT::v8i32,   4 },
2974     { ISD::BITREVERSE, MVT::v16i16,  4 },
2975     { ISD::BITREVERSE, MVT::v32i8,   4 },
2976     { ISD::BITREVERSE, MVT::v2i64,   1 },
2977     { ISD::BITREVERSE, MVT::v4i32,   1 },
2978     { ISD::BITREVERSE, MVT::v8i16,   1 },
2979     { ISD::BITREVERSE, MVT::v16i8,   1 },
2980     { ISD::BITREVERSE, MVT::i64,     3 },
2981     { ISD::BITREVERSE, MVT::i32,     3 },
2982     { ISD::BITREVERSE, MVT::i16,     3 },
2983     { ISD::BITREVERSE, MVT::i8,      3 }
2984   };
2985   static const CostTblEntry AVX2CostTbl[] = {
2986     { ISD::ABS,        MVT::v4i64,   2 }, // VBLENDVPD(X,VPSUBQ(0,X),X)
2987     { ISD::ABS,        MVT::v8i32,   1 },
2988     { ISD::ABS,        MVT::v16i16,  1 },
2989     { ISD::ABS,        MVT::v32i8,   1 },
2990     { ISD::BITREVERSE, MVT::v2i64,   3 },
2991     { ISD::BITREVERSE, MVT::v4i64,   3 },
2992     { ISD::BITREVERSE, MVT::v4i32,   3 },
2993     { ISD::BITREVERSE, MVT::v8i32,   3 },
2994     { ISD::BITREVERSE, MVT::v8i16,   3 },
2995     { ISD::BITREVERSE, MVT::v16i16,  3 },
2996     { ISD::BITREVERSE, MVT::v16i8,   3 },
2997     { ISD::BITREVERSE, MVT::v32i8,   3 },
2998     { ISD::BSWAP,      MVT::v4i64,   1 },
2999     { ISD::BSWAP,      MVT::v8i32,   1 },
3000     { ISD::BSWAP,      MVT::v16i16,  1 },
3001     { ISD::CTLZ,       MVT::v2i64,   7 },
3002     { ISD::CTLZ,       MVT::v4i64,   7 },
3003     { ISD::CTLZ,       MVT::v4i32,   5 },
3004     { ISD::CTLZ,       MVT::v8i32,   5 },
3005     { ISD::CTLZ,       MVT::v8i16,   4 },
3006     { ISD::CTLZ,       MVT::v16i16,  4 },
3007     { ISD::CTLZ,       MVT::v16i8,   3 },
3008     { ISD::CTLZ,       MVT::v32i8,   3 },
3009     { ISD::CTPOP,      MVT::v2i64,   3 },
3010     { ISD::CTPOP,      MVT::v4i64,   3 },
3011     { ISD::CTPOP,      MVT::v4i32,   7 },
3012     { ISD::CTPOP,      MVT::v8i32,   7 },
3013     { ISD::CTPOP,      MVT::v8i16,   3 },
3014     { ISD::CTPOP,      MVT::v16i16,  3 },
3015     { ISD::CTPOP,      MVT::v16i8,   2 },
3016     { ISD::CTPOP,      MVT::v32i8,   2 },
3017     { ISD::CTTZ,       MVT::v2i64,   4 },
3018     { ISD::CTTZ,       MVT::v4i64,   4 },
3019     { ISD::CTTZ,       MVT::v4i32,   7 },
3020     { ISD::CTTZ,       MVT::v8i32,   7 },
3021     { ISD::CTTZ,       MVT::v8i16,   4 },
3022     { ISD::CTTZ,       MVT::v16i16,  4 },
3023     { ISD::CTTZ,       MVT::v16i8,   3 },
3024     { ISD::CTTZ,       MVT::v32i8,   3 },
3025     { ISD::SADDSAT,    MVT::v16i16,  1 },
3026     { ISD::SADDSAT,    MVT::v32i8,   1 },
3027     { ISD::SMAX,       MVT::v8i32,   1 },
3028     { ISD::SMAX,       MVT::v16i16,  1 },
3029     { ISD::SMAX,       MVT::v32i8,   1 },
3030     { ISD::SMIN,       MVT::v8i32,   1 },
3031     { ISD::SMIN,       MVT::v16i16,  1 },
3032     { ISD::SMIN,       MVT::v32i8,   1 },
3033     { ISD::SSUBSAT,    MVT::v16i16,  1 },
3034     { ISD::SSUBSAT,    MVT::v32i8,   1 },
3035     { ISD::UADDSAT,    MVT::v16i16,  1 },
3036     { ISD::UADDSAT,    MVT::v32i8,   1 },
3037     { ISD::UADDSAT,    MVT::v8i32,   3 }, // not + pminud + paddd
3038     { ISD::UMAX,       MVT::v8i32,   1 },
3039     { ISD::UMAX,       MVT::v16i16,  1 },
3040     { ISD::UMAX,       MVT::v32i8,   1 },
3041     { ISD::UMIN,       MVT::v8i32,   1 },
3042     { ISD::UMIN,       MVT::v16i16,  1 },
3043     { ISD::UMIN,       MVT::v32i8,   1 },
3044     { ISD::USUBSAT,    MVT::v16i16,  1 },
3045     { ISD::USUBSAT,    MVT::v32i8,   1 },
3046     { ISD::USUBSAT,    MVT::v8i32,   2 }, // pmaxud + psubd
3047     { ISD::FMAXNUM,    MVT::v8f32,   3 }, // MAXPS + CMPUNORDPS + BLENDVPS
3048     { ISD::FMAXNUM,    MVT::v4f64,   3 }, // MAXPD + CMPUNORDPD + BLENDVPD
3049     { ISD::FSQRT,      MVT::f32,     7 }, // Haswell from http://www.agner.org/
3050     { ISD::FSQRT,      MVT::v4f32,   7 }, // Haswell from http://www.agner.org/
3051     { ISD::FSQRT,      MVT::v8f32,  14 }, // Haswell from http://www.agner.org/
3052     { ISD::FSQRT,      MVT::f64,    14 }, // Haswell from http://www.agner.org/
3053     { ISD::FSQRT,      MVT::v2f64,  14 }, // Haswell from http://www.agner.org/
3054     { ISD::FSQRT,      MVT::v4f64,  28 }, // Haswell from http://www.agner.org/
3055   };
3056   static const CostTblEntry AVX1CostTbl[] = {
3057     { ISD::ABS,        MVT::v4i64,   5 }, // VBLENDVPD(X,VPSUBQ(0,X),X)
3058     { ISD::ABS,        MVT::v8i32,   3 },
3059     { ISD::ABS,        MVT::v16i16,  3 },
3060     { ISD::ABS,        MVT::v32i8,   3 },
3061     { ISD::BITREVERSE, MVT::v4i64,  12 }, // 2 x 128-bit Op + extract/insert
3062     { ISD::BITREVERSE, MVT::v8i32,  12 }, // 2 x 128-bit Op + extract/insert
3063     { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert
3064     { ISD::BITREVERSE, MVT::v32i8,  12 }, // 2 x 128-bit Op + extract/insert
3065     { ISD::BSWAP,      MVT::v4i64,   4 },
3066     { ISD::BSWAP,      MVT::v8i32,   4 },
3067     { ISD::BSWAP,      MVT::v16i16,  4 },
3068     { ISD::CTLZ,       MVT::v4i64,  48 }, // 2 x 128-bit Op + extract/insert
3069     { ISD::CTLZ,       MVT::v8i32,  38 }, // 2 x 128-bit Op + extract/insert
3070     { ISD::CTLZ,       MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert
3071     { ISD::CTLZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
3072     { ISD::CTPOP,      MVT::v4i64,  16 }, // 2 x 128-bit Op + extract/insert
3073     { ISD::CTPOP,      MVT::v8i32,  24 }, // 2 x 128-bit Op + extract/insert
3074     { ISD::CTPOP,      MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert
3075     { ISD::CTPOP,      MVT::v32i8,  14 }, // 2 x 128-bit Op + extract/insert
3076     { ISD::CTTZ,       MVT::v4i64,  22 }, // 2 x 128-bit Op + extract/insert
3077     { ISD::CTTZ,       MVT::v8i32,  30 }, // 2 x 128-bit Op + extract/insert
3078     { ISD::CTTZ,       MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert
3079     { ISD::CTTZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
3080     { ISD::SADDSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3081     { ISD::SADDSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3082     { ISD::SMAX,       MVT::v8i32,   4 }, // 2 x 128-bit Op + extract/insert
3083     { ISD::SMAX,       MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3084     { ISD::SMAX,       MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3085     { ISD::SMIN,       MVT::v8i32,   4 }, // 2 x 128-bit Op + extract/insert
3086     { ISD::SMIN,       MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3087     { ISD::SMIN,       MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3088     { ISD::SSUBSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3089     { ISD::SSUBSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3090     { ISD::UADDSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3091     { ISD::UADDSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3092     { ISD::UADDSAT,    MVT::v8i32,   8 }, // 2 x 128-bit Op + extract/insert
3093     { ISD::UMAX,       MVT::v8i32,   4 }, // 2 x 128-bit Op + extract/insert
3094     { ISD::UMAX,       MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3095     { ISD::UMAX,       MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3096     { ISD::UMIN,       MVT::v8i32,   4 }, // 2 x 128-bit Op + extract/insert
3097     { ISD::UMIN,       MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3098     { ISD::UMIN,       MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3099     { ISD::USUBSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3100     { ISD::USUBSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3101     { ISD::USUBSAT,    MVT::v8i32,   6 }, // 2 x 128-bit Op + extract/insert
3102     { ISD::FMAXNUM,    MVT::f32,     3 }, // MAXSS + CMPUNORDSS + BLENDVPS
3103     { ISD::FMAXNUM,    MVT::v4f32,   3 }, // MAXPS + CMPUNORDPS + BLENDVPS
3104     { ISD::FMAXNUM,    MVT::v8f32,   5 }, // MAXPS + CMPUNORDPS + BLENDVPS + ?
3105     { ISD::FMAXNUM,    MVT::f64,     3 }, // MAXSD + CMPUNORDSD + BLENDVPD
3106     { ISD::FMAXNUM,    MVT::v2f64,   3 }, // MAXPD + CMPUNORDPD + BLENDVPD
3107     { ISD::FMAXNUM,    MVT::v4f64,   5 }, // MAXPD + CMPUNORDPD + BLENDVPD + ?
3108     { ISD::FSQRT,      MVT::f32,    14 }, // SNB from http://www.agner.org/
3109     { ISD::FSQRT,      MVT::v4f32,  14 }, // SNB from http://www.agner.org/
3110     { ISD::FSQRT,      MVT::v8f32,  28 }, // SNB from http://www.agner.org/
3111     { ISD::FSQRT,      MVT::f64,    21 }, // SNB from http://www.agner.org/
3112     { ISD::FSQRT,      MVT::v2f64,  21 }, // SNB from http://www.agner.org/
3113     { ISD::FSQRT,      MVT::v4f64,  43 }, // SNB from http://www.agner.org/
3114   };
3115   static const CostTblEntry GLMCostTbl[] = {
3116     { ISD::FSQRT, MVT::f32,   19 }, // sqrtss
3117     { ISD::FSQRT, MVT::v4f32, 37 }, // sqrtps
3118     { ISD::FSQRT, MVT::f64,   34 }, // sqrtsd
3119     { ISD::FSQRT, MVT::v2f64, 67 }, // sqrtpd
3120   };
3121   static const CostTblEntry SLMCostTbl[] = {
3122     { ISD::FSQRT, MVT::f32,   20 }, // sqrtss
3123     { ISD::FSQRT, MVT::v4f32, 40 }, // sqrtps
3124     { ISD::FSQRT, MVT::f64,   35 }, // sqrtsd
3125     { ISD::FSQRT, MVT::v2f64, 70 }, // sqrtpd
3126   };
3127   static const CostTblEntry SSE42CostTbl[] = {
3128     { ISD::USUBSAT,    MVT::v4i32,   2 }, // pmaxud + psubd
3129     { ISD::UADDSAT,    MVT::v4i32,   3 }, // not + pminud + paddd
3130     { ISD::FSQRT,      MVT::f32,    18 }, // Nehalem from http://www.agner.org/
3131     { ISD::FSQRT,      MVT::v4f32,  18 }, // Nehalem from http://www.agner.org/
3132   };
3133   static const CostTblEntry SSE41CostTbl[] = {
3134     { ISD::ABS,        MVT::v2i64,   2 }, // BLENDVPD(X,PSUBQ(0,X),X)
3135     { ISD::SMAX,       MVT::v4i32,   1 },
3136     { ISD::SMAX,       MVT::v16i8,   1 },
3137     { ISD::SMIN,       MVT::v4i32,   1 },
3138     { ISD::SMIN,       MVT::v16i8,   1 },
3139     { ISD::UMAX,       MVT::v4i32,   1 },
3140     { ISD::UMAX,       MVT::v8i16,   1 },
3141     { ISD::UMIN,       MVT::v4i32,   1 },
3142     { ISD::UMIN,       MVT::v8i16,   1 },
3143   };
3144   static const CostTblEntry SSSE3CostTbl[] = {
3145     { ISD::ABS,        MVT::v4i32,   1 },
3146     { ISD::ABS,        MVT::v8i16,   1 },
3147     { ISD::ABS,        MVT::v16i8,   1 },
3148     { ISD::BITREVERSE, MVT::v2i64,   5 },
3149     { ISD::BITREVERSE, MVT::v4i32,   5 },
3150     { ISD::BITREVERSE, MVT::v8i16,   5 },
3151     { ISD::BITREVERSE, MVT::v16i8,   5 },
3152     { ISD::BSWAP,      MVT::v2i64,   1 },
3153     { ISD::BSWAP,      MVT::v4i32,   1 },
3154     { ISD::BSWAP,      MVT::v8i16,   1 },
3155     { ISD::CTLZ,       MVT::v2i64,  23 },
3156     { ISD::CTLZ,       MVT::v4i32,  18 },
3157     { ISD::CTLZ,       MVT::v8i16,  14 },
3158     { ISD::CTLZ,       MVT::v16i8,   9 },
3159     { ISD::CTPOP,      MVT::v2i64,   7 },
3160     { ISD::CTPOP,      MVT::v4i32,  11 },
3161     { ISD::CTPOP,      MVT::v8i16,   9 },
3162     { ISD::CTPOP,      MVT::v16i8,   6 },
3163     { ISD::CTTZ,       MVT::v2i64,  10 },
3164     { ISD::CTTZ,       MVT::v4i32,  14 },
3165     { ISD::CTTZ,       MVT::v8i16,  12 },
3166     { ISD::CTTZ,       MVT::v16i8,   9 }
3167   };
3168   static const CostTblEntry SSE2CostTbl[] = {
3169     { ISD::ABS,        MVT::v2i64,   4 },
3170     { ISD::ABS,        MVT::v4i32,   3 },
3171     { ISD::ABS,        MVT::v8i16,   2 },
3172     { ISD::ABS,        MVT::v16i8,   2 },
3173     { ISD::BITREVERSE, MVT::v2i64,  29 },
3174     { ISD::BITREVERSE, MVT::v4i32,  27 },
3175     { ISD::BITREVERSE, MVT::v8i16,  27 },
3176     { ISD::BITREVERSE, MVT::v16i8,  20 },
3177     { ISD::BSWAP,      MVT::v2i64,   7 },
3178     { ISD::BSWAP,      MVT::v4i32,   7 },
3179     { ISD::BSWAP,      MVT::v8i16,   7 },
3180     { ISD::CTLZ,       MVT::v2i64,  25 },
3181     { ISD::CTLZ,       MVT::v4i32,  26 },
3182     { ISD::CTLZ,       MVT::v8i16,  20 },
3183     { ISD::CTLZ,       MVT::v16i8,  17 },
3184     { ISD::CTPOP,      MVT::v2i64,  12 },
3185     { ISD::CTPOP,      MVT::v4i32,  15 },
3186     { ISD::CTPOP,      MVT::v8i16,  13 },
3187     { ISD::CTPOP,      MVT::v16i8,  10 },
3188     { ISD::CTTZ,       MVT::v2i64,  14 },
3189     { ISD::CTTZ,       MVT::v4i32,  18 },
3190     { ISD::CTTZ,       MVT::v8i16,  16 },
3191     { ISD::CTTZ,       MVT::v16i8,  13 },
3192     { ISD::SADDSAT,    MVT::v8i16,   1 },
3193     { ISD::SADDSAT,    MVT::v16i8,   1 },
3194     { ISD::SMAX,       MVT::v8i16,   1 },
3195     { ISD::SMIN,       MVT::v8i16,   1 },
3196     { ISD::SSUBSAT,    MVT::v8i16,   1 },
3197     { ISD::SSUBSAT,    MVT::v16i8,   1 },
3198     { ISD::UADDSAT,    MVT::v8i16,   1 },
3199     { ISD::UADDSAT,    MVT::v16i8,   1 },
3200     { ISD::UMAX,       MVT::v8i16,   2 },
3201     { ISD::UMAX,       MVT::v16i8,   1 },
3202     { ISD::UMIN,       MVT::v8i16,   2 },
3203     { ISD::UMIN,       MVT::v16i8,   1 },
3204     { ISD::USUBSAT,    MVT::v8i16,   1 },
3205     { ISD::USUBSAT,    MVT::v16i8,   1 },
3206     { ISD::FMAXNUM,    MVT::f64,     4 },
3207     { ISD::FMAXNUM,    MVT::v2f64,   4 },
3208     { ISD::FSQRT,      MVT::f64,    32 }, // Nehalem from http://www.agner.org/
3209     { ISD::FSQRT,      MVT::v2f64,  32 }, // Nehalem from http://www.agner.org/
3210   };
3211   static const CostTblEntry SSE1CostTbl[] = {
3212     { ISD::FMAXNUM,    MVT::f32,     4 },
3213     { ISD::FMAXNUM,    MVT::v4f32,   4 },
3214     { ISD::FSQRT,      MVT::f32,    28 }, // Pentium III from http://www.agner.org/
3215     { ISD::FSQRT,      MVT::v4f32,  56 }, // Pentium III from http://www.agner.org/
3216   };
3217   static const CostTblEntry BMI64CostTbl[] = { // 64-bit targets
3218     { ISD::CTTZ,       MVT::i64,     1 },
3219   };
3220   static const CostTblEntry BMI32CostTbl[] = { // 32 or 64-bit targets
3221     { ISD::CTTZ,       MVT::i32,     1 },
3222     { ISD::CTTZ,       MVT::i16,     1 },
3223     { ISD::CTTZ,       MVT::i8,      1 },
3224   };
3225   static const CostTblEntry LZCNT64CostTbl[] = { // 64-bit targets
3226     { ISD::CTLZ,       MVT::i64,     1 },
3227   };
3228   static const CostTblEntry LZCNT32CostTbl[] = { // 32 or 64-bit targets
3229     { ISD::CTLZ,       MVT::i32,     1 },
3230     { ISD::CTLZ,       MVT::i16,     1 },
3231     { ISD::CTLZ,       MVT::i8,      1 },
3232   };
3233   static const CostTblEntry POPCNT64CostTbl[] = { // 64-bit targets
3234     { ISD::CTPOP,      MVT::i64,     1 },
3235   };
3236   static const CostTblEntry POPCNT32CostTbl[] = { // 32 or 64-bit targets
3237     { ISD::CTPOP,      MVT::i32,     1 },
3238     { ISD::CTPOP,      MVT::i16,     1 },
3239     { ISD::CTPOP,      MVT::i8,      1 },
3240   };
3241   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
3242     { ISD::ABS,        MVT::i64,     2 }, // SUB+CMOV
3243     { ISD::BITREVERSE, MVT::i64,    14 },
3244     { ISD::BSWAP,      MVT::i64,     1 },
3245     { ISD::CTLZ,       MVT::i64,     4 }, // BSR+XOR or BSR+XOR+CMOV
3246     { ISD::CTTZ,       MVT::i64,     3 }, // TEST+BSF+CMOV/BRANCH
3247     { ISD::CTPOP,      MVT::i64,    10 },
3248     { ISD::SADDO,      MVT::i64,     1 },
3249     { ISD::UADDO,      MVT::i64,     1 },
3250     { ISD::UMULO,      MVT::i64,     2 }, // mulq + seto
3251   };
3252   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
3253     { ISD::ABS,        MVT::i32,     2 }, // SUB+CMOV
3254     { ISD::ABS,        MVT::i16,     2 }, // SUB+CMOV
3255     { ISD::BITREVERSE, MVT::i32,    14 },
3256     { ISD::BITREVERSE, MVT::i16,    14 },
3257     { ISD::BITREVERSE, MVT::i8,     11 },
3258     { ISD::BSWAP,      MVT::i32,     1 },
3259     { ISD::BSWAP,      MVT::i16,     1 }, // ROL
3260     { ISD::CTLZ,       MVT::i32,     4 }, // BSR+XOR or BSR+XOR+CMOV
3261     { ISD::CTLZ,       MVT::i16,     4 }, // BSR+XOR or BSR+XOR+CMOV
3262     { ISD::CTLZ,       MVT::i8,      4 }, // BSR+XOR or BSR+XOR+CMOV
3263     { ISD::CTTZ,       MVT::i32,     3 }, // TEST+BSF+CMOV/BRANCH
3264     { ISD::CTTZ,       MVT::i16,     3 }, // TEST+BSF+CMOV/BRANCH
3265     { ISD::CTTZ,       MVT::i8,      3 }, // TEST+BSF+CMOV/BRANCH
3266     { ISD::CTPOP,      MVT::i32,     8 },
3267     { ISD::CTPOP,      MVT::i16,     9 },
3268     { ISD::CTPOP,      MVT::i8,      7 },
3269     { ISD::SADDO,      MVT::i32,     1 },
3270     { ISD::SADDO,      MVT::i16,     1 },
3271     { ISD::SADDO,      MVT::i8,      1 },
3272     { ISD::UADDO,      MVT::i32,     1 },
3273     { ISD::UADDO,      MVT::i16,     1 },
3274     { ISD::UADDO,      MVT::i8,      1 },
3275     { ISD::UMULO,      MVT::i32,     2 }, // mul + seto
3276     { ISD::UMULO,      MVT::i16,     2 },
3277     { ISD::UMULO,      MVT::i8,      2 },
3278   };
3279 
3280   Type *RetTy = ICA.getReturnType();
3281   Type *OpTy = RetTy;
3282   Intrinsic::ID IID = ICA.getID();
3283   unsigned ISD = ISD::DELETED_NODE;
3284   switch (IID) {
3285   default:
3286     break;
3287   case Intrinsic::abs:
3288     ISD = ISD::ABS;
3289     break;
3290   case Intrinsic::bitreverse:
3291     ISD = ISD::BITREVERSE;
3292     break;
3293   case Intrinsic::bswap:
3294     ISD = ISD::BSWAP;
3295     break;
3296   case Intrinsic::ctlz:
3297     ISD = ISD::CTLZ;
3298     break;
3299   case Intrinsic::ctpop:
3300     ISD = ISD::CTPOP;
3301     break;
3302   case Intrinsic::cttz:
3303     ISD = ISD::CTTZ;
3304     break;
3305   case Intrinsic::maxnum:
3306   case Intrinsic::minnum:
3307     // FMINNUM has same costs so don't duplicate.
3308     ISD = ISD::FMAXNUM;
3309     break;
3310   case Intrinsic::sadd_sat:
3311     ISD = ISD::SADDSAT;
3312     break;
3313   case Intrinsic::smax:
3314     ISD = ISD::SMAX;
3315     break;
3316   case Intrinsic::smin:
3317     ISD = ISD::SMIN;
3318     break;
3319   case Intrinsic::ssub_sat:
3320     ISD = ISD::SSUBSAT;
3321     break;
3322   case Intrinsic::uadd_sat:
3323     ISD = ISD::UADDSAT;
3324     break;
3325   case Intrinsic::umax:
3326     ISD = ISD::UMAX;
3327     break;
3328   case Intrinsic::umin:
3329     ISD = ISD::UMIN;
3330     break;
3331   case Intrinsic::usub_sat:
3332     ISD = ISD::USUBSAT;
3333     break;
3334   case Intrinsic::sqrt:
3335     ISD = ISD::FSQRT;
3336     break;
3337   case Intrinsic::sadd_with_overflow:
3338   case Intrinsic::ssub_with_overflow:
3339     // SSUBO has same costs so don't duplicate.
3340     ISD = ISD::SADDO;
3341     OpTy = RetTy->getContainedType(0);
3342     break;
3343   case Intrinsic::uadd_with_overflow:
3344   case Intrinsic::usub_with_overflow:
3345     // USUBO has same costs so don't duplicate.
3346     ISD = ISD::UADDO;
3347     OpTy = RetTy->getContainedType(0);
3348     break;
3349   case Intrinsic::umul_with_overflow:
3350   case Intrinsic::smul_with_overflow:
3351     // SMULO has same costs so don't duplicate.
3352     ISD = ISD::UMULO;
3353     OpTy = RetTy->getContainedType(0);
3354     break;
3355   }
3356 
3357   if (ISD != ISD::DELETED_NODE) {
3358     // Legalize the type.
3359     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, OpTy);
3360     MVT MTy = LT.second;
3361 
3362     // Attempt to lookup cost.
3363     if (ISD == ISD::BITREVERSE && ST->hasGFNI() && ST->hasSSSE3() &&
3364         MTy.isVector()) {
3365       // With PSHUFB the code is very similar for all types. If we have integer
3366       // byte operations, we just need a GF2P8AFFINEQB for vXi8. For other types
3367       // we also need a PSHUFB.
3368       unsigned Cost = MTy.getVectorElementType() == MVT::i8 ? 1 : 2;
3369 
3370       // Without byte operations, we need twice as many GF2P8AFFINEQB and PSHUFB
3371       // instructions. We also need an extract and an insert.
3372       if (!(MTy.is128BitVector() || (ST->hasAVX2() && MTy.is256BitVector()) ||
3373             (ST->hasBWI() && MTy.is512BitVector())))
3374         Cost = Cost * 2 + 2;
3375 
3376       return LT.first * Cost;
3377     }
3378 
3379     auto adjustTableCost = [](const CostTblEntry &Entry,
3380                               InstructionCost LegalizationCost,
3381                               FastMathFlags FMF) {
3382       // If there are no NANs to deal with, then these are reduced to a
3383       // single MIN** or MAX** instruction instead of the MIN/CMP/SELECT that we
3384       // assume is used in the non-fast case.
3385       if (Entry.ISD == ISD::FMAXNUM || Entry.ISD == ISD::FMINNUM) {
3386         if (FMF.noNaNs())
3387           return LegalizationCost * 1;
3388       }
3389       return LegalizationCost * (int)Entry.Cost;
3390     };
3391 
3392     if (ST->useGLMDivSqrtCosts())
3393       if (const auto *Entry = CostTableLookup(GLMCostTbl, ISD, MTy))
3394         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3395 
3396     if (ST->useSLMArithCosts())
3397       if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy))
3398         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3399 
3400     if (ST->hasBITALG())
3401       if (const auto *Entry = CostTableLookup(AVX512BITALGCostTbl, ISD, MTy))
3402         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3403 
3404     if (ST->hasVPOPCNTDQ())
3405       if (const auto *Entry = CostTableLookup(AVX512VPOPCNTDQCostTbl, ISD, MTy))
3406         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3407 
3408     if (ST->hasCDI())
3409       if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy))
3410         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3411 
3412     if (ST->hasBWI())
3413       if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
3414         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3415 
3416     if (ST->hasAVX512())
3417       if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
3418         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3419 
3420     if (ST->hasXOP())
3421       if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
3422         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3423 
3424     if (ST->hasAVX2())
3425       if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
3426         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3427 
3428     if (ST->hasAVX())
3429       if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
3430         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3431 
3432     if (ST->hasSSE42())
3433       if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
3434         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3435 
3436     if (ST->hasSSE41())
3437       if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
3438         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3439 
3440     if (ST->hasSSSE3())
3441       if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy))
3442         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3443 
3444     if (ST->hasSSE2())
3445       if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
3446         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3447 
3448     if (ST->hasSSE1())
3449       if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
3450         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3451 
3452     if (ST->hasBMI()) {
3453       if (ST->is64Bit())
3454         if (const auto *Entry = CostTableLookup(BMI64CostTbl, ISD, MTy))
3455           return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3456 
3457       if (const auto *Entry = CostTableLookup(BMI32CostTbl, ISD, MTy))
3458         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3459     }
3460 
3461     if (ST->hasLZCNT()) {
3462       if (ST->is64Bit())
3463         if (const auto *Entry = CostTableLookup(LZCNT64CostTbl, ISD, MTy))
3464           return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3465 
3466       if (const auto *Entry = CostTableLookup(LZCNT32CostTbl, ISD, MTy))
3467         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3468     }
3469 
3470     if (ST->hasPOPCNT()) {
3471       if (ST->is64Bit())
3472         if (const auto *Entry = CostTableLookup(POPCNT64CostTbl, ISD, MTy))
3473           return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3474 
3475       if (const auto *Entry = CostTableLookup(POPCNT32CostTbl, ISD, MTy))
3476         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3477     }
3478 
3479     if (ISD == ISD::BSWAP && ST->hasMOVBE() && ST->hasFastMOVBE()) {
3480       if (const Instruction *II = ICA.getInst()) {
3481         if (II->hasOneUse() && isa<StoreInst>(II->user_back()))
3482           return TTI::TCC_Free;
3483         if (auto *LI = dyn_cast<LoadInst>(II->getOperand(0))) {
3484           if (LI->hasOneUse())
3485             return TTI::TCC_Free;
3486         }
3487       }
3488     }
3489 
3490     if (ST->is64Bit())
3491       if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
3492         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3493 
3494     if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
3495       return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3496   }
3497 
3498   return BaseT::getIntrinsicInstrCost(ICA, CostKind);
3499 }
3500 
3501 InstructionCost
3502 X86TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
3503                                   TTI::TargetCostKind CostKind) {
3504   if (ICA.isTypeBasedOnly())
3505     return getTypeBasedIntrinsicInstrCost(ICA, CostKind);
3506 
3507   static const CostTblEntry AVX512BWCostTbl[] = {
3508     { ISD::ROTL,       MVT::v32i16,  2 },
3509     { ISD::ROTL,       MVT::v16i16,  2 },
3510     { ISD::ROTL,       MVT::v8i16,   2 },
3511     { ISD::ROTL,       MVT::v64i8,   5 },
3512     { ISD::ROTL,       MVT::v32i8,   5 },
3513     { ISD::ROTL,       MVT::v16i8,   5 },
3514     { ISD::ROTR,       MVT::v32i16,  2 },
3515     { ISD::ROTR,       MVT::v16i16,  2 },
3516     { ISD::ROTR,       MVT::v8i16,   2 },
3517     { ISD::ROTR,       MVT::v64i8,   5 },
3518     { ISD::ROTR,       MVT::v32i8,   5 },
3519     { ISD::ROTR,       MVT::v16i8,   5 }
3520   };
3521   static const CostTblEntry AVX512CostTbl[] = {
3522     { ISD::ROTL,       MVT::v8i64,   1 },
3523     { ISD::ROTL,       MVT::v4i64,   1 },
3524     { ISD::ROTL,       MVT::v2i64,   1 },
3525     { ISD::ROTL,       MVT::v16i32,  1 },
3526     { ISD::ROTL,       MVT::v8i32,   1 },
3527     { ISD::ROTL,       MVT::v4i32,   1 },
3528     { ISD::ROTR,       MVT::v8i64,   1 },
3529     { ISD::ROTR,       MVT::v4i64,   1 },
3530     { ISD::ROTR,       MVT::v2i64,   1 },
3531     { ISD::ROTR,       MVT::v16i32,  1 },
3532     { ISD::ROTR,       MVT::v8i32,   1 },
3533     { ISD::ROTR,       MVT::v4i32,   1 }
3534   };
3535   // XOP: ROTL = VPROT(X,Y), ROTR = VPROT(X,SUB(0,Y))
3536   static const CostTblEntry XOPCostTbl[] = {
3537     { ISD::ROTL,       MVT::v4i64,   4 },
3538     { ISD::ROTL,       MVT::v8i32,   4 },
3539     { ISD::ROTL,       MVT::v16i16,  4 },
3540     { ISD::ROTL,       MVT::v32i8,   4 },
3541     { ISD::ROTL,       MVT::v2i64,   1 },
3542     { ISD::ROTL,       MVT::v4i32,   1 },
3543     { ISD::ROTL,       MVT::v8i16,   1 },
3544     { ISD::ROTL,       MVT::v16i8,   1 },
3545     { ISD::ROTR,       MVT::v4i64,   6 },
3546     { ISD::ROTR,       MVT::v8i32,   6 },
3547     { ISD::ROTR,       MVT::v16i16,  6 },
3548     { ISD::ROTR,       MVT::v32i8,   6 },
3549     { ISD::ROTR,       MVT::v2i64,   2 },
3550     { ISD::ROTR,       MVT::v4i32,   2 },
3551     { ISD::ROTR,       MVT::v8i16,   2 },
3552     { ISD::ROTR,       MVT::v16i8,   2 }
3553   };
3554   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
3555     { ISD::ROTL,       MVT::i64,     1 },
3556     { ISD::ROTR,       MVT::i64,     1 },
3557     { ISD::FSHL,       MVT::i64,     4 }
3558   };
3559   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
3560     { ISD::ROTL,       MVT::i32,     1 },
3561     { ISD::ROTL,       MVT::i16,     1 },
3562     { ISD::ROTL,       MVT::i8,      1 },
3563     { ISD::ROTR,       MVT::i32,     1 },
3564     { ISD::ROTR,       MVT::i16,     1 },
3565     { ISD::ROTR,       MVT::i8,      1 },
3566     { ISD::FSHL,       MVT::i32,     4 },
3567     { ISD::FSHL,       MVT::i16,     4 },
3568     { ISD::FSHL,       MVT::i8,      4 }
3569   };
3570 
3571   Intrinsic::ID IID = ICA.getID();
3572   Type *RetTy = ICA.getReturnType();
3573   const SmallVectorImpl<const Value *> &Args = ICA.getArgs();
3574   unsigned ISD = ISD::DELETED_NODE;
3575   switch (IID) {
3576   default:
3577     break;
3578   case Intrinsic::fshl:
3579     ISD = ISD::FSHL;
3580     if (Args[0] == Args[1])
3581       ISD = ISD::ROTL;
3582     break;
3583   case Intrinsic::fshr:
3584     // FSHR has same costs so don't duplicate.
3585     ISD = ISD::FSHL;
3586     if (Args[0] == Args[1])
3587       ISD = ISD::ROTR;
3588     break;
3589   }
3590 
3591   if (ISD != ISD::DELETED_NODE) {
3592     // Legalize the type.
3593     std::pair<InstructionCost, MVT> LT =
3594         TLI->getTypeLegalizationCost(DL, RetTy);
3595     MVT MTy = LT.second;
3596 
3597     // Attempt to lookup cost.
3598     if (ST->hasBWI())
3599       if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
3600         return LT.first * Entry->Cost;
3601 
3602     if (ST->hasAVX512())
3603       if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
3604         return LT.first * Entry->Cost;
3605 
3606     if (ST->hasXOP())
3607       if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
3608         return LT.first * Entry->Cost;
3609 
3610     if (ST->is64Bit())
3611       if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
3612         return LT.first * Entry->Cost;
3613 
3614     if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
3615       return LT.first * Entry->Cost;
3616   }
3617 
3618   return BaseT::getIntrinsicInstrCost(ICA, CostKind);
3619 }
3620 
3621 InstructionCost X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
3622                                                unsigned Index) {
3623   static const CostTblEntry SLMCostTbl[] = {
3624      { ISD::EXTRACT_VECTOR_ELT,       MVT::i8,      4 },
3625      { ISD::EXTRACT_VECTOR_ELT,       MVT::i16,     4 },
3626      { ISD::EXTRACT_VECTOR_ELT,       MVT::i32,     4 },
3627      { ISD::EXTRACT_VECTOR_ELT,       MVT::i64,     7 }
3628    };
3629 
3630   assert(Val->isVectorTy() && "This must be a vector type");
3631   Type *ScalarType = Val->getScalarType();
3632   int RegisterFileMoveCost = 0;
3633 
3634   // Non-immediate extraction/insertion can be handled as a sequence of
3635   // aliased loads+stores via the stack.
3636   if (Index == -1U && (Opcode == Instruction::ExtractElement ||
3637                        Opcode == Instruction::InsertElement)) {
3638     // TODO: On some SSE41+ targets, we expand to cmp+splat+select patterns:
3639     // inselt N0, N1, N2 --> select (SplatN2 == {0,1,2...}) ? SplatN1 : N0.
3640 
3641     // TODO: Move this to BasicTTIImpl.h? We'd need better gep + index handling.
3642     assert(isa<FixedVectorType>(Val) && "Fixed vector type expected");
3643     Align VecAlign = DL.getPrefTypeAlign(Val);
3644     Align SclAlign = DL.getPrefTypeAlign(ScalarType);
3645 
3646     // Extract - store vector to stack, load scalar.
3647     if (Opcode == Instruction::ExtractElement) {
3648       return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0,
3649                              TTI::TargetCostKind::TCK_RecipThroughput) +
3650              getMemoryOpCost(Instruction::Load, ScalarType, SclAlign, 0,
3651                              TTI::TargetCostKind::TCK_RecipThroughput);
3652     }
3653     // Insert - store vector to stack, store scalar, load vector.
3654     if (Opcode == Instruction::InsertElement) {
3655       return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0,
3656                              TTI::TargetCostKind::TCK_RecipThroughput) +
3657              getMemoryOpCost(Instruction::Store, ScalarType, SclAlign, 0,
3658                              TTI::TargetCostKind::TCK_RecipThroughput) +
3659              getMemoryOpCost(Instruction::Load, Val, VecAlign, 0,
3660                              TTI::TargetCostKind::TCK_RecipThroughput);
3661     }
3662   }
3663 
3664   if (Index != -1U && (Opcode == Instruction::ExtractElement ||
3665                        Opcode == Instruction::InsertElement)) {
3666     // Extraction of vXi1 elements are now efficiently handled by MOVMSK.
3667     if (Opcode == Instruction::ExtractElement &&
3668         ScalarType->getScalarSizeInBits() == 1 &&
3669         cast<FixedVectorType>(Val)->getNumElements() > 1)
3670       return 1;
3671 
3672     // Legalize the type.
3673     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Val);
3674 
3675     // This type is legalized to a scalar type.
3676     if (!LT.second.isVector())
3677       return 0;
3678 
3679     // The type may be split. Normalize the index to the new type.
3680     unsigned SizeInBits = LT.second.getSizeInBits();
3681     unsigned NumElts = LT.second.getVectorNumElements();
3682     unsigned SubNumElts = NumElts;
3683     Index = Index % NumElts;
3684 
3685     // For >128-bit vectors, we need to extract higher 128-bit subvectors.
3686     // For inserts, we also need to insert the subvector back.
3687     if (SizeInBits > 128) {
3688       assert((SizeInBits % 128) == 0 && "Illegal vector");
3689       unsigned NumSubVecs = SizeInBits / 128;
3690       SubNumElts = NumElts / NumSubVecs;
3691       if (SubNumElts <= Index) {
3692         RegisterFileMoveCost += (Opcode == Instruction::InsertElement ? 2 : 1);
3693         Index %= SubNumElts;
3694       }
3695     }
3696 
3697     if (Index == 0) {
3698       // Floating point scalars are already located in index #0.
3699       // Many insertions to #0 can fold away for scalar fp-ops, so let's assume
3700       // true for all.
3701       if (ScalarType->isFloatingPointTy())
3702         return RegisterFileMoveCost;
3703 
3704       // Assume movd/movq XMM -> GPR is relatively cheap on all targets.
3705       if (ScalarType->isIntegerTy() && Opcode == Instruction::ExtractElement)
3706         return 1 + RegisterFileMoveCost;
3707     }
3708 
3709     int ISD = TLI->InstructionOpcodeToISD(Opcode);
3710     assert(ISD && "Unexpected vector opcode");
3711     MVT MScalarTy = LT.second.getScalarType();
3712     if (ST->useSLMArithCosts())
3713       if (auto *Entry = CostTableLookup(SLMCostTbl, ISD, MScalarTy))
3714         return Entry->Cost + RegisterFileMoveCost;
3715 
3716     // Assume pinsr/pextr XMM <-> GPR is relatively cheap on all targets.
3717     if ((MScalarTy == MVT::i16 && ST->hasSSE2()) ||
3718         (MScalarTy.isInteger() && ST->hasSSE41()))
3719       return 1 + RegisterFileMoveCost;
3720 
3721     // Assume insertps is relatively cheap on all targets.
3722     if (MScalarTy == MVT::f32 && ST->hasSSE41() &&
3723         Opcode == Instruction::InsertElement)
3724       return 1 + RegisterFileMoveCost;
3725 
3726     // For extractions we just need to shuffle the element to index 0, which
3727     // should be very cheap (assume cost = 1). For insertions we need to shuffle
3728     // the elements to its destination. In both cases we must handle the
3729     // subvector move(s).
3730     // If the vector type is already less than 128-bits then don't reduce it.
3731     // TODO: Under what circumstances should we shuffle using the full width?
3732     InstructionCost ShuffleCost = 1;
3733     if (Opcode == Instruction::InsertElement) {
3734       auto *SubTy = cast<VectorType>(Val);
3735       EVT VT = TLI->getValueType(DL, Val);
3736       if (VT.getScalarType() != MScalarTy || VT.getSizeInBits() >= 128)
3737         SubTy = FixedVectorType::get(ScalarType, SubNumElts);
3738       ShuffleCost =
3739           getShuffleCost(TTI::SK_PermuteTwoSrc, SubTy, None, 0, SubTy);
3740     }
3741     int IntOrFpCost = ScalarType->isFloatingPointTy() ? 0 : 1;
3742     return ShuffleCost + IntOrFpCost + RegisterFileMoveCost;
3743   }
3744 
3745   // Add to the base cost if we know that the extracted element of a vector is
3746   // destined to be moved to and used in the integer register file.
3747   if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy())
3748     RegisterFileMoveCost += 1;
3749 
3750   return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost;
3751 }
3752 
3753 InstructionCost X86TTIImpl::getScalarizationOverhead(VectorType *Ty,
3754                                                      const APInt &DemandedElts,
3755                                                      bool Insert,
3756                                                      bool Extract) {
3757   InstructionCost Cost = 0;
3758 
3759   // For insertions, a ISD::BUILD_VECTOR style vector initialization can be much
3760   // cheaper than an accumulation of ISD::INSERT_VECTOR_ELT.
3761   if (Insert) {
3762     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
3763     MVT MScalarTy = LT.second.getScalarType();
3764     unsigned SizeInBits = LT.second.getSizeInBits();
3765 
3766     if ((MScalarTy == MVT::i16 && ST->hasSSE2()) ||
3767         (MScalarTy.isInteger() && ST->hasSSE41()) ||
3768         (MScalarTy == MVT::f32 && ST->hasSSE41())) {
3769       // For types we can insert directly, insertion into 128-bit sub vectors is
3770       // cheap, followed by a cheap chain of concatenations.
3771       if (SizeInBits <= 128) {
3772         Cost +=
3773             BaseT::getScalarizationOverhead(Ty, DemandedElts, Insert, false);
3774       } else {
3775         // In each 128-lane, if at least one index is demanded but not all
3776         // indices are demanded and this 128-lane is not the first 128-lane of
3777         // the legalized-vector, then this 128-lane needs a extracti128; If in
3778         // each 128-lane, there is at least one demanded index, this 128-lane
3779         // needs a inserti128.
3780 
3781         // The following cases will help you build a better understanding:
3782         // Assume we insert several elements into a v8i32 vector in avx2,
3783         // Case#1: inserting into 1th index needs vpinsrd + inserti128.
3784         // Case#2: inserting into 5th index needs extracti128 + vpinsrd +
3785         // inserti128.
3786         // Case#3: inserting into 4,5,6,7 index needs 4*vpinsrd + inserti128.
3787         const int CostValue = *LT.first.getValue();
3788         assert(CostValue >= 0 && "Negative cost!");
3789         unsigned Num128Lanes = SizeInBits / 128 * CostValue;
3790         unsigned NumElts = LT.second.getVectorNumElements() * CostValue;
3791         APInt WidenedDemandedElts = DemandedElts.zextOrSelf(NumElts);
3792         unsigned Scale = NumElts / Num128Lanes;
3793         // We iterate each 128-lane, and check if we need a
3794         // extracti128/inserti128 for this 128-lane.
3795         for (unsigned I = 0; I < NumElts; I += Scale) {
3796           APInt Mask = WidenedDemandedElts.getBitsSet(NumElts, I, I + Scale);
3797           APInt MaskedDE = Mask & WidenedDemandedElts;
3798           unsigned Population = MaskedDE.countPopulation();
3799           Cost += (Population > 0 && Population != Scale &&
3800                    I % LT.second.getVectorNumElements() != 0);
3801           Cost += Population > 0;
3802         }
3803         Cost += DemandedElts.countPopulation();
3804 
3805         // For vXf32 cases, insertion into the 0'th index in each v4f32
3806         // 128-bit vector is free.
3807         // NOTE: This assumes legalization widens vXf32 vectors.
3808         if (MScalarTy == MVT::f32)
3809           for (unsigned i = 0, e = cast<FixedVectorType>(Ty)->getNumElements();
3810                i < e; i += 4)
3811             if (DemandedElts[i])
3812               Cost--;
3813       }
3814     } else if (LT.second.isVector()) {
3815       // Without fast insertion, we need to use MOVD/MOVQ to pass each demanded
3816       // integer element as a SCALAR_TO_VECTOR, then we build the vector as a
3817       // series of UNPCK followed by CONCAT_VECTORS - all of these can be
3818       // considered cheap.
3819       if (Ty->isIntOrIntVectorTy())
3820         Cost += DemandedElts.countPopulation();
3821 
3822       // Get the smaller of the legalized or original pow2-extended number of
3823       // vector elements, which represents the number of unpacks we'll end up
3824       // performing.
3825       unsigned NumElts = LT.second.getVectorNumElements();
3826       unsigned Pow2Elts =
3827           PowerOf2Ceil(cast<FixedVectorType>(Ty)->getNumElements());
3828       Cost += (std::min<unsigned>(NumElts, Pow2Elts) - 1) * LT.first;
3829     }
3830   }
3831 
3832   // TODO: Use default extraction for now, but we should investigate extending this
3833   // to handle repeated subvector extraction.
3834   if (Extract)
3835     Cost += BaseT::getScalarizationOverhead(Ty, DemandedElts, false, Extract);
3836 
3837   return Cost;
3838 }
3839 
3840 InstructionCost
3841 X86TTIImpl::getReplicationShuffleCost(Type *EltTy, int ReplicationFactor,
3842                                       int VF, const APInt &DemandedDstElts,
3843                                       TTI::TargetCostKind CostKind) {
3844   const unsigned EltTyBits = DL.getTypeSizeInBits(EltTy);
3845   // We don't differentiate element types here, only element bit width.
3846   EltTy = IntegerType::getIntNTy(EltTy->getContext(), EltTyBits);
3847 
3848   auto bailout = [&]() {
3849     return BaseT::getReplicationShuffleCost(EltTy, ReplicationFactor, VF,
3850                                             DemandedDstElts, CostKind);
3851   };
3852 
3853   // For now, only deal with AVX512 cases.
3854   if (!ST->hasAVX512())
3855     return bailout();
3856 
3857   // Do we have a native shuffle for this element type, or should we promote?
3858   unsigned PromEltTyBits = EltTyBits;
3859   switch (EltTyBits) {
3860   case 32:
3861   case 64:
3862     break; // AVX512F.
3863   case 16:
3864     if (!ST->hasBWI())
3865       PromEltTyBits = 32; // promote to i32, AVX512F.
3866     break;                // AVX512BW
3867   case 8:
3868     if (!ST->hasVBMI())
3869       PromEltTyBits = 32; // promote to i32, AVX512F.
3870     break;                // AVX512VBMI
3871   case 1:
3872     // There is no support for shuffling i1 elements. We *must* promote.
3873     if (ST->hasBWI()) {
3874       if (ST->hasVBMI())
3875         PromEltTyBits = 8; // promote to i8, AVX512VBMI.
3876       else
3877         PromEltTyBits = 16; // promote to i16, AVX512BW.
3878       break;
3879     }
3880     if (ST->hasDQI()) {
3881       PromEltTyBits = 32; // promote to i32, AVX512F.
3882       break;
3883     }
3884     return bailout();
3885   default:
3886     return bailout();
3887   }
3888   auto *PromEltTy = IntegerType::getIntNTy(EltTy->getContext(), PromEltTyBits);
3889 
3890   auto *SrcVecTy = FixedVectorType::get(EltTy, VF);
3891   auto *PromSrcVecTy = FixedVectorType::get(PromEltTy, VF);
3892 
3893   int NumDstElements = VF * ReplicationFactor;
3894   auto *PromDstVecTy = FixedVectorType::get(PromEltTy, NumDstElements);
3895   auto *DstVecTy = FixedVectorType::get(EltTy, NumDstElements);
3896 
3897   // Legalize the types.
3898   MVT LegalSrcVecTy = TLI->getTypeLegalizationCost(DL, SrcVecTy).second;
3899   MVT LegalPromSrcVecTy = TLI->getTypeLegalizationCost(DL, PromSrcVecTy).second;
3900   MVT LegalPromDstVecTy = TLI->getTypeLegalizationCost(DL, PromDstVecTy).second;
3901   MVT LegalDstVecTy = TLI->getTypeLegalizationCost(DL, DstVecTy).second;
3902   // They should have legalized into vector types.
3903   if (!LegalSrcVecTy.isVector() || !LegalPromSrcVecTy.isVector() ||
3904       !LegalPromDstVecTy.isVector() || !LegalDstVecTy.isVector())
3905     return bailout();
3906 
3907   if (PromEltTyBits != EltTyBits) {
3908     // If we have to perform the shuffle with wider elt type than our data type,
3909     // then we will first need to anyext (we don't care about the new bits)
3910     // the source elements, and then truncate Dst elements.
3911     InstructionCost PromotionCost;
3912     PromotionCost += getCastInstrCost(
3913         Instruction::SExt, /*Dst=*/PromSrcVecTy, /*Src=*/SrcVecTy,
3914         TargetTransformInfo::CastContextHint::None, CostKind);
3915     PromotionCost +=
3916         getCastInstrCost(Instruction::Trunc, /*Dst=*/DstVecTy,
3917                          /*Src=*/PromDstVecTy,
3918                          TargetTransformInfo::CastContextHint::None, CostKind);
3919     return PromotionCost + getReplicationShuffleCost(PromEltTy,
3920                                                      ReplicationFactor, VF,
3921                                                      DemandedDstElts, CostKind);
3922   }
3923 
3924   assert(LegalSrcVecTy.getScalarSizeInBits() == EltTyBits &&
3925          LegalSrcVecTy.getScalarType() == LegalDstVecTy.getScalarType() &&
3926          "We expect that the legalization doesn't affect the element width, "
3927          "doesn't coalesce/split elements.");
3928 
3929   unsigned NumEltsPerDstVec = LegalDstVecTy.getVectorNumElements();
3930   unsigned NumDstVectors =
3931       divideCeil(DstVecTy->getNumElements(), NumEltsPerDstVec);
3932 
3933   auto *SingleDstVecTy = FixedVectorType::get(EltTy, NumEltsPerDstVec);
3934 
3935   // Not all the produced Dst elements may be demanded. In our case,
3936   // given that a single Dst vector is formed by a single shuffle,
3937   // if all elements that will form a single Dst vector aren't demanded,
3938   // then we won't need to do that shuffle, so adjust the cost accordingly.
3939   APInt DemandedDstVectors = APIntOps::ScaleBitMask(
3940       DemandedDstElts.zextOrSelf(NumDstVectors * NumEltsPerDstVec),
3941       NumDstVectors);
3942   unsigned NumDstVectorsDemanded = DemandedDstVectors.countPopulation();
3943 
3944   InstructionCost SingleShuffleCost =
3945       getShuffleCost(TTI::SK_PermuteSingleSrc, SingleDstVecTy,
3946                      /*Mask=*/None, /*Index=*/0, /*SubTp=*/nullptr);
3947   return NumDstVectorsDemanded * SingleShuffleCost;
3948 }
3949 
3950 InstructionCost X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
3951                                             MaybeAlign Alignment,
3952                                             unsigned AddressSpace,
3953                                             TTI::TargetCostKind CostKind,
3954                                             const Instruction *I) {
3955   // TODO: Handle other cost kinds.
3956   if (CostKind != TTI::TCK_RecipThroughput) {
3957     if (auto *SI = dyn_cast_or_null<StoreInst>(I)) {
3958       // Store instruction with index and scale costs 2 Uops.
3959       // Check the preceding GEP to identify non-const indices.
3960       if (auto *GEP = dyn_cast<GetElementPtrInst>(SI->getPointerOperand())) {
3961         if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); }))
3962           return TTI::TCC_Basic * 2;
3963       }
3964     }
3965     return TTI::TCC_Basic;
3966   }
3967 
3968   assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
3969          "Invalid Opcode");
3970   // Type legalization can't handle structs
3971   if (TLI->getValueType(DL, Src, true) == MVT::Other)
3972     return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
3973                                   CostKind);
3974 
3975   // Legalize the type.
3976   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
3977 
3978   auto *VTy = dyn_cast<FixedVectorType>(Src);
3979 
3980   // Handle the simple case of non-vectors.
3981   // NOTE: this assumes that legalization never creates vector from scalars!
3982   if (!VTy || !LT.second.isVector())
3983     // Each load/store unit costs 1.
3984     return LT.first * 1;
3985 
3986   bool IsLoad = Opcode == Instruction::Load;
3987 
3988   Type *EltTy = VTy->getElementType();
3989 
3990   const int EltTyBits = DL.getTypeSizeInBits(EltTy);
3991 
3992   InstructionCost Cost = 0;
3993 
3994   // Source of truth: how many elements were there in the original IR vector?
3995   const unsigned SrcNumElt = VTy->getNumElements();
3996 
3997   // How far have we gotten?
3998   int NumEltRemaining = SrcNumElt;
3999   // Note that we intentionally capture by-reference, NumEltRemaining changes.
4000   auto NumEltDone = [&]() { return SrcNumElt - NumEltRemaining; };
4001 
4002   const int MaxLegalOpSizeBytes = divideCeil(LT.second.getSizeInBits(), 8);
4003 
4004   // Note that even if we can store 64 bits of an XMM, we still operate on XMM.
4005   const unsigned XMMBits = 128;
4006   if (XMMBits % EltTyBits != 0)
4007     // Vector size must be a multiple of the element size. I.e. no padding.
4008     return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
4009                                   CostKind);
4010   const int NumEltPerXMM = XMMBits / EltTyBits;
4011 
4012   auto *XMMVecTy = FixedVectorType::get(EltTy, NumEltPerXMM);
4013 
4014   for (int CurrOpSizeBytes = MaxLegalOpSizeBytes, SubVecEltsLeft = 0;
4015        NumEltRemaining > 0; CurrOpSizeBytes /= 2) {
4016     // How many elements would a single op deal with at once?
4017     if ((8 * CurrOpSizeBytes) % EltTyBits != 0)
4018       // Vector size must be a multiple of the element size. I.e. no padding.
4019       return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
4020                                     CostKind);
4021     int CurrNumEltPerOp = (8 * CurrOpSizeBytes) / EltTyBits;
4022 
4023     assert(CurrOpSizeBytes > 0 && CurrNumEltPerOp > 0 && "How'd we get here?");
4024     assert((((NumEltRemaining * EltTyBits) < (2 * 8 * CurrOpSizeBytes)) ||
4025             (CurrOpSizeBytes == MaxLegalOpSizeBytes)) &&
4026            "Unless we haven't halved the op size yet, "
4027            "we have less than two op's sized units of work left.");
4028 
4029     auto *CurrVecTy = CurrNumEltPerOp > NumEltPerXMM
4030                           ? FixedVectorType::get(EltTy, CurrNumEltPerOp)
4031                           : XMMVecTy;
4032 
4033     assert(CurrVecTy->getNumElements() % CurrNumEltPerOp == 0 &&
4034            "After halving sizes, the vector elt count is no longer a multiple "
4035            "of number of elements per operation?");
4036     auto *CoalescedVecTy =
4037         CurrNumEltPerOp == 1
4038             ? CurrVecTy
4039             : FixedVectorType::get(
4040                   IntegerType::get(Src->getContext(),
4041                                    EltTyBits * CurrNumEltPerOp),
4042                   CurrVecTy->getNumElements() / CurrNumEltPerOp);
4043     assert(DL.getTypeSizeInBits(CoalescedVecTy) ==
4044                DL.getTypeSizeInBits(CurrVecTy) &&
4045            "coalesciing elements doesn't change vector width.");
4046 
4047     while (NumEltRemaining > 0) {
4048       assert(SubVecEltsLeft >= 0 && "Subreg element count overconsumtion?");
4049 
4050       // Can we use this vector size, as per the remaining element count?
4051       // Iff the vector is naturally aligned, we can do a wide load regardless.
4052       if (NumEltRemaining < CurrNumEltPerOp &&
4053           (!IsLoad || Alignment.valueOrOne() < CurrOpSizeBytes) &&
4054           CurrOpSizeBytes != 1)
4055         break; // Try smalled vector size.
4056 
4057       bool Is0thSubVec = (NumEltDone() % LT.second.getVectorNumElements()) == 0;
4058 
4059       // If we have fully processed the previous reg, we need to replenish it.
4060       if (SubVecEltsLeft == 0) {
4061         SubVecEltsLeft += CurrVecTy->getNumElements();
4062         // And that's free only for the 0'th subvector of a legalized vector.
4063         if (!Is0thSubVec)
4064           Cost += getShuffleCost(IsLoad ? TTI::ShuffleKind::SK_InsertSubvector
4065                                         : TTI::ShuffleKind::SK_ExtractSubvector,
4066                                  VTy, None, NumEltDone(), CurrVecTy);
4067       }
4068 
4069       // While we can directly load/store ZMM, YMM, and 64-bit halves of XMM,
4070       // for smaller widths (32/16/8) we have to insert/extract them separately.
4071       // Again, it's free for the 0'th subreg (if op is 32/64 bit wide,
4072       // but let's pretend that it is also true for 16/8 bit wide ops...)
4073       if (CurrOpSizeBytes <= 32 / 8 && !Is0thSubVec) {
4074         int NumEltDoneInCurrXMM = NumEltDone() % NumEltPerXMM;
4075         assert(NumEltDoneInCurrXMM % CurrNumEltPerOp == 0 && "");
4076         int CoalescedVecEltIdx = NumEltDoneInCurrXMM / CurrNumEltPerOp;
4077         APInt DemandedElts =
4078             APInt::getBitsSet(CoalescedVecTy->getNumElements(),
4079                               CoalescedVecEltIdx, CoalescedVecEltIdx + 1);
4080         assert(DemandedElts.countPopulation() == 1 && "Inserting single value");
4081         Cost += getScalarizationOverhead(CoalescedVecTy, DemandedElts, IsLoad,
4082                                          !IsLoad);
4083       }
4084 
4085       // This isn't exactly right. We're using slow unaligned 32-byte accesses
4086       // as a proxy for a double-pumped AVX memory interface such as on
4087       // Sandybridge.
4088       if (CurrOpSizeBytes == 32 && ST->isUnalignedMem32Slow())
4089         Cost += 2;
4090       else
4091         Cost += 1;
4092 
4093       SubVecEltsLeft -= CurrNumEltPerOp;
4094       NumEltRemaining -= CurrNumEltPerOp;
4095       Alignment = commonAlignment(Alignment.valueOrOne(), CurrOpSizeBytes);
4096     }
4097   }
4098 
4099   assert(NumEltRemaining <= 0 && "Should have processed all the elements.");
4100 
4101   return Cost;
4102 }
4103 
4104 InstructionCost
4105 X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy, Align Alignment,
4106                                   unsigned AddressSpace,
4107                                   TTI::TargetCostKind CostKind) {
4108   bool IsLoad = (Instruction::Load == Opcode);
4109   bool IsStore = (Instruction::Store == Opcode);
4110 
4111   auto *SrcVTy = dyn_cast<FixedVectorType>(SrcTy);
4112   if (!SrcVTy)
4113     // To calculate scalar take the regular cost, without mask
4114     return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace, CostKind);
4115 
4116   unsigned NumElem = SrcVTy->getNumElements();
4117   auto *MaskTy =
4118       FixedVectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem);
4119   if ((IsLoad && !isLegalMaskedLoad(SrcVTy, Alignment)) ||
4120       (IsStore && !isLegalMaskedStore(SrcVTy, Alignment))) {
4121     // Scalarization
4122     APInt DemandedElts = APInt::getAllOnes(NumElem);
4123     InstructionCost MaskSplitCost =
4124         getScalarizationOverhead(MaskTy, DemandedElts, false, true);
4125     InstructionCost ScalarCompareCost = getCmpSelInstrCost(
4126         Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr,
4127         CmpInst::BAD_ICMP_PREDICATE, CostKind);
4128     InstructionCost BranchCost = getCFInstrCost(Instruction::Br, CostKind);
4129     InstructionCost MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost);
4130     InstructionCost ValueSplitCost =
4131         getScalarizationOverhead(SrcVTy, DemandedElts, IsLoad, IsStore);
4132     InstructionCost MemopCost =
4133         NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
4134                                          Alignment, AddressSpace, CostKind);
4135     return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost;
4136   }
4137 
4138   // Legalize the type.
4139   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy);
4140   auto VT = TLI->getValueType(DL, SrcVTy);
4141   InstructionCost Cost = 0;
4142   if (VT.isSimple() && LT.second != VT.getSimpleVT() &&
4143       LT.second.getVectorNumElements() == NumElem)
4144     // Promotion requires extend/truncate for data and a shuffle for mask.
4145     Cost += getShuffleCost(TTI::SK_PermuteTwoSrc, SrcVTy, None, 0, nullptr) +
4146             getShuffleCost(TTI::SK_PermuteTwoSrc, MaskTy, None, 0, nullptr);
4147 
4148   else if (LT.first * LT.second.getVectorNumElements() > NumElem) {
4149     auto *NewMaskTy = FixedVectorType::get(MaskTy->getElementType(),
4150                                            LT.second.getVectorNumElements());
4151     // Expanding requires fill mask with zeroes
4152     Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, None, 0, MaskTy);
4153   }
4154 
4155   // Pre-AVX512 - each maskmov load costs 2 + store costs ~8.
4156   if (!ST->hasAVX512())
4157     return Cost + LT.first * (IsLoad ? 2 : 8);
4158 
4159   // AVX-512 masked load/store is cheapper
4160   return Cost + LT.first;
4161 }
4162 
4163 InstructionCost X86TTIImpl::getAddressComputationCost(Type *Ty,
4164                                                       ScalarEvolution *SE,
4165                                                       const SCEV *Ptr) {
4166   // Address computations in vectorized code with non-consecutive addresses will
4167   // likely result in more instructions compared to scalar code where the
4168   // computation can more often be merged into the index mode. The resulting
4169   // extra micro-ops can significantly decrease throughput.
4170   const unsigned NumVectorInstToHideOverhead = 10;
4171 
4172   // Cost modeling of Strided Access Computation is hidden by the indexing
4173   // modes of X86 regardless of the stride value. We dont believe that there
4174   // is a difference between constant strided access in gerenal and constant
4175   // strided value which is less than or equal to 64.
4176   // Even in the case of (loop invariant) stride whose value is not known at
4177   // compile time, the address computation will not incur more than one extra
4178   // ADD instruction.
4179   if (Ty->isVectorTy() && SE && !ST->hasAVX2()) {
4180     // TODO: AVX2 is the current cut-off because we don't have correct
4181     //       interleaving costs for prior ISA's.
4182     if (!BaseT::isStridedAccess(Ptr))
4183       return NumVectorInstToHideOverhead;
4184     if (!BaseT::getConstantStrideStep(SE, Ptr))
4185       return 1;
4186   }
4187 
4188   return BaseT::getAddressComputationCost(Ty, SE, Ptr);
4189 }
4190 
4191 InstructionCost
4192 X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy,
4193                                        Optional<FastMathFlags> FMF,
4194                                        TTI::TargetCostKind CostKind) {
4195   if (TTI::requiresOrderedReduction(FMF))
4196     return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
4197 
4198   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
4199   // and make it as the cost.
4200 
4201   static const CostTblEntry SLMCostTblNoPairWise[] = {
4202     { ISD::FADD,  MVT::v2f64,   3 },
4203     { ISD::ADD,   MVT::v2i64,   5 },
4204   };
4205 
4206   static const CostTblEntry SSE2CostTblNoPairWise[] = {
4207     { ISD::FADD,  MVT::v2f64,   2 },
4208     { ISD::FADD,  MVT::v2f32,   2 },
4209     { ISD::FADD,  MVT::v4f32,   4 },
4210     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
4211     { ISD::ADD,   MVT::v2i32,   2 }, // FIXME: chosen to be less than v4i32
4212     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.3".
4213     { ISD::ADD,   MVT::v2i16,   2 },      // The data reported by the IACA tool is "4.3".
4214     { ISD::ADD,   MVT::v4i16,   3 },      // The data reported by the IACA tool is "4.3".
4215     { ISD::ADD,   MVT::v8i16,   4 },      // The data reported by the IACA tool is "4.3".
4216     { ISD::ADD,   MVT::v2i8,    2 },
4217     { ISD::ADD,   MVT::v4i8,    2 },
4218     { ISD::ADD,   MVT::v8i8,    2 },
4219     { ISD::ADD,   MVT::v16i8,   3 },
4220   };
4221 
4222   static const CostTblEntry AVX1CostTblNoPairWise[] = {
4223     { ISD::FADD,  MVT::v4f64,   3 },
4224     { ISD::FADD,  MVT::v4f32,   3 },
4225     { ISD::FADD,  MVT::v8f32,   4 },
4226     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
4227     { ISD::ADD,   MVT::v4i64,   3 },
4228     { ISD::ADD,   MVT::v8i32,   5 },
4229     { ISD::ADD,   MVT::v16i16,  5 },
4230     { ISD::ADD,   MVT::v32i8,   4 },
4231   };
4232 
4233   int ISD = TLI->InstructionOpcodeToISD(Opcode);
4234   assert(ISD && "Invalid opcode");
4235 
4236   // Before legalizing the type, give a chance to look up illegal narrow types
4237   // in the table.
4238   // FIXME: Is there a better way to do this?
4239   EVT VT = TLI->getValueType(DL, ValTy);
4240   if (VT.isSimple()) {
4241     MVT MTy = VT.getSimpleVT();
4242     if (ST->useSLMArithCosts())
4243       if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy))
4244         return Entry->Cost;
4245 
4246     if (ST->hasAVX())
4247       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
4248         return Entry->Cost;
4249 
4250     if (ST->hasSSE2())
4251       if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
4252         return Entry->Cost;
4253   }
4254 
4255   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
4256 
4257   MVT MTy = LT.second;
4258 
4259   auto *ValVTy = cast<FixedVectorType>(ValTy);
4260 
4261   // Special case: vXi8 mul reductions are performed as vXi16.
4262   if (ISD == ISD::MUL && MTy.getScalarType() == MVT::i8) {
4263     auto *WideSclTy = IntegerType::get(ValVTy->getContext(), 16);
4264     auto *WideVecTy = FixedVectorType::get(WideSclTy, ValVTy->getNumElements());
4265     return getCastInstrCost(Instruction::ZExt, WideVecTy, ValTy,
4266                             TargetTransformInfo::CastContextHint::None,
4267                             CostKind) +
4268            getArithmeticReductionCost(Opcode, WideVecTy, FMF, CostKind);
4269   }
4270 
4271   InstructionCost ArithmeticCost = 0;
4272   if (LT.first != 1 && MTy.isVector() &&
4273       MTy.getVectorNumElements() < ValVTy->getNumElements()) {
4274     // Type needs to be split. We need LT.first - 1 arithmetic ops.
4275     auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(),
4276                                             MTy.getVectorNumElements());
4277     ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind);
4278     ArithmeticCost *= LT.first - 1;
4279   }
4280 
4281   if (ST->useSLMArithCosts())
4282     if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy))
4283       return ArithmeticCost + Entry->Cost;
4284 
4285   if (ST->hasAVX())
4286     if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
4287       return ArithmeticCost + Entry->Cost;
4288 
4289   if (ST->hasSSE2())
4290     if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
4291       return ArithmeticCost + Entry->Cost;
4292 
4293   // FIXME: These assume a naive kshift+binop lowering, which is probably
4294   // conservative in most cases.
4295   static const CostTblEntry AVX512BoolReduction[] = {
4296     { ISD::AND,  MVT::v2i1,   3 },
4297     { ISD::AND,  MVT::v4i1,   5 },
4298     { ISD::AND,  MVT::v8i1,   7 },
4299     { ISD::AND,  MVT::v16i1,  9 },
4300     { ISD::AND,  MVT::v32i1, 11 },
4301     { ISD::AND,  MVT::v64i1, 13 },
4302     { ISD::OR,   MVT::v2i1,   3 },
4303     { ISD::OR,   MVT::v4i1,   5 },
4304     { ISD::OR,   MVT::v8i1,   7 },
4305     { ISD::OR,   MVT::v16i1,  9 },
4306     { ISD::OR,   MVT::v32i1, 11 },
4307     { ISD::OR,   MVT::v64i1, 13 },
4308   };
4309 
4310   static const CostTblEntry AVX2BoolReduction[] = {
4311     { ISD::AND,  MVT::v16i16,  2 }, // vpmovmskb + cmp
4312     { ISD::AND,  MVT::v32i8,   2 }, // vpmovmskb + cmp
4313     { ISD::OR,   MVT::v16i16,  2 }, // vpmovmskb + cmp
4314     { ISD::OR,   MVT::v32i8,   2 }, // vpmovmskb + cmp
4315   };
4316 
4317   static const CostTblEntry AVX1BoolReduction[] = {
4318     { ISD::AND,  MVT::v4i64,   2 }, // vmovmskpd + cmp
4319     { ISD::AND,  MVT::v8i32,   2 }, // vmovmskps + cmp
4320     { ISD::AND,  MVT::v16i16,  4 }, // vextractf128 + vpand + vpmovmskb + cmp
4321     { ISD::AND,  MVT::v32i8,   4 }, // vextractf128 + vpand + vpmovmskb + cmp
4322     { ISD::OR,   MVT::v4i64,   2 }, // vmovmskpd + cmp
4323     { ISD::OR,   MVT::v8i32,   2 }, // vmovmskps + cmp
4324     { ISD::OR,   MVT::v16i16,  4 }, // vextractf128 + vpor + vpmovmskb + cmp
4325     { ISD::OR,   MVT::v32i8,   4 }, // vextractf128 + vpor + vpmovmskb + cmp
4326   };
4327 
4328   static const CostTblEntry SSE2BoolReduction[] = {
4329     { ISD::AND,  MVT::v2i64,   2 }, // movmskpd + cmp
4330     { ISD::AND,  MVT::v4i32,   2 }, // movmskps + cmp
4331     { ISD::AND,  MVT::v8i16,   2 }, // pmovmskb + cmp
4332     { ISD::AND,  MVT::v16i8,   2 }, // pmovmskb + cmp
4333     { ISD::OR,   MVT::v2i64,   2 }, // movmskpd + cmp
4334     { ISD::OR,   MVT::v4i32,   2 }, // movmskps + cmp
4335     { ISD::OR,   MVT::v8i16,   2 }, // pmovmskb + cmp
4336     { ISD::OR,   MVT::v16i8,   2 }, // pmovmskb + cmp
4337   };
4338 
4339   // Handle bool allof/anyof patterns.
4340   if (ValVTy->getElementType()->isIntegerTy(1)) {
4341     InstructionCost ArithmeticCost = 0;
4342     if (LT.first != 1 && MTy.isVector() &&
4343         MTy.getVectorNumElements() < ValVTy->getNumElements()) {
4344       // Type needs to be split. We need LT.first - 1 arithmetic ops.
4345       auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(),
4346                                               MTy.getVectorNumElements());
4347       ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind);
4348       ArithmeticCost *= LT.first - 1;
4349     }
4350 
4351     if (ST->hasAVX512())
4352       if (const auto *Entry = CostTableLookup(AVX512BoolReduction, ISD, MTy))
4353         return ArithmeticCost + Entry->Cost;
4354     if (ST->hasAVX2())
4355       if (const auto *Entry = CostTableLookup(AVX2BoolReduction, ISD, MTy))
4356         return ArithmeticCost + Entry->Cost;
4357     if (ST->hasAVX())
4358       if (const auto *Entry = CostTableLookup(AVX1BoolReduction, ISD, MTy))
4359         return ArithmeticCost + Entry->Cost;
4360     if (ST->hasSSE2())
4361       if (const auto *Entry = CostTableLookup(SSE2BoolReduction, ISD, MTy))
4362         return ArithmeticCost + Entry->Cost;
4363 
4364     return BaseT::getArithmeticReductionCost(Opcode, ValVTy, FMF, CostKind);
4365   }
4366 
4367   unsigned NumVecElts = ValVTy->getNumElements();
4368   unsigned ScalarSize = ValVTy->getScalarSizeInBits();
4369 
4370   // Special case power of 2 reductions where the scalar type isn't changed
4371   // by type legalization.
4372   if (!isPowerOf2_32(NumVecElts) || ScalarSize != MTy.getScalarSizeInBits())
4373     return BaseT::getArithmeticReductionCost(Opcode, ValVTy, FMF, CostKind);
4374 
4375   InstructionCost ReductionCost = 0;
4376 
4377   auto *Ty = ValVTy;
4378   if (LT.first != 1 && MTy.isVector() &&
4379       MTy.getVectorNumElements() < ValVTy->getNumElements()) {
4380     // Type needs to be split. We need LT.first - 1 arithmetic ops.
4381     Ty = FixedVectorType::get(ValVTy->getElementType(),
4382                               MTy.getVectorNumElements());
4383     ReductionCost = getArithmeticInstrCost(Opcode, Ty, CostKind);
4384     ReductionCost *= LT.first - 1;
4385     NumVecElts = MTy.getVectorNumElements();
4386   }
4387 
4388   // Now handle reduction with the legal type, taking into account size changes
4389   // at each level.
4390   while (NumVecElts > 1) {
4391     // Determine the size of the remaining vector we need to reduce.
4392     unsigned Size = NumVecElts * ScalarSize;
4393     NumVecElts /= 2;
4394     // If we're reducing from 256/512 bits, use an extract_subvector.
4395     if (Size > 128) {
4396       auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts);
4397       ReductionCost +=
4398           getShuffleCost(TTI::SK_ExtractSubvector, Ty, None, NumVecElts, SubTy);
4399       Ty = SubTy;
4400     } else if (Size == 128) {
4401       // Reducing from 128 bits is a permute of v2f64/v2i64.
4402       FixedVectorType *ShufTy;
4403       if (ValVTy->isFloatingPointTy())
4404         ShufTy =
4405             FixedVectorType::get(Type::getDoubleTy(ValVTy->getContext()), 2);
4406       else
4407         ShufTy =
4408             FixedVectorType::get(Type::getInt64Ty(ValVTy->getContext()), 2);
4409       ReductionCost +=
4410           getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4411     } else if (Size == 64) {
4412       // Reducing from 64 bits is a shuffle of v4f32/v4i32.
4413       FixedVectorType *ShufTy;
4414       if (ValVTy->isFloatingPointTy())
4415         ShufTy =
4416             FixedVectorType::get(Type::getFloatTy(ValVTy->getContext()), 4);
4417       else
4418         ShufTy =
4419             FixedVectorType::get(Type::getInt32Ty(ValVTy->getContext()), 4);
4420       ReductionCost +=
4421           getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4422     } else {
4423       // Reducing from smaller size is a shift by immediate.
4424       auto *ShiftTy = FixedVectorType::get(
4425           Type::getIntNTy(ValVTy->getContext(), Size), 128 / Size);
4426       ReductionCost += getArithmeticInstrCost(
4427           Instruction::LShr, ShiftTy, CostKind,
4428           TargetTransformInfo::OK_AnyValue,
4429           TargetTransformInfo::OK_UniformConstantValue,
4430           TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
4431     }
4432 
4433     // Add the arithmetic op for this level.
4434     ReductionCost += getArithmeticInstrCost(Opcode, Ty, CostKind);
4435   }
4436 
4437   // Add the final extract element to the cost.
4438   return ReductionCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
4439 }
4440 
4441 InstructionCost X86TTIImpl::getMinMaxCost(Type *Ty, Type *CondTy,
4442                                           bool IsUnsigned) {
4443   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
4444 
4445   MVT MTy = LT.second;
4446 
4447   int ISD;
4448   if (Ty->isIntOrIntVectorTy()) {
4449     ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN;
4450   } else {
4451     assert(Ty->isFPOrFPVectorTy() &&
4452            "Expected float point or integer vector type.");
4453     ISD = ISD::FMINNUM;
4454   }
4455 
4456   static const CostTblEntry SSE1CostTbl[] = {
4457     {ISD::FMINNUM, MVT::v4f32, 1},
4458   };
4459 
4460   static const CostTblEntry SSE2CostTbl[] = {
4461     {ISD::FMINNUM, MVT::v2f64, 1},
4462     {ISD::SMIN,    MVT::v8i16, 1},
4463     {ISD::UMIN,    MVT::v16i8, 1},
4464   };
4465 
4466   static const CostTblEntry SSE41CostTbl[] = {
4467     {ISD::SMIN,    MVT::v4i32, 1},
4468     {ISD::UMIN,    MVT::v4i32, 1},
4469     {ISD::UMIN,    MVT::v8i16, 1},
4470     {ISD::SMIN,    MVT::v16i8, 1},
4471   };
4472 
4473   static const CostTblEntry SSE42CostTbl[] = {
4474     {ISD::UMIN,    MVT::v2i64, 3}, // xor+pcmpgtq+blendvpd
4475   };
4476 
4477   static const CostTblEntry AVX1CostTbl[] = {
4478     {ISD::FMINNUM, MVT::v8f32,  1},
4479     {ISD::FMINNUM, MVT::v4f64,  1},
4480     {ISD::SMIN,    MVT::v8i32,  3},
4481     {ISD::UMIN,    MVT::v8i32,  3},
4482     {ISD::SMIN,    MVT::v16i16, 3},
4483     {ISD::UMIN,    MVT::v16i16, 3},
4484     {ISD::SMIN,    MVT::v32i8,  3},
4485     {ISD::UMIN,    MVT::v32i8,  3},
4486   };
4487 
4488   static const CostTblEntry AVX2CostTbl[] = {
4489     {ISD::SMIN,    MVT::v8i32,  1},
4490     {ISD::UMIN,    MVT::v8i32,  1},
4491     {ISD::SMIN,    MVT::v16i16, 1},
4492     {ISD::UMIN,    MVT::v16i16, 1},
4493     {ISD::SMIN,    MVT::v32i8,  1},
4494     {ISD::UMIN,    MVT::v32i8,  1},
4495   };
4496 
4497   static const CostTblEntry AVX512CostTbl[] = {
4498     {ISD::FMINNUM, MVT::v16f32, 1},
4499     {ISD::FMINNUM, MVT::v8f64,  1},
4500     {ISD::SMIN,    MVT::v2i64,  1},
4501     {ISD::UMIN,    MVT::v2i64,  1},
4502     {ISD::SMIN,    MVT::v4i64,  1},
4503     {ISD::UMIN,    MVT::v4i64,  1},
4504     {ISD::SMIN,    MVT::v8i64,  1},
4505     {ISD::UMIN,    MVT::v8i64,  1},
4506     {ISD::SMIN,    MVT::v16i32, 1},
4507     {ISD::UMIN,    MVT::v16i32, 1},
4508   };
4509 
4510   static const CostTblEntry AVX512BWCostTbl[] = {
4511     {ISD::SMIN,    MVT::v32i16, 1},
4512     {ISD::UMIN,    MVT::v32i16, 1},
4513     {ISD::SMIN,    MVT::v64i8,  1},
4514     {ISD::UMIN,    MVT::v64i8,  1},
4515   };
4516 
4517   // If we have a native MIN/MAX instruction for this type, use it.
4518   if (ST->hasBWI())
4519     if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
4520       return LT.first * Entry->Cost;
4521 
4522   if (ST->hasAVX512())
4523     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
4524       return LT.first * Entry->Cost;
4525 
4526   if (ST->hasAVX2())
4527     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
4528       return LT.first * Entry->Cost;
4529 
4530   if (ST->hasAVX())
4531     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
4532       return LT.first * Entry->Cost;
4533 
4534   if (ST->hasSSE42())
4535     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
4536       return LT.first * Entry->Cost;
4537 
4538   if (ST->hasSSE41())
4539     if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
4540       return LT.first * Entry->Cost;
4541 
4542   if (ST->hasSSE2())
4543     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
4544       return LT.first * Entry->Cost;
4545 
4546   if (ST->hasSSE1())
4547     if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
4548       return LT.first * Entry->Cost;
4549 
4550   unsigned CmpOpcode;
4551   if (Ty->isFPOrFPVectorTy()) {
4552     CmpOpcode = Instruction::FCmp;
4553   } else {
4554     assert(Ty->isIntOrIntVectorTy() &&
4555            "expecting floating point or integer type for min/max reduction");
4556     CmpOpcode = Instruction::ICmp;
4557   }
4558 
4559   TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
4560   // Otherwise fall back to cmp+select.
4561   InstructionCost Result =
4562       getCmpSelInstrCost(CmpOpcode, Ty, CondTy, CmpInst::BAD_ICMP_PREDICATE,
4563                          CostKind) +
4564       getCmpSelInstrCost(Instruction::Select, Ty, CondTy,
4565                          CmpInst::BAD_ICMP_PREDICATE, CostKind);
4566   return Result;
4567 }
4568 
4569 InstructionCost
4570 X86TTIImpl::getMinMaxReductionCost(VectorType *ValTy, VectorType *CondTy,
4571                                    bool IsUnsigned,
4572                                    TTI::TargetCostKind CostKind) {
4573   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
4574 
4575   MVT MTy = LT.second;
4576 
4577   int ISD;
4578   if (ValTy->isIntOrIntVectorTy()) {
4579     ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN;
4580   } else {
4581     assert(ValTy->isFPOrFPVectorTy() &&
4582            "Expected float point or integer vector type.");
4583     ISD = ISD::FMINNUM;
4584   }
4585 
4586   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
4587   // and make it as the cost.
4588 
4589   static const CostTblEntry SSE2CostTblNoPairWise[] = {
4590       {ISD::UMIN, MVT::v2i16, 5}, // need pxors to use pminsw/pmaxsw
4591       {ISD::UMIN, MVT::v4i16, 7}, // need pxors to use pminsw/pmaxsw
4592       {ISD::UMIN, MVT::v8i16, 9}, // need pxors to use pminsw/pmaxsw
4593   };
4594 
4595   static const CostTblEntry SSE41CostTblNoPairWise[] = {
4596       {ISD::SMIN, MVT::v2i16, 3}, // same as sse2
4597       {ISD::SMIN, MVT::v4i16, 5}, // same as sse2
4598       {ISD::UMIN, MVT::v2i16, 5}, // same as sse2
4599       {ISD::UMIN, MVT::v4i16, 7}, // same as sse2
4600       {ISD::SMIN, MVT::v8i16, 4}, // phminposuw+xor
4601       {ISD::UMIN, MVT::v8i16, 4}, // FIXME: umin is cheaper than umax
4602       {ISD::SMIN, MVT::v2i8,  3}, // pminsb
4603       {ISD::SMIN, MVT::v4i8,  5}, // pminsb
4604       {ISD::SMIN, MVT::v8i8,  7}, // pminsb
4605       {ISD::SMIN, MVT::v16i8, 6},
4606       {ISD::UMIN, MVT::v2i8,  3}, // same as sse2
4607       {ISD::UMIN, MVT::v4i8,  5}, // same as sse2
4608       {ISD::UMIN, MVT::v8i8,  7}, // same as sse2
4609       {ISD::UMIN, MVT::v16i8, 6}, // FIXME: umin is cheaper than umax
4610   };
4611 
4612   static const CostTblEntry AVX1CostTblNoPairWise[] = {
4613       {ISD::SMIN, MVT::v16i16, 6},
4614       {ISD::UMIN, MVT::v16i16, 6}, // FIXME: umin is cheaper than umax
4615       {ISD::SMIN, MVT::v32i8, 8},
4616       {ISD::UMIN, MVT::v32i8, 8},
4617   };
4618 
4619   static const CostTblEntry AVX512BWCostTblNoPairWise[] = {
4620       {ISD::SMIN, MVT::v32i16, 8},
4621       {ISD::UMIN, MVT::v32i16, 8}, // FIXME: umin is cheaper than umax
4622       {ISD::SMIN, MVT::v64i8, 10},
4623       {ISD::UMIN, MVT::v64i8, 10},
4624   };
4625 
4626   // Before legalizing the type, give a chance to look up illegal narrow types
4627   // in the table.
4628   // FIXME: Is there a better way to do this?
4629   EVT VT = TLI->getValueType(DL, ValTy);
4630   if (VT.isSimple()) {
4631     MVT MTy = VT.getSimpleVT();
4632     if (ST->hasBWI())
4633       if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy))
4634         return Entry->Cost;
4635 
4636     if (ST->hasAVX())
4637       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
4638         return Entry->Cost;
4639 
4640     if (ST->hasSSE41())
4641       if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy))
4642         return Entry->Cost;
4643 
4644     if (ST->hasSSE2())
4645       if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
4646         return Entry->Cost;
4647   }
4648 
4649   auto *ValVTy = cast<FixedVectorType>(ValTy);
4650   unsigned NumVecElts = ValVTy->getNumElements();
4651 
4652   auto *Ty = ValVTy;
4653   InstructionCost MinMaxCost = 0;
4654   if (LT.first != 1 && MTy.isVector() &&
4655       MTy.getVectorNumElements() < ValVTy->getNumElements()) {
4656     // Type needs to be split. We need LT.first - 1 operations ops.
4657     Ty = FixedVectorType::get(ValVTy->getElementType(),
4658                               MTy.getVectorNumElements());
4659     auto *SubCondTy = FixedVectorType::get(CondTy->getElementType(),
4660                                            MTy.getVectorNumElements());
4661     MinMaxCost = getMinMaxCost(Ty, SubCondTy, IsUnsigned);
4662     MinMaxCost *= LT.first - 1;
4663     NumVecElts = MTy.getVectorNumElements();
4664   }
4665 
4666   if (ST->hasBWI())
4667     if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy))
4668       return MinMaxCost + Entry->Cost;
4669 
4670   if (ST->hasAVX())
4671     if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
4672       return MinMaxCost + Entry->Cost;
4673 
4674   if (ST->hasSSE41())
4675     if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy))
4676       return MinMaxCost + Entry->Cost;
4677 
4678   if (ST->hasSSE2())
4679     if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
4680       return MinMaxCost + Entry->Cost;
4681 
4682   unsigned ScalarSize = ValTy->getScalarSizeInBits();
4683 
4684   // Special case power of 2 reductions where the scalar type isn't changed
4685   // by type legalization.
4686   if (!isPowerOf2_32(ValVTy->getNumElements()) ||
4687       ScalarSize != MTy.getScalarSizeInBits())
4688     return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsUnsigned, CostKind);
4689 
4690   // Now handle reduction with the legal type, taking into account size changes
4691   // at each level.
4692   while (NumVecElts > 1) {
4693     // Determine the size of the remaining vector we need to reduce.
4694     unsigned Size = NumVecElts * ScalarSize;
4695     NumVecElts /= 2;
4696     // If we're reducing from 256/512 bits, use an extract_subvector.
4697     if (Size > 128) {
4698       auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts);
4699       MinMaxCost +=
4700           getShuffleCost(TTI::SK_ExtractSubvector, Ty, None, NumVecElts, SubTy);
4701       Ty = SubTy;
4702     } else if (Size == 128) {
4703       // Reducing from 128 bits is a permute of v2f64/v2i64.
4704       VectorType *ShufTy;
4705       if (ValTy->isFloatingPointTy())
4706         ShufTy =
4707             FixedVectorType::get(Type::getDoubleTy(ValTy->getContext()), 2);
4708       else
4709         ShufTy = FixedVectorType::get(Type::getInt64Ty(ValTy->getContext()), 2);
4710       MinMaxCost +=
4711           getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4712     } else if (Size == 64) {
4713       // Reducing from 64 bits is a shuffle of v4f32/v4i32.
4714       FixedVectorType *ShufTy;
4715       if (ValTy->isFloatingPointTy())
4716         ShufTy = FixedVectorType::get(Type::getFloatTy(ValTy->getContext()), 4);
4717       else
4718         ShufTy = FixedVectorType::get(Type::getInt32Ty(ValTy->getContext()), 4);
4719       MinMaxCost +=
4720           getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4721     } else {
4722       // Reducing from smaller size is a shift by immediate.
4723       auto *ShiftTy = FixedVectorType::get(
4724           Type::getIntNTy(ValTy->getContext(), Size), 128 / Size);
4725       MinMaxCost += getArithmeticInstrCost(
4726           Instruction::LShr, ShiftTy, TTI::TCK_RecipThroughput,
4727           TargetTransformInfo::OK_AnyValue,
4728           TargetTransformInfo::OK_UniformConstantValue,
4729           TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
4730     }
4731 
4732     // Add the arithmetic op for this level.
4733     auto *SubCondTy =
4734         FixedVectorType::get(CondTy->getElementType(), Ty->getNumElements());
4735     MinMaxCost += getMinMaxCost(Ty, SubCondTy, IsUnsigned);
4736   }
4737 
4738   // Add the final extract element to the cost.
4739   return MinMaxCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
4740 }
4741 
4742 /// Calculate the cost of materializing a 64-bit value. This helper
4743 /// method might only calculate a fraction of a larger immediate. Therefore it
4744 /// is valid to return a cost of ZERO.
4745 InstructionCost X86TTIImpl::getIntImmCost(int64_t Val) {
4746   if (Val == 0)
4747     return TTI::TCC_Free;
4748 
4749   if (isInt<32>(Val))
4750     return TTI::TCC_Basic;
4751 
4752   return 2 * TTI::TCC_Basic;
4753 }
4754 
4755 InstructionCost X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
4756                                           TTI::TargetCostKind CostKind) {
4757   assert(Ty->isIntegerTy());
4758 
4759   unsigned BitSize = Ty->getPrimitiveSizeInBits();
4760   if (BitSize == 0)
4761     return ~0U;
4762 
4763   // Never hoist constants larger than 128bit, because this might lead to
4764   // incorrect code generation or assertions in codegen.
4765   // Fixme: Create a cost model for types larger than i128 once the codegen
4766   // issues have been fixed.
4767   if (BitSize > 128)
4768     return TTI::TCC_Free;
4769 
4770   if (Imm == 0)
4771     return TTI::TCC_Free;
4772 
4773   // Sign-extend all constants to a multiple of 64-bit.
4774   APInt ImmVal = Imm;
4775   if (BitSize % 64 != 0)
4776     ImmVal = Imm.sext(alignTo(BitSize, 64));
4777 
4778   // Split the constant into 64-bit chunks and calculate the cost for each
4779   // chunk.
4780   InstructionCost Cost = 0;
4781   for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
4782     APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
4783     int64_t Val = Tmp.getSExtValue();
4784     Cost += getIntImmCost(Val);
4785   }
4786   // We need at least one instruction to materialize the constant.
4787   return std::max<InstructionCost>(1, Cost);
4788 }
4789 
4790 InstructionCost X86TTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
4791                                               const APInt &Imm, Type *Ty,
4792                                               TTI::TargetCostKind CostKind,
4793                                               Instruction *Inst) {
4794   assert(Ty->isIntegerTy());
4795 
4796   unsigned BitSize = Ty->getPrimitiveSizeInBits();
4797   // There is no cost model for constants with a bit size of 0. Return TCC_Free
4798   // here, so that constant hoisting will ignore this constant.
4799   if (BitSize == 0)
4800     return TTI::TCC_Free;
4801 
4802   unsigned ImmIdx = ~0U;
4803   switch (Opcode) {
4804   default:
4805     return TTI::TCC_Free;
4806   case Instruction::GetElementPtr:
4807     // Always hoist the base address of a GetElementPtr. This prevents the
4808     // creation of new constants for every base constant that gets constant
4809     // folded with the offset.
4810     if (Idx == 0)
4811       return 2 * TTI::TCC_Basic;
4812     return TTI::TCC_Free;
4813   case Instruction::Store:
4814     ImmIdx = 0;
4815     break;
4816   case Instruction::ICmp:
4817     // This is an imperfect hack to prevent constant hoisting of
4818     // compares that might be trying to check if a 64-bit value fits in
4819     // 32-bits. The backend can optimize these cases using a right shift by 32.
4820     // Ideally we would check the compare predicate here. There also other
4821     // similar immediates the backend can use shifts for.
4822     if (Idx == 1 && Imm.getBitWidth() == 64) {
4823       uint64_t ImmVal = Imm.getZExtValue();
4824       if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff)
4825         return TTI::TCC_Free;
4826     }
4827     ImmIdx = 1;
4828     break;
4829   case Instruction::And:
4830     // We support 64-bit ANDs with immediates with 32-bits of leading zeroes
4831     // by using a 32-bit operation with implicit zero extension. Detect such
4832     // immediates here as the normal path expects bit 31 to be sign extended.
4833     if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue()))
4834       return TTI::TCC_Free;
4835     ImmIdx = 1;
4836     break;
4837   case Instruction::Add:
4838   case Instruction::Sub:
4839     // For add/sub, we can use the opposite instruction for INT32_MIN.
4840     if (Idx == 1 && Imm.getBitWidth() == 64 && Imm.getZExtValue() == 0x80000000)
4841       return TTI::TCC_Free;
4842     ImmIdx = 1;
4843     break;
4844   case Instruction::UDiv:
4845   case Instruction::SDiv:
4846   case Instruction::URem:
4847   case Instruction::SRem:
4848     // Division by constant is typically expanded later into a different
4849     // instruction sequence. This completely changes the constants.
4850     // Report them as "free" to stop ConstantHoist from marking them as opaque.
4851     return TTI::TCC_Free;
4852   case Instruction::Mul:
4853   case Instruction::Or:
4854   case Instruction::Xor:
4855     ImmIdx = 1;
4856     break;
4857   // Always return TCC_Free for the shift value of a shift instruction.
4858   case Instruction::Shl:
4859   case Instruction::LShr:
4860   case Instruction::AShr:
4861     if (Idx == 1)
4862       return TTI::TCC_Free;
4863     break;
4864   case Instruction::Trunc:
4865   case Instruction::ZExt:
4866   case Instruction::SExt:
4867   case Instruction::IntToPtr:
4868   case Instruction::PtrToInt:
4869   case Instruction::BitCast:
4870   case Instruction::PHI:
4871   case Instruction::Call:
4872   case Instruction::Select:
4873   case Instruction::Ret:
4874   case Instruction::Load:
4875     break;
4876   }
4877 
4878   if (Idx == ImmIdx) {
4879     int NumConstants = divideCeil(BitSize, 64);
4880     InstructionCost Cost = X86TTIImpl::getIntImmCost(Imm, Ty, CostKind);
4881     return (Cost <= NumConstants * TTI::TCC_Basic)
4882                ? static_cast<int>(TTI::TCC_Free)
4883                : Cost;
4884   }
4885 
4886   return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind);
4887 }
4888 
4889 InstructionCost X86TTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
4890                                                 const APInt &Imm, Type *Ty,
4891                                                 TTI::TargetCostKind CostKind) {
4892   assert(Ty->isIntegerTy());
4893 
4894   unsigned BitSize = Ty->getPrimitiveSizeInBits();
4895   // There is no cost model for constants with a bit size of 0. Return TCC_Free
4896   // here, so that constant hoisting will ignore this constant.
4897   if (BitSize == 0)
4898     return TTI::TCC_Free;
4899 
4900   switch (IID) {
4901   default:
4902     return TTI::TCC_Free;
4903   case Intrinsic::sadd_with_overflow:
4904   case Intrinsic::uadd_with_overflow:
4905   case Intrinsic::ssub_with_overflow:
4906   case Intrinsic::usub_with_overflow:
4907   case Intrinsic::smul_with_overflow:
4908   case Intrinsic::umul_with_overflow:
4909     if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
4910       return TTI::TCC_Free;
4911     break;
4912   case Intrinsic::experimental_stackmap:
4913     if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
4914       return TTI::TCC_Free;
4915     break;
4916   case Intrinsic::experimental_patchpoint_void:
4917   case Intrinsic::experimental_patchpoint_i64:
4918     if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
4919       return TTI::TCC_Free;
4920     break;
4921   }
4922   return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind);
4923 }
4924 
4925 InstructionCost X86TTIImpl::getCFInstrCost(unsigned Opcode,
4926                                            TTI::TargetCostKind CostKind,
4927                                            const Instruction *I) {
4928   if (CostKind != TTI::TCK_RecipThroughput)
4929     return Opcode == Instruction::PHI ? 0 : 1;
4930   // Branches are assumed to be predicted.
4931   return 0;
4932 }
4933 
4934 int X86TTIImpl::getGatherOverhead() const {
4935   // Some CPUs have more overhead for gather. The specified overhead is relative
4936   // to the Load operation. "2" is the number provided by Intel architects. This
4937   // parameter is used for cost estimation of Gather Op and comparison with
4938   // other alternatives.
4939   // TODO: Remove the explicit hasAVX512()?, That would mean we would only
4940   // enable gather with a -march.
4941   if (ST->hasAVX512() || (ST->hasAVX2() && ST->hasFastGather()))
4942     return 2;
4943 
4944   return 1024;
4945 }
4946 
4947 int X86TTIImpl::getScatterOverhead() const {
4948   if (ST->hasAVX512())
4949     return 2;
4950 
4951   return 1024;
4952 }
4953 
4954 // Return an average cost of Gather / Scatter instruction, maybe improved later.
4955 // FIXME: Add TargetCostKind support.
4956 InstructionCost X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy,
4957                                             const Value *Ptr, Align Alignment,
4958                                             unsigned AddressSpace) {
4959 
4960   assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost");
4961   unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements();
4962 
4963   // Try to reduce index size from 64 bit (default for GEP)
4964   // to 32. It is essential for VF 16. If the index can't be reduced to 32, the
4965   // operation will use 16 x 64 indices which do not fit in a zmm and needs
4966   // to split. Also check that the base pointer is the same for all lanes,
4967   // and that there's at most one variable index.
4968   auto getIndexSizeInBits = [](const Value *Ptr, const DataLayout &DL) {
4969     unsigned IndexSize = DL.getPointerSizeInBits();
4970     const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
4971     if (IndexSize < 64 || !GEP)
4972       return IndexSize;
4973 
4974     unsigned NumOfVarIndices = 0;
4975     const Value *Ptrs = GEP->getPointerOperand();
4976     if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs))
4977       return IndexSize;
4978     for (unsigned i = 1; i < GEP->getNumOperands(); ++i) {
4979       if (isa<Constant>(GEP->getOperand(i)))
4980         continue;
4981       Type *IndxTy = GEP->getOperand(i)->getType();
4982       if (auto *IndexVTy = dyn_cast<VectorType>(IndxTy))
4983         IndxTy = IndexVTy->getElementType();
4984       if ((IndxTy->getPrimitiveSizeInBits() == 64 &&
4985           !isa<SExtInst>(GEP->getOperand(i))) ||
4986          ++NumOfVarIndices > 1)
4987         return IndexSize; // 64
4988     }
4989     return (unsigned)32;
4990   };
4991 
4992   // Trying to reduce IndexSize to 32 bits for vector 16.
4993   // By default the IndexSize is equal to pointer size.
4994   unsigned IndexSize = (ST->hasAVX512() && VF >= 16)
4995                            ? getIndexSizeInBits(Ptr, DL)
4996                            : DL.getPointerSizeInBits();
4997 
4998   auto *IndexVTy = FixedVectorType::get(
4999       IntegerType::get(SrcVTy->getContext(), IndexSize), VF);
5000   std::pair<InstructionCost, MVT> IdxsLT =
5001       TLI->getTypeLegalizationCost(DL, IndexVTy);
5002   std::pair<InstructionCost, MVT> SrcLT =
5003       TLI->getTypeLegalizationCost(DL, SrcVTy);
5004   InstructionCost::CostType SplitFactor =
5005       *std::max(IdxsLT.first, SrcLT.first).getValue();
5006   if (SplitFactor > 1) {
5007     // Handle splitting of vector of pointers
5008     auto *SplitSrcTy =
5009         FixedVectorType::get(SrcVTy->getScalarType(), VF / SplitFactor);
5010     return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment,
5011                                          AddressSpace);
5012   }
5013 
5014   // The gather / scatter cost is given by Intel architects. It is a rough
5015   // number since we are looking at one instruction in a time.
5016   const int GSOverhead = (Opcode == Instruction::Load)
5017                              ? getGatherOverhead()
5018                              : getScatterOverhead();
5019   return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
5020                                            MaybeAlign(Alignment), AddressSpace,
5021                                            TTI::TCK_RecipThroughput);
5022 }
5023 
5024 /// Return the cost of full scalarization of gather / scatter operation.
5025 ///
5026 /// Opcode - Load or Store instruction.
5027 /// SrcVTy - The type of the data vector that should be gathered or scattered.
5028 /// VariableMask - The mask is non-constant at compile time.
5029 /// Alignment - Alignment for one element.
5030 /// AddressSpace - pointer[s] address space.
5031 ///
5032 /// FIXME: Add TargetCostKind support.
5033 InstructionCost X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy,
5034                                             bool VariableMask, Align Alignment,
5035                                             unsigned AddressSpace) {
5036   Type *ScalarTy = SrcVTy->getScalarType();
5037   unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements();
5038   APInt DemandedElts = APInt::getAllOnes(VF);
5039   TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
5040 
5041   InstructionCost MaskUnpackCost = 0;
5042   if (VariableMask) {
5043     auto *MaskTy =
5044         FixedVectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF);
5045     MaskUnpackCost = getScalarizationOverhead(
5046         MaskTy, DemandedElts, /*Insert=*/false, /*Extract=*/true);
5047     InstructionCost ScalarCompareCost = getCmpSelInstrCost(
5048         Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()), nullptr,
5049         CmpInst::BAD_ICMP_PREDICATE, CostKind);
5050     InstructionCost BranchCost = getCFInstrCost(Instruction::Br, CostKind);
5051     MaskUnpackCost += VF * (BranchCost + ScalarCompareCost);
5052   }
5053 
5054   InstructionCost AddressUnpackCost = getScalarizationOverhead(
5055       FixedVectorType::get(ScalarTy->getPointerTo(), VF), DemandedElts,
5056       /*Insert=*/false, /*Extract=*/true);
5057 
5058   // The cost of the scalar loads/stores.
5059   InstructionCost MemoryOpCost =
5060       VF * getMemoryOpCost(Opcode, ScalarTy, MaybeAlign(Alignment),
5061                            AddressSpace, CostKind);
5062 
5063   // The cost of forming the vector from loaded scalars/
5064   // scalarizing the vector to perform scalar stores.
5065   InstructionCost InsertExtractCost =
5066       getScalarizationOverhead(cast<FixedVectorType>(SrcVTy), DemandedElts,
5067                                /*Insert=*/Opcode == Instruction::Load,
5068                                /*Extract=*/Opcode == Instruction::Store);
5069 
5070   return AddressUnpackCost + MemoryOpCost + MaskUnpackCost + InsertExtractCost;
5071 }
5072 
5073 /// Calculate the cost of Gather / Scatter operation
5074 InstructionCost X86TTIImpl::getGatherScatterOpCost(
5075     unsigned Opcode, Type *SrcVTy, const Value *Ptr, bool VariableMask,
5076     Align Alignment, TTI::TargetCostKind CostKind,
5077     const Instruction *I = nullptr) {
5078   if (CostKind != TTI::TCK_RecipThroughput) {
5079     if ((Opcode == Instruction::Load &&
5080          isLegalMaskedGather(SrcVTy, Align(Alignment)) &&
5081          !forceScalarizeMaskedGather(cast<VectorType>(SrcVTy),
5082                                      Align(Alignment))) ||
5083         (Opcode == Instruction::Store &&
5084          isLegalMaskedScatter(SrcVTy, Align(Alignment)) &&
5085          !forceScalarizeMaskedScatter(cast<VectorType>(SrcVTy),
5086                                       Align(Alignment))))
5087       return 1;
5088     return BaseT::getGatherScatterOpCost(Opcode, SrcVTy, Ptr, VariableMask,
5089                                          Alignment, CostKind, I);
5090   }
5091 
5092   assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter");
5093   PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
5094   if (!PtrTy && Ptr->getType()->isVectorTy())
5095     PtrTy = dyn_cast<PointerType>(
5096         cast<VectorType>(Ptr->getType())->getElementType());
5097   assert(PtrTy && "Unexpected type for Ptr argument");
5098   unsigned AddressSpace = PtrTy->getAddressSpace();
5099 
5100   if ((Opcode == Instruction::Load &&
5101        (!isLegalMaskedGather(SrcVTy, Align(Alignment)) ||
5102         forceScalarizeMaskedGather(cast<VectorType>(SrcVTy),
5103                                    Align(Alignment)))) ||
5104       (Opcode == Instruction::Store &&
5105        (!isLegalMaskedScatter(SrcVTy, Align(Alignment)) ||
5106         forceScalarizeMaskedScatter(cast<VectorType>(SrcVTy),
5107                                     Align(Alignment)))))
5108     return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment,
5109                            AddressSpace);
5110 
5111   return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace);
5112 }
5113 
5114 bool X86TTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1,
5115                                TargetTransformInfo::LSRCost &C2) {
5116     // X86 specific here are "instruction number 1st priority".
5117     return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost,
5118                     C1.NumIVMuls, C1.NumBaseAdds,
5119                     C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
5120            std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost,
5121                     C2.NumIVMuls, C2.NumBaseAdds,
5122                     C2.ScaleCost, C2.ImmCost, C2.SetupCost);
5123 }
5124 
5125 bool X86TTIImpl::canMacroFuseCmp() {
5126   return ST->hasMacroFusion() || ST->hasBranchFusion();
5127 }
5128 
5129 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy, Align Alignment) {
5130   if (!ST->hasAVX())
5131     return false;
5132 
5133   // The backend can't handle a single element vector.
5134   if (isa<VectorType>(DataTy) &&
5135       cast<FixedVectorType>(DataTy)->getNumElements() == 1)
5136     return false;
5137   Type *ScalarTy = DataTy->getScalarType();
5138 
5139   if (ScalarTy->isPointerTy())
5140     return true;
5141 
5142   if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
5143     return true;
5144 
5145   if (ScalarTy->isHalfTy() && ST->hasBWI() && ST->hasFP16())
5146     return true;
5147 
5148   if (!ScalarTy->isIntegerTy())
5149     return false;
5150 
5151   unsigned IntWidth = ScalarTy->getIntegerBitWidth();
5152   return IntWidth == 32 || IntWidth == 64 ||
5153          ((IntWidth == 8 || IntWidth == 16) && ST->hasBWI());
5154 }
5155 
5156 bool X86TTIImpl::isLegalMaskedStore(Type *DataType, Align Alignment) {
5157   return isLegalMaskedLoad(DataType, Alignment);
5158 }
5159 
5160 bool X86TTIImpl::isLegalNTLoad(Type *DataType, Align Alignment) {
5161   unsigned DataSize = DL.getTypeStoreSize(DataType);
5162   // The only supported nontemporal loads are for aligned vectors of 16 or 32
5163   // bytes.  Note that 32-byte nontemporal vector loads are supported by AVX2
5164   // (the equivalent stores only require AVX).
5165   if (Alignment >= DataSize && (DataSize == 16 || DataSize == 32))
5166     return DataSize == 16 ?  ST->hasSSE1() : ST->hasAVX2();
5167 
5168   return false;
5169 }
5170 
5171 bool X86TTIImpl::isLegalNTStore(Type *DataType, Align Alignment) {
5172   unsigned DataSize = DL.getTypeStoreSize(DataType);
5173 
5174   // SSE4A supports nontemporal stores of float and double at arbitrary
5175   // alignment.
5176   if (ST->hasSSE4A() && (DataType->isFloatTy() || DataType->isDoubleTy()))
5177     return true;
5178 
5179   // Besides the SSE4A subtarget exception above, only aligned stores are
5180   // available nontemporaly on any other subtarget.  And only stores with a size
5181   // of 4..32 bytes (powers of 2, only) are permitted.
5182   if (Alignment < DataSize || DataSize < 4 || DataSize > 32 ||
5183       !isPowerOf2_32(DataSize))
5184     return false;
5185 
5186   // 32-byte vector nontemporal stores are supported by AVX (the equivalent
5187   // loads require AVX2).
5188   if (DataSize == 32)
5189     return ST->hasAVX();
5190   if (DataSize == 16)
5191     return ST->hasSSE1();
5192   return true;
5193 }
5194 
5195 bool X86TTIImpl::isLegalBroadcastLoad(Type *ElementTy,
5196                                       ElementCount NumElements) const {
5197   // movddup
5198   return ST->hasSSE3() && !NumElements.isScalable() &&
5199          NumElements.getFixedValue() == 2 &&
5200          ElementTy == Type::getDoubleTy(ElementTy->getContext());
5201 }
5202 
5203 bool X86TTIImpl::isLegalMaskedExpandLoad(Type *DataTy) {
5204   if (!isa<VectorType>(DataTy))
5205     return false;
5206 
5207   if (!ST->hasAVX512())
5208     return false;
5209 
5210   // The backend can't handle a single element vector.
5211   if (cast<FixedVectorType>(DataTy)->getNumElements() == 1)
5212     return false;
5213 
5214   Type *ScalarTy = cast<VectorType>(DataTy)->getElementType();
5215 
5216   if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
5217     return true;
5218 
5219   if (!ScalarTy->isIntegerTy())
5220     return false;
5221 
5222   unsigned IntWidth = ScalarTy->getIntegerBitWidth();
5223   return IntWidth == 32 || IntWidth == 64 ||
5224          ((IntWidth == 8 || IntWidth == 16) && ST->hasVBMI2());
5225 }
5226 
5227 bool X86TTIImpl::isLegalMaskedCompressStore(Type *DataTy) {
5228   return isLegalMaskedExpandLoad(DataTy);
5229 }
5230 
5231 bool X86TTIImpl::supportsGather() const {
5232   // Some CPUs have better gather performance than others.
5233   // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only
5234   // enable gather with a -march.
5235   return ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2());
5236 }
5237 
5238 bool X86TTIImpl::forceScalarizeMaskedGather(VectorType *VTy, Align Alignment) {
5239   // Gather / Scatter for vector 2 is not profitable on KNL / SKX
5240   // Vector-4 of gather/scatter instruction does not exist on KNL. We can extend
5241   // it to 8 elements, but zeroing upper bits of the mask vector will add more
5242   // instructions. Right now we give the scalar cost of vector-4 for KNL. TODO:
5243   // Check, maybe the gather/scatter instruction is better in the VariableMask
5244   // case.
5245   unsigned NumElts = cast<FixedVectorType>(VTy)->getNumElements();
5246   return NumElts == 1 ||
5247          (ST->hasAVX512() && (NumElts == 2 || (NumElts == 4 && !ST->hasVLX())));
5248 }
5249 
5250 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy, Align Alignment) {
5251   if (!supportsGather())
5252     return false;
5253   Type *ScalarTy = DataTy->getScalarType();
5254   if (ScalarTy->isPointerTy())
5255     return true;
5256 
5257   if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
5258     return true;
5259 
5260   if (!ScalarTy->isIntegerTy())
5261     return false;
5262 
5263   unsigned IntWidth = ScalarTy->getIntegerBitWidth();
5264   return IntWidth == 32 || IntWidth == 64;
5265 }
5266 
5267 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType, Align Alignment) {
5268   // AVX2 doesn't support scatter
5269   if (!ST->hasAVX512())
5270     return false;
5271   return isLegalMaskedGather(DataType, Alignment);
5272 }
5273 
5274 bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) {
5275   EVT VT = TLI->getValueType(DL, DataType);
5276   return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT);
5277 }
5278 
5279 bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
5280   return false;
5281 }
5282 
5283 bool X86TTIImpl::areInlineCompatible(const Function *Caller,
5284                                      const Function *Callee) const {
5285   const TargetMachine &TM = getTLI()->getTargetMachine();
5286 
5287   // Work this as a subsetting of subtarget features.
5288   const FeatureBitset &CallerBits =
5289       TM.getSubtargetImpl(*Caller)->getFeatureBits();
5290   const FeatureBitset &CalleeBits =
5291       TM.getSubtargetImpl(*Callee)->getFeatureBits();
5292 
5293   // Check whether features are the same (apart from the ignore list).
5294   FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList;
5295   FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList;
5296   if (RealCallerBits == RealCalleeBits)
5297     return true;
5298 
5299   // If the features are a subset, we need to additionally check for calls
5300   // that may become ABI-incompatible as a result of inlining.
5301   if ((RealCallerBits & RealCalleeBits) != RealCalleeBits)
5302     return false;
5303 
5304   for (const Instruction &I : instructions(Callee)) {
5305     if (const auto *CB = dyn_cast<CallBase>(&I)) {
5306       SmallVector<Type *, 8> Types;
5307       for (Value *Arg : CB->args())
5308         Types.push_back(Arg->getType());
5309       if (!CB->getType()->isVoidTy())
5310         Types.push_back(CB->getType());
5311 
5312       // Simple types are always ABI compatible.
5313       auto IsSimpleTy = [](Type *Ty) {
5314         return !Ty->isVectorTy() && !Ty->isAggregateType();
5315       };
5316       if (all_of(Types, IsSimpleTy))
5317         continue;
5318 
5319       if (Function *NestedCallee = CB->getCalledFunction()) {
5320         // Assume that intrinsics are always ABI compatible.
5321         if (NestedCallee->isIntrinsic())
5322           continue;
5323 
5324         // Do a precise compatibility check.
5325         if (!areTypesABICompatible(Caller, NestedCallee, Types))
5326           return false;
5327       } else {
5328         // We don't know the target features of the callee,
5329         // assume it is incompatible.
5330         return false;
5331       }
5332     }
5333   }
5334   return true;
5335 }
5336 
5337 bool X86TTIImpl::areTypesABICompatible(const Function *Caller,
5338                                        const Function *Callee,
5339                                        const ArrayRef<Type *> &Types) const {
5340   if (!BaseT::areTypesABICompatible(Caller, Callee, Types))
5341     return false;
5342 
5343   // If we get here, we know the target features match. If one function
5344   // considers 512-bit vectors legal and the other does not, consider them
5345   // incompatible.
5346   const TargetMachine &TM = getTLI()->getTargetMachine();
5347 
5348   if (TM.getSubtarget<X86Subtarget>(*Caller).useAVX512Regs() ==
5349       TM.getSubtarget<X86Subtarget>(*Callee).useAVX512Regs())
5350     return true;
5351 
5352   // Consider the arguments compatible if they aren't vectors or aggregates.
5353   // FIXME: Look at the size of vectors.
5354   // FIXME: Look at the element types of aggregates to see if there are vectors.
5355   return llvm::none_of(Types,
5356       [](Type *T) { return T->isVectorTy() || T->isAggregateType(); });
5357 }
5358 
5359 X86TTIImpl::TTI::MemCmpExpansionOptions
5360 X86TTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
5361   TTI::MemCmpExpansionOptions Options;
5362   Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize);
5363   Options.NumLoadsPerBlock = 2;
5364   // All GPR and vector loads can be unaligned.
5365   Options.AllowOverlappingLoads = true;
5366   if (IsZeroCmp) {
5367     // Only enable vector loads for equality comparison. Right now the vector
5368     // version is not as fast for three way compare (see #33329).
5369     const unsigned PreferredWidth = ST->getPreferVectorWidth();
5370     if (PreferredWidth >= 512 && ST->hasAVX512()) Options.LoadSizes.push_back(64);
5371     if (PreferredWidth >= 256 && ST->hasAVX()) Options.LoadSizes.push_back(32);
5372     if (PreferredWidth >= 128 && ST->hasSSE2()) Options.LoadSizes.push_back(16);
5373   }
5374   if (ST->is64Bit()) {
5375     Options.LoadSizes.push_back(8);
5376   }
5377   Options.LoadSizes.push_back(4);
5378   Options.LoadSizes.push_back(2);
5379   Options.LoadSizes.push_back(1);
5380   return Options;
5381 }
5382 
5383 bool X86TTIImpl::prefersVectorizedAddressing() const {
5384   return supportsGather();
5385 }
5386 
5387 bool X86TTIImpl::supportsEfficientVectorElementLoadStore() const {
5388   return false;
5389 }
5390 
5391 bool X86TTIImpl::enableInterleavedAccessVectorization() {
5392   // TODO: We expect this to be beneficial regardless of arch,
5393   // but there are currently some unexplained performance artifacts on Atom.
5394   // As a temporary solution, disable on Atom.
5395   return !(ST->isAtom());
5396 }
5397 
5398 // Get estimation for interleaved load/store operations and strided load.
5399 // \p Indices contains indices for strided load.
5400 // \p Factor - the factor of interleaving.
5401 // AVX-512 provides 3-src shuffles that significantly reduces the cost.
5402 InstructionCost X86TTIImpl::getInterleavedMemoryOpCostAVX512(
5403     unsigned Opcode, FixedVectorType *VecTy, unsigned Factor,
5404     ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace,
5405     TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) {
5406   // VecTy for interleave memop is <VF*Factor x Elt>.
5407   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
5408   // VecTy = <12 x i32>.
5409 
5410   // Calculate the number of memory operations (NumOfMemOps), required
5411   // for load/store the VecTy.
5412   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
5413   unsigned VecTySize = DL.getTypeStoreSize(VecTy);
5414   unsigned LegalVTSize = LegalVT.getStoreSize();
5415   unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize;
5416 
5417   // Get the cost of one memory operation.
5418   auto *SingleMemOpTy = FixedVectorType::get(VecTy->getElementType(),
5419                                              LegalVT.getVectorNumElements());
5420   InstructionCost MemOpCost;
5421   bool UseMaskedMemOp = UseMaskForCond || UseMaskForGaps;
5422   if (UseMaskedMemOp)
5423     MemOpCost = getMaskedMemoryOpCost(Opcode, SingleMemOpTy, Alignment,
5424                                       AddressSpace, CostKind);
5425   else
5426     MemOpCost = getMemoryOpCost(Opcode, SingleMemOpTy, MaybeAlign(Alignment),
5427                                 AddressSpace, CostKind);
5428 
5429   unsigned VF = VecTy->getNumElements() / Factor;
5430   MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF);
5431 
5432   InstructionCost MaskCost;
5433   if (UseMaskedMemOp) {
5434     APInt DemandedLoadStoreElts = APInt::getZero(VecTy->getNumElements());
5435     for (unsigned Index : Indices) {
5436       assert(Index < Factor && "Invalid index for interleaved memory op");
5437       for (unsigned Elm = 0; Elm < VF; Elm++)
5438         DemandedLoadStoreElts.setBit(Index + Elm * Factor);
5439     }
5440 
5441     Type *I1Type = Type::getInt1Ty(VecTy->getContext());
5442 
5443     MaskCost = getReplicationShuffleCost(
5444         I1Type, Factor, VF,
5445         UseMaskForGaps ? DemandedLoadStoreElts
5446                        : APInt::getAllOnes(VecTy->getNumElements()),
5447         CostKind);
5448 
5449     // The Gaps mask is invariant and created outside the loop, therefore the
5450     // cost of creating it is not accounted for here. However if we have both
5451     // a MaskForGaps and some other mask that guards the execution of the
5452     // memory access, we need to account for the cost of And-ing the two masks
5453     // inside the loop.
5454     if (UseMaskForGaps) {
5455       auto *MaskVT = FixedVectorType::get(I1Type, VecTy->getNumElements());
5456       MaskCost += getArithmeticInstrCost(BinaryOperator::And, MaskVT, CostKind);
5457     }
5458   }
5459 
5460   if (Opcode == Instruction::Load) {
5461     // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl)
5462     // contain the cost of the optimized shuffle sequence that the
5463     // X86InterleavedAccess pass will generate.
5464     // The cost of loads and stores are computed separately from the table.
5465 
5466     // X86InterleavedAccess support only the following interleaved-access group.
5467     static const CostTblEntry AVX512InterleavedLoadTbl[] = {
5468         {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8
5469         {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8
5470         {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8
5471     };
5472 
5473     if (const auto *Entry =
5474             CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT))
5475       return MaskCost + NumOfMemOps * MemOpCost + Entry->Cost;
5476     //If an entry does not exist, fallback to the default implementation.
5477 
5478     // Kind of shuffle depends on number of loaded values.
5479     // If we load the entire data in one register, we can use a 1-src shuffle.
5480     // Otherwise, we'll merge 2 sources in each operation.
5481     TTI::ShuffleKind ShuffleKind =
5482         (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc;
5483 
5484     InstructionCost ShuffleCost =
5485         getShuffleCost(ShuffleKind, SingleMemOpTy, None, 0, nullptr);
5486 
5487     unsigned NumOfLoadsInInterleaveGrp =
5488         Indices.size() ? Indices.size() : Factor;
5489     auto *ResultTy = FixedVectorType::get(VecTy->getElementType(),
5490                                           VecTy->getNumElements() / Factor);
5491     InstructionCost NumOfResults =
5492         getTLI()->getTypeLegalizationCost(DL, ResultTy).first *
5493         NumOfLoadsInInterleaveGrp;
5494 
5495     // About a half of the loads may be folded in shuffles when we have only
5496     // one result. If we have more than one result, or the loads are masked,
5497     // we do not fold loads at all.
5498     unsigned NumOfUnfoldedLoads =
5499         UseMaskedMemOp || NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2;
5500 
5501     // Get a number of shuffle operations per result.
5502     unsigned NumOfShufflesPerResult =
5503         std::max((unsigned)1, (unsigned)(NumOfMemOps - 1));
5504 
5505     // The SK_MergeTwoSrc shuffle clobbers one of src operands.
5506     // When we have more than one destination, we need additional instructions
5507     // to keep sources.
5508     InstructionCost NumOfMoves = 0;
5509     if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc)
5510       NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2;
5511 
5512     InstructionCost Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost +
5513                            MaskCost + NumOfUnfoldedLoads * MemOpCost +
5514                            NumOfMoves;
5515 
5516     return Cost;
5517   }
5518 
5519   // Store.
5520   assert(Opcode == Instruction::Store &&
5521          "Expected Store Instruction at this  point");
5522   // X86InterleavedAccess support only the following interleaved-access group.
5523   static const CostTblEntry AVX512InterleavedStoreTbl[] = {
5524       {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store)
5525       {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store)
5526       {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store)
5527 
5528       {4, MVT::v8i8, 10},  // interleave 4 x 8i8  into 32i8  (and store)
5529       {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8  (and store)
5530       {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store)
5531       {4, MVT::v64i8, 24}  // interleave 4 x 32i8 into 256i8 (and store)
5532   };
5533 
5534   if (const auto *Entry =
5535           CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT))
5536     return MaskCost + NumOfMemOps * MemOpCost + Entry->Cost;
5537   //If an entry does not exist, fallback to the default implementation.
5538 
5539   // There is no strided stores meanwhile. And store can't be folded in
5540   // shuffle.
5541   unsigned NumOfSources = Factor; // The number of values to be merged.
5542   InstructionCost ShuffleCost =
5543       getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, None, 0, nullptr);
5544   unsigned NumOfShufflesPerStore = NumOfSources - 1;
5545 
5546   // The SK_MergeTwoSrc shuffle clobbers one of src operands.
5547   // We need additional instructions to keep sources.
5548   unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2;
5549   InstructionCost Cost =
5550       MaskCost +
5551       NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) +
5552       NumOfMoves;
5553   return Cost;
5554 }
5555 
5556 InstructionCost X86TTIImpl::getInterleavedMemoryOpCost(
5557     unsigned Opcode, Type *BaseTy, unsigned Factor, ArrayRef<unsigned> Indices,
5558     Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
5559     bool UseMaskForCond, bool UseMaskForGaps) {
5560   auto *VecTy = cast<FixedVectorType>(BaseTy);
5561 
5562   auto isSupportedOnAVX512 = [&](Type *VecTy, bool HasBW) {
5563     Type *EltTy = cast<VectorType>(VecTy)->getElementType();
5564     if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) ||
5565         EltTy->isIntegerTy(32) || EltTy->isPointerTy())
5566       return true;
5567     if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8) ||
5568         (!ST->useSoftFloat() && ST->hasFP16() && EltTy->isHalfTy()))
5569       return HasBW;
5570     return false;
5571   };
5572   if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI()))
5573     return getInterleavedMemoryOpCostAVX512(
5574         Opcode, VecTy, Factor, Indices, Alignment,
5575         AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps);
5576 
5577   if (UseMaskForCond || UseMaskForGaps)
5578     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5579                                              Alignment, AddressSpace, CostKind,
5580                                              UseMaskForCond, UseMaskForGaps);
5581 
5582   // Get estimation for interleaved load/store operations for SSE-AVX2.
5583   // As opposed to AVX-512, SSE-AVX2 do not have generic shuffles that allow
5584   // computing the cost using a generic formula as a function of generic
5585   // shuffles. We therefore use a lookup table instead, filled according to
5586   // the instruction sequences that codegen currently generates.
5587 
5588   // VecTy for interleave memop is <VF*Factor x Elt>.
5589   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
5590   // VecTy = <12 x i32>.
5591   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
5592 
5593   // This function can be called with VecTy=<6xi128>, Factor=3, in which case
5594   // the VF=2, while v2i128 is an unsupported MVT vector type
5595   // (see MachineValueType.h::getVectorVT()).
5596   if (!LegalVT.isVector())
5597     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5598                                              Alignment, AddressSpace, CostKind);
5599 
5600   unsigned VF = VecTy->getNumElements() / Factor;
5601   Type *ScalarTy = VecTy->getElementType();
5602   // Deduplicate entries, model floats/pointers as appropriately-sized integers.
5603   if (!ScalarTy->isIntegerTy())
5604     ScalarTy =
5605         Type::getIntNTy(ScalarTy->getContext(), DL.getTypeSizeInBits(ScalarTy));
5606 
5607   // Get the cost of all the memory operations.
5608   // FIXME: discount dead loads.
5609   InstructionCost MemOpCosts = getMemoryOpCost(
5610       Opcode, VecTy, MaybeAlign(Alignment), AddressSpace, CostKind);
5611 
5612   auto *VT = FixedVectorType::get(ScalarTy, VF);
5613   EVT ETy = TLI->getValueType(DL, VT);
5614   if (!ETy.isSimple())
5615     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5616                                              Alignment, AddressSpace, CostKind);
5617 
5618   // TODO: Complete for other data-types and strides.
5619   // Each combination of Stride, element bit width and VF results in a different
5620   // sequence; The cost tables are therefore accessed with:
5621   // Factor (stride) and VectorType=VFxiN.
5622   // The Cost accounts only for the shuffle sequence;
5623   // The cost of the loads/stores is accounted for separately.
5624   //
5625   static const CostTblEntry AVX2InterleavedLoadTbl[] = {
5626       {2, MVT::v2i8, 2},  // (load 4i8 and) deinterleave into 2 x 2i8
5627       {2, MVT::v4i8, 2},  // (load 8i8 and) deinterleave into 2 x 4i8
5628       {2, MVT::v8i8, 2},  // (load 16i8 and) deinterleave into 2 x 8i8
5629       {2, MVT::v16i8, 4}, // (load 32i8 and) deinterleave into 2 x 16i8
5630       {2, MVT::v32i8, 6}, // (load 64i8 and) deinterleave into 2 x 32i8
5631 
5632       {2, MVT::v8i16, 6},   // (load 16i16 and) deinterleave into 2 x 8i16
5633       {2, MVT::v16i16, 9},  // (load 32i16 and) deinterleave into 2 x 16i16
5634       {2, MVT::v32i16, 18}, // (load 64i16 and) deinterleave into 2 x 32i16
5635 
5636       {2, MVT::v8i32, 4},   // (load 16i32 and) deinterleave into 2 x 8i32
5637       {2, MVT::v16i32, 8},  // (load 32i32 and) deinterleave into 2 x 16i32
5638       {2, MVT::v32i32, 16}, // (load 64i32 and) deinterleave into 2 x 32i32
5639 
5640       {2, MVT::v4i64, 4},   // (load 8i64 and) deinterleave into 2 x 4i64
5641       {2, MVT::v8i64, 8},   // (load 16i64 and) deinterleave into 2 x 8i64
5642       {2, MVT::v16i64, 16}, // (load 32i64 and) deinterleave into 2 x 16i64
5643       {2, MVT::v32i64, 32}, // (load 64i64 and) deinterleave into 2 x 32i64
5644 
5645       {3, MVT::v2i8, 3},   // (load 6i8 and) deinterleave into 3 x 2i8
5646       {3, MVT::v4i8, 3},   // (load 12i8 and) deinterleave into 3 x 4i8
5647       {3, MVT::v8i8, 6},   // (load 24i8 and) deinterleave into 3 x 8i8
5648       {3, MVT::v16i8, 11}, // (load 48i8 and) deinterleave into 3 x 16i8
5649       {3, MVT::v32i8, 14}, // (load 96i8 and) deinterleave into 3 x 32i8
5650 
5651       {3, MVT::v2i16, 5},   // (load 6i16 and) deinterleave into 3 x 2i16
5652       {3, MVT::v4i16, 7},   // (load 12i16 and) deinterleave into 3 x 4i16
5653       {3, MVT::v8i16, 9},   // (load 24i16 and) deinterleave into 3 x 8i16
5654       {3, MVT::v16i16, 28}, // (load 48i16 and) deinterleave into 3 x 16i16
5655       {3, MVT::v32i16, 56}, // (load 96i16 and) deinterleave into 3 x 32i16
5656 
5657       {3, MVT::v2i32, 3},   // (load 6i32 and) deinterleave into 3 x 2i32
5658       {3, MVT::v4i32, 3},   // (load 12i32 and) deinterleave into 3 x 4i32
5659       {3, MVT::v8i32, 7},   // (load 24i32 and) deinterleave into 3 x 8i32
5660       {3, MVT::v16i32, 14}, // (load 48i32 and) deinterleave into 3 x 16i32
5661       {3, MVT::v32i32, 32}, // (load 96i32 and) deinterleave into 3 x 32i32
5662 
5663       {3, MVT::v2i64, 1},   // (load 6i64 and) deinterleave into 3 x 2i64
5664       {3, MVT::v4i64, 5},   // (load 12i64 and) deinterleave into 3 x 4i64
5665       {3, MVT::v8i64, 10},  // (load 24i64 and) deinterleave into 3 x 8i64
5666       {3, MVT::v16i64, 20}, // (load 48i64 and) deinterleave into 3 x 16i64
5667 
5668       {4, MVT::v2i8, 4},   // (load 8i8 and) deinterleave into 4 x 2i8
5669       {4, MVT::v4i8, 4},   // (load 16i8 and) deinterleave into 4 x 4i8
5670       {4, MVT::v8i8, 12},  // (load 32i8 and) deinterleave into 4 x 8i8
5671       {4, MVT::v16i8, 24}, // (load 64i8 and) deinterleave into 4 x 16i8
5672       {4, MVT::v32i8, 56}, // (load 128i8 and) deinterleave into 4 x 32i8
5673 
5674       {4, MVT::v2i16, 6},    // (load 8i16 and) deinterleave into 4 x 2i16
5675       {4, MVT::v4i16, 17},   // (load 16i16 and) deinterleave into 4 x 4i16
5676       {4, MVT::v8i16, 33},   // (load 32i16 and) deinterleave into 4 x 8i16
5677       {4, MVT::v16i16, 75},  // (load 64i16 and) deinterleave into 4 x 16i16
5678       {4, MVT::v32i16, 150}, // (load 128i16 and) deinterleave into 4 x 32i16
5679 
5680       {4, MVT::v2i32, 4},   // (load 8i32 and) deinterleave into 4 x 2i32
5681       {4, MVT::v4i32, 8},   // (load 16i32 and) deinterleave into 4 x 4i32
5682       {4, MVT::v8i32, 16},  // (load 32i32 and) deinterleave into 4 x 8i32
5683       {4, MVT::v16i32, 32}, // (load 64i32 and) deinterleave into 4 x 16i32
5684       {4, MVT::v32i32, 68}, // (load 128i32 and) deinterleave into 4 x 32i32
5685 
5686       {4, MVT::v2i64, 6},  // (load 8i64 and) deinterleave into 4 x 2i64
5687       {4, MVT::v4i64, 8},  // (load 16i64 and) deinterleave into 4 x 4i64
5688       {4, MVT::v8i64, 20}, // (load 32i64 and) deinterleave into 4 x 8i64
5689       {4, MVT::v16i64, 40}, // (load 64i64 and) deinterleave into 4 x 16i64
5690 
5691       {6, MVT::v2i8, 6},   // (load 12i8 and) deinterleave into 6 x 2i8
5692       {6, MVT::v4i8, 14},  // (load 24i8 and) deinterleave into 6 x 4i8
5693       {6, MVT::v8i8, 18},  // (load 48i8 and) deinterleave into 6 x 8i8
5694       {6, MVT::v16i8, 43}, // (load 96i8 and) deinterleave into 6 x 16i8
5695       {6, MVT::v32i8, 82}, // (load 192i8 and) deinterleave into 6 x 32i8
5696 
5697       {6, MVT::v2i16, 13},   // (load 12i16 and) deinterleave into 6 x 2i16
5698       {6, MVT::v4i16, 9},    // (load 24i16 and) deinterleave into 6 x 4i16
5699       {6, MVT::v8i16, 39},   // (load 48i16 and) deinterleave into 6 x 8i16
5700       {6, MVT::v16i16, 106}, // (load 96i16 and) deinterleave into 6 x 16i16
5701       {6, MVT::v32i16, 212}, // (load 192i16 and) deinterleave into 6 x 32i16
5702 
5703       {6, MVT::v2i32, 6},   // (load 12i32 and) deinterleave into 6 x 2i32
5704       {6, MVT::v4i32, 15},  // (load 24i32 and) deinterleave into 6 x 4i32
5705       {6, MVT::v8i32, 31},  // (load 48i32 and) deinterleave into 6 x 8i32
5706       {6, MVT::v16i32, 64}, // (load 96i32 and) deinterleave into 6 x 16i32
5707 
5708       {6, MVT::v2i64, 6},  // (load 12i64 and) deinterleave into 6 x 2i64
5709       {6, MVT::v4i64, 18}, // (load 24i64 and) deinterleave into 6 x 4i64
5710       {6, MVT::v8i64, 36}, // (load 48i64 and) deinterleave into 6 x 8i64
5711 
5712       {8, MVT::v8i32, 40} // (load 64i32 and) deinterleave into 8 x 8i32
5713   };
5714 
5715   static const CostTblEntry SSSE3InterleavedLoadTbl[] = {
5716       {2, MVT::v4i16, 2},   // (load 8i16 and) deinterleave into 2 x 4i16
5717   };
5718 
5719   static const CostTblEntry SSE2InterleavedLoadTbl[] = {
5720       {2, MVT::v2i16, 2},   // (load 4i16 and) deinterleave into 2 x 2i16
5721       {2, MVT::v4i16, 7},   // (load 8i16 and) deinterleave into 2 x 4i16
5722 
5723       {2, MVT::v2i32, 2},   // (load 4i32 and) deinterleave into 2 x 2i32
5724       {2, MVT::v4i32, 2},   // (load 8i32 and) deinterleave into 2 x 4i32
5725 
5726       {2, MVT::v2i64, 2},   // (load 4i64 and) deinterleave into 2 x 2i64
5727   };
5728 
5729   static const CostTblEntry AVX2InterleavedStoreTbl[] = {
5730       {2, MVT::v16i8, 3}, // interleave 2 x 16i8 into 32i8 (and store)
5731       {2, MVT::v32i8, 4}, // interleave 2 x 32i8 into 64i8 (and store)
5732 
5733       {2, MVT::v8i16, 3},  // interleave 2 x 8i16 into 16i16 (and store)
5734       {2, MVT::v16i16, 4}, // interleave 2 x 16i16 into 32i16 (and store)
5735       {2, MVT::v32i16, 8}, // interleave 2 x 32i16 into 64i16 (and store)
5736 
5737       {2, MVT::v4i32, 2},   // interleave 2 x 4i32 into 8i32 (and store)
5738       {2, MVT::v8i32, 4},   // interleave 2 x 8i32 into 16i32 (and store)
5739       {2, MVT::v16i32, 8},  // interleave 2 x 16i32 into 32i32 (and store)
5740       {2, MVT::v32i32, 16}, // interleave 2 x 32i32 into 64i32 (and store)
5741 
5742       {2, MVT::v2i64, 2},   // interleave 2 x 2i64 into 4i64 (and store)
5743       {2, MVT::v4i64, 4},   // interleave 2 x 4i64 into 8i64 (and store)
5744       {2, MVT::v8i64, 8},   // interleave 2 x 8i64 into 16i64 (and store)
5745       {2, MVT::v16i64, 16}, // interleave 2 x 16i64 into 32i64 (and store)
5746       {2, MVT::v32i64, 32}, // interleave 2 x 32i64 into 64i64 (and store)
5747 
5748       {3, MVT::v2i8, 4},   // interleave 3 x 2i8 into 6i8 (and store)
5749       {3, MVT::v4i8, 4},   // interleave 3 x 4i8 into 12i8 (and store)
5750       {3, MVT::v8i8, 6},   // interleave 3 x 8i8 into 24i8 (and store)
5751       {3, MVT::v16i8, 11}, // interleave 3 x 16i8 into 48i8 (and store)
5752       {3, MVT::v32i8, 13}, // interleave 3 x 32i8 into 96i8 (and store)
5753 
5754       {3, MVT::v2i16, 4},   // interleave 3 x 2i16 into 6i16 (and store)
5755       {3, MVT::v4i16, 6},   // interleave 3 x 4i16 into 12i16 (and store)
5756       {3, MVT::v8i16, 12},  // interleave 3 x 8i16 into 24i16 (and store)
5757       {3, MVT::v16i16, 27}, // interleave 3 x 16i16 into 48i16 (and store)
5758       {3, MVT::v32i16, 54}, // interleave 3 x 32i16 into 96i16 (and store)
5759 
5760       {3, MVT::v2i32, 4},   // interleave 3 x 2i32 into 6i32 (and store)
5761       {3, MVT::v4i32, 5},   // interleave 3 x 4i32 into 12i32 (and store)
5762       {3, MVT::v8i32, 11},  // interleave 3 x 8i32 into 24i32 (and store)
5763       {3, MVT::v16i32, 22}, // interleave 3 x 16i32 into 48i32 (and store)
5764       {3, MVT::v32i32, 48}, // interleave 3 x 32i32 into 96i32 (and store)
5765 
5766       {3, MVT::v2i64, 4},   // interleave 3 x 2i64 into 6i64 (and store)
5767       {3, MVT::v4i64, 6},   // interleave 3 x 4i64 into 12i64 (and store)
5768       {3, MVT::v8i64, 12},  // interleave 3 x 8i64 into 24i64 (and store)
5769       {3, MVT::v16i64, 24}, // interleave 3 x 16i64 into 48i64 (and store)
5770 
5771       {4, MVT::v2i8, 4},   // interleave 4 x 2i8 into 8i8 (and store)
5772       {4, MVT::v4i8, 4},   // interleave 4 x 4i8 into 16i8 (and store)
5773       {4, MVT::v8i8, 4},   // interleave 4 x 8i8 into 32i8 (and store)
5774       {4, MVT::v16i8, 8},  // interleave 4 x 16i8 into 64i8 (and store)
5775       {4, MVT::v32i8, 12}, // interleave 4 x 32i8 into 128i8 (and store)
5776 
5777       {4, MVT::v2i16, 2},   // interleave 4 x 2i16 into 8i16 (and store)
5778       {4, MVT::v4i16, 6},   // interleave 4 x 4i16 into 16i16 (and store)
5779       {4, MVT::v8i16, 10},  // interleave 4 x 8i16 into 32i16 (and store)
5780       {4, MVT::v16i16, 32}, // interleave 4 x 16i16 into 64i16 (and store)
5781       {4, MVT::v32i16, 64}, // interleave 4 x 32i16 into 128i16 (and store)
5782 
5783       {4, MVT::v2i32, 5},   // interleave 4 x 2i32 into 8i32 (and store)
5784       {4, MVT::v4i32, 6},   // interleave 4 x 4i32 into 16i32 (and store)
5785       {4, MVT::v8i32, 16},  // interleave 4 x 8i32 into 32i32 (and store)
5786       {4, MVT::v16i32, 32}, // interleave 4 x 16i32 into 64i32 (and store)
5787       {4, MVT::v32i32, 64}, // interleave 4 x 32i32 into 128i32 (and store)
5788 
5789       {4, MVT::v2i64, 6},  // interleave 4 x 2i64 into 8i64 (and store)
5790       {4, MVT::v4i64, 8},  // interleave 4 x 4i64 into 16i64 (and store)
5791       {4, MVT::v8i64, 20}, // interleave 4 x 8i64 into 32i64 (and store)
5792       {4, MVT::v16i64, 40}, // interleave 4 x 16i64 into 64i64 (and store)
5793 
5794       {6, MVT::v2i8, 7},   // interleave 6 x 2i8 into 12i8 (and store)
5795       {6, MVT::v4i8, 9},   // interleave 6 x 4i8 into 24i8 (and store)
5796       {6, MVT::v8i8, 16},  // interleave 6 x 8i8 into 48i8 (and store)
5797       {6, MVT::v16i8, 27}, // interleave 6 x 16i8 into 96i8 (and store)
5798       {6, MVT::v32i8, 90}, // interleave 6 x 32i8 into 192i8 (and store)
5799 
5800       {6, MVT::v2i16, 10},  // interleave 6 x 2i16 into 12i16 (and store)
5801       {6, MVT::v4i16, 15},  // interleave 6 x 4i16 into 24i16 (and store)
5802       {6, MVT::v8i16, 21},  // interleave 6 x 8i16 into 48i16 (and store)
5803       {6, MVT::v16i16, 58}, // interleave 6 x 16i16 into 96i16 (and store)
5804       {6, MVT::v32i16, 90}, // interleave 6 x 32i16 into 192i16 (and store)
5805 
5806       {6, MVT::v2i32, 9},   // interleave 6 x 2i32 into 12i32 (and store)
5807       {6, MVT::v4i32, 12},  // interleave 6 x 4i32 into 24i32 (and store)
5808       {6, MVT::v8i32, 33},  // interleave 6 x 8i32 into 48i32 (and store)
5809       {6, MVT::v16i32, 66}, // interleave 6 x 16i32 into 96i32 (and store)
5810 
5811       {6, MVT::v2i64, 8},  // interleave 6 x 2i64 into 12i64 (and store)
5812       {6, MVT::v4i64, 15}, // interleave 6 x 4i64 into 24i64 (and store)
5813       {6, MVT::v8i64, 30}, // interleave 6 x 8i64 into 48i64 (and store)
5814   };
5815 
5816   static const CostTblEntry SSE2InterleavedStoreTbl[] = {
5817       {2, MVT::v2i8, 1},   // interleave 2 x 2i8 into 4i8 (and store)
5818       {2, MVT::v4i8, 1},   // interleave 2 x 4i8 into 8i8 (and store)
5819       {2, MVT::v8i8, 1},   // interleave 2 x 8i8 into 16i8 (and store)
5820 
5821       {2, MVT::v2i16, 1},  // interleave 2 x 2i16 into 4i16 (and store)
5822       {2, MVT::v4i16, 1},  // interleave 2 x 4i16 into 8i16 (and store)
5823 
5824       {2, MVT::v2i32, 1},  // interleave 2 x 2i32 into 4i32 (and store)
5825   };
5826 
5827   if (Opcode == Instruction::Load) {
5828     auto GetDiscountedCost = [Factor, NumMembers = Indices.size(),
5829                               MemOpCosts](const CostTblEntry *Entry) {
5830       // NOTE: this is just an approximation!
5831       //       It can over/under -estimate the cost!
5832       return MemOpCosts + divideCeil(NumMembers * Entry->Cost, Factor);
5833     };
5834 
5835     if (ST->hasAVX2())
5836       if (const auto *Entry = CostTableLookup(AVX2InterleavedLoadTbl, Factor,
5837                                               ETy.getSimpleVT()))
5838         return GetDiscountedCost(Entry);
5839 
5840     if (ST->hasSSSE3())
5841       if (const auto *Entry = CostTableLookup(SSSE3InterleavedLoadTbl, Factor,
5842                                               ETy.getSimpleVT()))
5843         return GetDiscountedCost(Entry);
5844 
5845     if (ST->hasSSE2())
5846       if (const auto *Entry = CostTableLookup(SSE2InterleavedLoadTbl, Factor,
5847                                               ETy.getSimpleVT()))
5848         return GetDiscountedCost(Entry);
5849   } else {
5850     assert(Opcode == Instruction::Store &&
5851            "Expected Store Instruction at this point");
5852     assert((!Indices.size() || Indices.size() == Factor) &&
5853            "Interleaved store only supports fully-interleaved groups.");
5854     if (ST->hasAVX2())
5855       if (const auto *Entry = CostTableLookup(AVX2InterleavedStoreTbl, Factor,
5856                                               ETy.getSimpleVT()))
5857         return MemOpCosts + Entry->Cost;
5858 
5859     if (ST->hasSSE2())
5860       if (const auto *Entry = CostTableLookup(SSE2InterleavedStoreTbl, Factor,
5861                                               ETy.getSimpleVT()))
5862         return MemOpCosts + Entry->Cost;
5863   }
5864 
5865   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5866                                            Alignment, AddressSpace, CostKind,
5867                                            UseMaskForCond, UseMaskForGaps);
5868 }
5869