xref: /freebsd-src/contrib/llvm-project/llvm/lib/Target/X86/X86TargetTransformInfo.cpp (revision 753f127f3ace09432b2baeffd71a308760641a62)
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 we're truncating to the same legalized type - just assume its free.
2522   if (ISD == ISD::TRUNCATE && LTSrc.second == LTDest.second)
2523     return TTI::TCC_Free;
2524 
2525   if (ST->useAVX512Regs()) {
2526     if (ST->hasBWI())
2527       if (const auto *Entry = ConvertCostTableLookup(
2528               AVX512BWConversionTbl, ISD, LTDest.second, LTSrc.second))
2529         return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2530 
2531     if (ST->hasDQI())
2532       if (const auto *Entry = ConvertCostTableLookup(
2533               AVX512DQConversionTbl, ISD, LTDest.second, LTSrc.second))
2534         return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2535 
2536     if (ST->hasAVX512())
2537       if (const auto *Entry = ConvertCostTableLookup(
2538               AVX512FConversionTbl, ISD, LTDest.second, LTSrc.second))
2539         return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2540   }
2541 
2542   if (ST->hasBWI())
2543     if (const auto *Entry = ConvertCostTableLookup(AVX512BWVLConversionTbl, ISD,
2544                                                    LTDest.second, LTSrc.second))
2545       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2546 
2547   if (ST->hasDQI())
2548     if (const auto *Entry = ConvertCostTableLookup(AVX512DQVLConversionTbl, ISD,
2549                                                    LTDest.second, LTSrc.second))
2550       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2551 
2552   if (ST->hasAVX512())
2553     if (const auto *Entry = ConvertCostTableLookup(AVX512VLConversionTbl, ISD,
2554                                                    LTDest.second, LTSrc.second))
2555       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2556 
2557   if (ST->hasAVX2())
2558     if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
2559                                                    LTDest.second, LTSrc.second))
2560       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2561 
2562   if (ST->hasAVX())
2563     if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
2564                                                    LTDest.second, LTSrc.second))
2565       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2566 
2567   if (ST->hasSSE41())
2568     if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
2569                                                    LTDest.second, LTSrc.second))
2570       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2571 
2572   if (ST->hasSSE2())
2573     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
2574                                                    LTDest.second, LTSrc.second))
2575       return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2576 
2577   // Fallback, for i8/i16 sitofp/uitofp cases we need to extend to i32 for
2578   // sitofp.
2579   if ((ISD == ISD::SINT_TO_FP || ISD == ISD::UINT_TO_FP) &&
2580       1 < Src->getScalarSizeInBits() && Src->getScalarSizeInBits() < 32) {
2581     Type *ExtSrc = Src->getWithNewBitWidth(32);
2582     unsigned ExtOpc =
2583         (ISD == ISD::SINT_TO_FP) ? Instruction::SExt : Instruction::ZExt;
2584 
2585     // For scalar loads the extend would be free.
2586     InstructionCost ExtCost = 0;
2587     if (!(Src->isIntegerTy() && I && isa<LoadInst>(I->getOperand(0))))
2588       ExtCost = getCastInstrCost(ExtOpc, ExtSrc, Src, CCH, CostKind);
2589 
2590     return ExtCost + getCastInstrCost(Instruction::SIToFP, Dst, ExtSrc,
2591                                       TTI::CastContextHint::None, CostKind);
2592   }
2593 
2594   // Fallback for fptosi/fptoui i8/i16 cases we need to truncate from fptosi
2595   // i32.
2596   if ((ISD == ISD::FP_TO_SINT || ISD == ISD::FP_TO_UINT) &&
2597       1 < Dst->getScalarSizeInBits() && Dst->getScalarSizeInBits() < 32) {
2598     Type *TruncDst = Dst->getWithNewBitWidth(32);
2599     return getCastInstrCost(Instruction::FPToSI, TruncDst, Src, CCH, CostKind) +
2600            getCastInstrCost(Instruction::Trunc, Dst, TruncDst,
2601                             TTI::CastContextHint::None, CostKind);
2602   }
2603 
2604   return AdjustCost(
2605       BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I));
2606 }
2607 
2608 InstructionCost X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
2609                                                Type *CondTy,
2610                                                CmpInst::Predicate VecPred,
2611                                                TTI::TargetCostKind CostKind,
2612                                                const Instruction *I) {
2613   // TODO: Handle other cost kinds.
2614   if (CostKind != TTI::TCK_RecipThroughput)
2615     return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
2616                                      I);
2617 
2618   // Legalize the type.
2619   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
2620 
2621   MVT MTy = LT.second;
2622 
2623   int ISD = TLI->InstructionOpcodeToISD(Opcode);
2624   assert(ISD && "Invalid opcode");
2625 
2626   InstructionCost ExtraCost = 0;
2627   if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) {
2628     // Some vector comparison predicates cost extra instructions.
2629     // TODO: Should we invert this and assume worst case cmp costs
2630     // and reduce for particular predicates?
2631     if (MTy.isVector() &&
2632         !((ST->hasXOP() && (!ST->hasAVX2() || MTy.is128BitVector())) ||
2633           (ST->hasAVX512() && 32 <= MTy.getScalarSizeInBits()) ||
2634           ST->hasBWI())) {
2635       // Fallback to I if a specific predicate wasn't specified.
2636       CmpInst::Predicate Pred = VecPred;
2637       if (I && (Pred == CmpInst::BAD_ICMP_PREDICATE ||
2638                 Pred == CmpInst::BAD_FCMP_PREDICATE))
2639         Pred = cast<CmpInst>(I)->getPredicate();
2640 
2641       switch (Pred) {
2642       case CmpInst::Predicate::ICMP_NE:
2643         // xor(cmpeq(x,y),-1)
2644         ExtraCost = 1;
2645         break;
2646       case CmpInst::Predicate::ICMP_SGE:
2647       case CmpInst::Predicate::ICMP_SLE:
2648         // xor(cmpgt(x,y),-1)
2649         ExtraCost = 1;
2650         break;
2651       case CmpInst::Predicate::ICMP_ULT:
2652       case CmpInst::Predicate::ICMP_UGT:
2653         // cmpgt(xor(x,signbit),xor(y,signbit))
2654         // xor(cmpeq(pmaxu(x,y),x),-1)
2655         ExtraCost = 2;
2656         break;
2657       case CmpInst::Predicate::ICMP_ULE:
2658       case CmpInst::Predicate::ICMP_UGE:
2659         if ((ST->hasSSE41() && MTy.getScalarSizeInBits() == 32) ||
2660             (ST->hasSSE2() && MTy.getScalarSizeInBits() < 32)) {
2661           // cmpeq(psubus(x,y),0)
2662           // cmpeq(pminu(x,y),x)
2663           ExtraCost = 1;
2664         } else {
2665           // xor(cmpgt(xor(x,signbit),xor(y,signbit)),-1)
2666           ExtraCost = 3;
2667         }
2668         break;
2669       case CmpInst::Predicate::BAD_ICMP_PREDICATE:
2670       case CmpInst::Predicate::BAD_FCMP_PREDICATE:
2671         // Assume worst case scenario and add the maximum extra cost.
2672         ExtraCost = 3;
2673         break;
2674       default:
2675         break;
2676       }
2677     }
2678   }
2679 
2680   static const CostTblEntry SLMCostTbl[] = {
2681     // slm pcmpeq/pcmpgt throughput is 2
2682     { ISD::SETCC,   MVT::v2i64,   2 },
2683   };
2684 
2685   static const CostTblEntry AVX512BWCostTbl[] = {
2686     { ISD::SETCC,   MVT::v32i16,  1 },
2687     { ISD::SETCC,   MVT::v64i8,   1 },
2688 
2689     { ISD::SELECT,  MVT::v32i16,  1 },
2690     { ISD::SELECT,  MVT::v64i8,   1 },
2691   };
2692 
2693   static const CostTblEntry AVX512CostTbl[] = {
2694     { ISD::SETCC,   MVT::v8i64,   1 },
2695     { ISD::SETCC,   MVT::v16i32,  1 },
2696     { ISD::SETCC,   MVT::v8f64,   1 },
2697     { ISD::SETCC,   MVT::v16f32,  1 },
2698 
2699     { ISD::SELECT,  MVT::v8i64,   1 },
2700     { ISD::SELECT,  MVT::v4i64,   1 },
2701     { ISD::SELECT,  MVT::v2i64,   1 },
2702     { ISD::SELECT,  MVT::v16i32,  1 },
2703     { ISD::SELECT,  MVT::v8i32,   1 },
2704     { ISD::SELECT,  MVT::v4i32,   1 },
2705     { ISD::SELECT,  MVT::v8f64,   1 },
2706     { ISD::SELECT,  MVT::v4f64,   1 },
2707     { ISD::SELECT,  MVT::v2f64,   1 },
2708     { ISD::SELECT,  MVT::f64,     1 },
2709     { ISD::SELECT,  MVT::v16f32,  1 },
2710     { ISD::SELECT,  MVT::v8f32 ,  1 },
2711     { ISD::SELECT,  MVT::v4f32,   1 },
2712     { ISD::SELECT,  MVT::f32  ,   1 },
2713 
2714     { ISD::SETCC,   MVT::v32i16,  2 }, // FIXME: should probably be 4
2715     { ISD::SETCC,   MVT::v64i8,   2 }, // FIXME: should probably be 4
2716 
2717     { ISD::SELECT,  MVT::v32i16,  2 },
2718     { ISD::SELECT,  MVT::v16i16,  1 },
2719     { ISD::SELECT,  MVT::v8i16,   1 },
2720     { ISD::SELECT,  MVT::v64i8,   2 },
2721     { ISD::SELECT,  MVT::v32i8,   1 },
2722     { ISD::SELECT,  MVT::v16i8,   1 },
2723   };
2724 
2725   static const CostTblEntry AVX2CostTbl[] = {
2726     { ISD::SETCC,   MVT::v4i64,   1 },
2727     { ISD::SETCC,   MVT::v8i32,   1 },
2728     { ISD::SETCC,   MVT::v16i16,  1 },
2729     { ISD::SETCC,   MVT::v32i8,   1 },
2730 
2731     { ISD::SELECT,  MVT::v4f64,   2 }, // vblendvpd
2732     { ISD::SELECT,  MVT::v8f32,   2 }, // vblendvps
2733     { ISD::SELECT,  MVT::v4i64,   2 }, // pblendvb
2734     { ISD::SELECT,  MVT::v8i32,   2 }, // pblendvb
2735     { ISD::SELECT,  MVT::v16i16,  2 }, // pblendvb
2736     { ISD::SELECT,  MVT::v32i8,   2 }, // pblendvb
2737   };
2738 
2739   static const CostTblEntry AVX1CostTbl[] = {
2740     { ISD::SETCC,   MVT::v4f64,   1 },
2741     { ISD::SETCC,   MVT::v8f32,   1 },
2742     // AVX1 does not support 8-wide integer compare.
2743     { ISD::SETCC,   MVT::v4i64,   4 },
2744     { ISD::SETCC,   MVT::v8i32,   4 },
2745     { ISD::SETCC,   MVT::v16i16,  4 },
2746     { ISD::SETCC,   MVT::v32i8,   4 },
2747 
2748     { ISD::SELECT,  MVT::v4f64,   3 }, // vblendvpd
2749     { ISD::SELECT,  MVT::v8f32,   3 }, // vblendvps
2750     { ISD::SELECT,  MVT::v4i64,   3 }, // vblendvpd
2751     { ISD::SELECT,  MVT::v8i32,   3 }, // vblendvps
2752     { ISD::SELECT,  MVT::v16i16,  3 }, // vandps + vandnps + vorps
2753     { ISD::SELECT,  MVT::v32i8,   3 }, // vandps + vandnps + vorps
2754   };
2755 
2756   static const CostTblEntry SSE42CostTbl[] = {
2757     { ISD::SETCC,   MVT::v2i64,   1 },
2758   };
2759 
2760   static const CostTblEntry SSE41CostTbl[] = {
2761     { ISD::SETCC,   MVT::v2f64,   1 },
2762     { ISD::SETCC,   MVT::v4f32,   1 },
2763 
2764     { ISD::SELECT,  MVT::v2f64,   2 }, // blendvpd
2765     { ISD::SELECT,  MVT::f64,     2 }, // blendvpd
2766     { ISD::SELECT,  MVT::v4f32,   2 }, // blendvps
2767     { ISD::SELECT,  MVT::f32  ,   2 }, // blendvps
2768     { ISD::SELECT,  MVT::v2i64,   2 }, // pblendvb
2769     { ISD::SELECT,  MVT::v4i32,   2 }, // pblendvb
2770     { ISD::SELECT,  MVT::v8i16,   2 }, // pblendvb
2771     { ISD::SELECT,  MVT::v16i8,   2 }, // pblendvb
2772   };
2773 
2774   static const CostTblEntry SSE2CostTbl[] = {
2775     { ISD::SETCC,   MVT::v2f64,   2 },
2776     { ISD::SETCC,   MVT::f64,     1 },
2777     { ISD::SETCC,   MVT::v2i64,   5 }, // pcmpeqd/pcmpgtd expansion
2778     { ISD::SETCC,   MVT::v4i32,   1 },
2779     { ISD::SETCC,   MVT::v8i16,   1 },
2780     { ISD::SETCC,   MVT::v16i8,   1 },
2781 
2782     { ISD::SELECT,  MVT::v2f64,   2 }, // andpd + andnpd + orpd
2783     { ISD::SELECT,  MVT::f64,     2 }, // andpd + andnpd + orpd
2784     { ISD::SELECT,  MVT::v2i64,   2 }, // pand + pandn + por
2785     { ISD::SELECT,  MVT::v4i32,   2 }, // pand + pandn + por
2786     { ISD::SELECT,  MVT::v8i16,   2 }, // pand + pandn + por
2787     { ISD::SELECT,  MVT::v16i8,   2 }, // pand + pandn + por
2788   };
2789 
2790   static const CostTblEntry SSE1CostTbl[] = {
2791     { ISD::SETCC,   MVT::v4f32,   2 },
2792     { ISD::SETCC,   MVT::f32,     1 },
2793 
2794     { ISD::SELECT,  MVT::v4f32,   2 }, // andps + andnps + orps
2795     { ISD::SELECT,  MVT::f32,     2 }, // andps + andnps + orps
2796   };
2797 
2798   if (ST->useSLMArithCosts())
2799     if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy))
2800       return LT.first * (ExtraCost + Entry->Cost);
2801 
2802   if (ST->hasBWI())
2803     if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
2804       return LT.first * (ExtraCost + Entry->Cost);
2805 
2806   if (ST->hasAVX512())
2807     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
2808       return LT.first * (ExtraCost + Entry->Cost);
2809 
2810   if (ST->hasAVX2())
2811     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
2812       return LT.first * (ExtraCost + Entry->Cost);
2813 
2814   if (ST->hasAVX())
2815     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
2816       return LT.first * (ExtraCost + Entry->Cost);
2817 
2818   if (ST->hasSSE42())
2819     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
2820       return LT.first * (ExtraCost + Entry->Cost);
2821 
2822   if (ST->hasSSE41())
2823     if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
2824       return LT.first * (ExtraCost + Entry->Cost);
2825 
2826   if (ST->hasSSE2())
2827     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
2828       return LT.first * (ExtraCost + Entry->Cost);
2829 
2830   if (ST->hasSSE1())
2831     if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
2832       return LT.first * (ExtraCost + Entry->Cost);
2833 
2834   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I);
2835 }
2836 
2837 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; }
2838 
2839 InstructionCost
2840 X86TTIImpl::getTypeBasedIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
2841                                            TTI::TargetCostKind CostKind) {
2842 
2843   // Costs should match the codegen from:
2844   // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll
2845   // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll
2846   // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll
2847   // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll
2848   // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll
2849 
2850   // TODO: Overflow intrinsics (*ADDO, *SUBO, *MULO) with vector types are not
2851   //       specialized in these tables yet.
2852   static const CostTblEntry AVX512BITALGCostTbl[] = {
2853     { ISD::CTPOP,      MVT::v32i16,  1 },
2854     { ISD::CTPOP,      MVT::v64i8,   1 },
2855     { ISD::CTPOP,      MVT::v16i16,  1 },
2856     { ISD::CTPOP,      MVT::v32i8,   1 },
2857     { ISD::CTPOP,      MVT::v8i16,   1 },
2858     { ISD::CTPOP,      MVT::v16i8,   1 },
2859   };
2860   static const CostTblEntry AVX512VPOPCNTDQCostTbl[] = {
2861     { ISD::CTPOP,      MVT::v8i64,   1 },
2862     { ISD::CTPOP,      MVT::v16i32,  1 },
2863     { ISD::CTPOP,      MVT::v4i64,   1 },
2864     { ISD::CTPOP,      MVT::v8i32,   1 },
2865     { ISD::CTPOP,      MVT::v2i64,   1 },
2866     { ISD::CTPOP,      MVT::v4i32,   1 },
2867   };
2868   static const CostTblEntry AVX512CDCostTbl[] = {
2869     { ISD::CTLZ,       MVT::v8i64,   1 },
2870     { ISD::CTLZ,       MVT::v16i32,  1 },
2871     { ISD::CTLZ,       MVT::v32i16,  8 },
2872     { ISD::CTLZ,       MVT::v64i8,  20 },
2873     { ISD::CTLZ,       MVT::v4i64,   1 },
2874     { ISD::CTLZ,       MVT::v8i32,   1 },
2875     { ISD::CTLZ,       MVT::v16i16,  4 },
2876     { ISD::CTLZ,       MVT::v32i8,  10 },
2877     { ISD::CTLZ,       MVT::v2i64,   1 },
2878     { ISD::CTLZ,       MVT::v4i32,   1 },
2879     { ISD::CTLZ,       MVT::v8i16,   4 },
2880     { ISD::CTLZ,       MVT::v16i8,   4 },
2881   };
2882   static const CostTblEntry AVX512BWCostTbl[] = {
2883     { ISD::ABS,        MVT::v32i16,  1 },
2884     { ISD::ABS,        MVT::v64i8,   1 },
2885     { ISD::BITREVERSE, MVT::v8i64,   3 },
2886     { ISD::BITREVERSE, MVT::v16i32,  3 },
2887     { ISD::BITREVERSE, MVT::v32i16,  3 },
2888     { ISD::BITREVERSE, MVT::v64i8,   2 },
2889     { ISD::BSWAP,      MVT::v8i64,   1 },
2890     { ISD::BSWAP,      MVT::v16i32,  1 },
2891     { ISD::BSWAP,      MVT::v32i16,  1 },
2892     { ISD::CTLZ,       MVT::v8i64,  23 },
2893     { ISD::CTLZ,       MVT::v16i32, 22 },
2894     { ISD::CTLZ,       MVT::v32i16, 18 },
2895     { ISD::CTLZ,       MVT::v64i8,  17 },
2896     { ISD::CTPOP,      MVT::v8i64,   7 },
2897     { ISD::CTPOP,      MVT::v16i32, 11 },
2898     { ISD::CTPOP,      MVT::v32i16,  9 },
2899     { ISD::CTPOP,      MVT::v64i8,   6 },
2900     { ISD::CTTZ,       MVT::v8i64,  10 },
2901     { ISD::CTTZ,       MVT::v16i32, 14 },
2902     { ISD::CTTZ,       MVT::v32i16, 12 },
2903     { ISD::CTTZ,       MVT::v64i8,   9 },
2904     { ISD::SADDSAT,    MVT::v32i16,  1 },
2905     { ISD::SADDSAT,    MVT::v64i8,   1 },
2906     { ISD::SMAX,       MVT::v32i16,  1 },
2907     { ISD::SMAX,       MVT::v64i8,   1 },
2908     { ISD::SMIN,       MVT::v32i16,  1 },
2909     { ISD::SMIN,       MVT::v64i8,   1 },
2910     { ISD::SSUBSAT,    MVT::v32i16,  1 },
2911     { ISD::SSUBSAT,    MVT::v64i8,   1 },
2912     { ISD::UADDSAT,    MVT::v32i16,  1 },
2913     { ISD::UADDSAT,    MVT::v64i8,   1 },
2914     { ISD::UMAX,       MVT::v32i16,  1 },
2915     { ISD::UMAX,       MVT::v64i8,   1 },
2916     { ISD::UMIN,       MVT::v32i16,  1 },
2917     { ISD::UMIN,       MVT::v64i8,   1 },
2918     { ISD::USUBSAT,    MVT::v32i16,  1 },
2919     { ISD::USUBSAT,    MVT::v64i8,   1 },
2920   };
2921   static const CostTblEntry AVX512CostTbl[] = {
2922     { ISD::ABS,        MVT::v8i64,   1 },
2923     { ISD::ABS,        MVT::v16i32,  1 },
2924     { ISD::ABS,        MVT::v32i16,  2 },
2925     { ISD::ABS,        MVT::v64i8,   2 },
2926     { ISD::ABS,        MVT::v4i64,   1 },
2927     { ISD::ABS,        MVT::v2i64,   1 },
2928     { ISD::BITREVERSE, MVT::v8i64,  36 },
2929     { ISD::BITREVERSE, MVT::v16i32, 24 },
2930     { ISD::BITREVERSE, MVT::v32i16, 10 },
2931     { ISD::BITREVERSE, MVT::v64i8,  10 },
2932     { ISD::BSWAP,      MVT::v8i64,   4 },
2933     { ISD::BSWAP,      MVT::v16i32,  4 },
2934     { ISD::BSWAP,      MVT::v32i16,  4 },
2935     { ISD::CTLZ,       MVT::v8i64,  29 },
2936     { ISD::CTLZ,       MVT::v16i32, 35 },
2937     { ISD::CTLZ,       MVT::v32i16, 28 },
2938     { ISD::CTLZ,       MVT::v64i8,  18 },
2939     { ISD::CTPOP,      MVT::v8i64,  16 },
2940     { ISD::CTPOP,      MVT::v16i32, 24 },
2941     { ISD::CTPOP,      MVT::v32i16, 18 },
2942     { ISD::CTPOP,      MVT::v64i8,  12 },
2943     { ISD::CTTZ,       MVT::v8i64,  20 },
2944     { ISD::CTTZ,       MVT::v16i32, 28 },
2945     { ISD::CTTZ,       MVT::v32i16, 24 },
2946     { ISD::CTTZ,       MVT::v64i8,  18 },
2947     { ISD::SMAX,       MVT::v8i64,   1 },
2948     { ISD::SMAX,       MVT::v16i32,  1 },
2949     { ISD::SMAX,       MVT::v32i16,  2 },
2950     { ISD::SMAX,       MVT::v64i8,   2 },
2951     { ISD::SMAX,       MVT::v4i64,   1 },
2952     { ISD::SMAX,       MVT::v2i64,   1 },
2953     { ISD::SMIN,       MVT::v8i64,   1 },
2954     { ISD::SMIN,       MVT::v16i32,  1 },
2955     { ISD::SMIN,       MVT::v32i16,  2 },
2956     { ISD::SMIN,       MVT::v64i8,   2 },
2957     { ISD::SMIN,       MVT::v4i64,   1 },
2958     { ISD::SMIN,       MVT::v2i64,   1 },
2959     { ISD::UMAX,       MVT::v8i64,   1 },
2960     { ISD::UMAX,       MVT::v16i32,  1 },
2961     { ISD::UMAX,       MVT::v32i16,  2 },
2962     { ISD::UMAX,       MVT::v64i8,   2 },
2963     { ISD::UMAX,       MVT::v4i64,   1 },
2964     { ISD::UMAX,       MVT::v2i64,   1 },
2965     { ISD::UMIN,       MVT::v8i64,   1 },
2966     { ISD::UMIN,       MVT::v16i32,  1 },
2967     { ISD::UMIN,       MVT::v32i16,  2 },
2968     { ISD::UMIN,       MVT::v64i8,   2 },
2969     { ISD::UMIN,       MVT::v4i64,   1 },
2970     { ISD::UMIN,       MVT::v2i64,   1 },
2971     { ISD::USUBSAT,    MVT::v16i32,  2 }, // pmaxud + psubd
2972     { ISD::USUBSAT,    MVT::v2i64,   2 }, // pmaxuq + psubq
2973     { ISD::USUBSAT,    MVT::v4i64,   2 }, // pmaxuq + psubq
2974     { ISD::USUBSAT,    MVT::v8i64,   2 }, // pmaxuq + psubq
2975     { ISD::UADDSAT,    MVT::v16i32,  3 }, // not + pminud + paddd
2976     { ISD::UADDSAT,    MVT::v2i64,   3 }, // not + pminuq + paddq
2977     { ISD::UADDSAT,    MVT::v4i64,   3 }, // not + pminuq + paddq
2978     { ISD::UADDSAT,    MVT::v8i64,   3 }, // not + pminuq + paddq
2979     { ISD::SADDSAT,    MVT::v32i16,  2 },
2980     { ISD::SADDSAT,    MVT::v64i8,   2 },
2981     { ISD::SSUBSAT,    MVT::v32i16,  2 },
2982     { ISD::SSUBSAT,    MVT::v64i8,   2 },
2983     { ISD::UADDSAT,    MVT::v32i16,  2 },
2984     { ISD::UADDSAT,    MVT::v64i8,   2 },
2985     { ISD::USUBSAT,    MVT::v32i16,  2 },
2986     { ISD::USUBSAT,    MVT::v64i8,   2 },
2987     { ISD::FMAXNUM,    MVT::f32,     2 },
2988     { ISD::FMAXNUM,    MVT::v4f32,   2 },
2989     { ISD::FMAXNUM,    MVT::v8f32,   2 },
2990     { ISD::FMAXNUM,    MVT::v16f32,  2 },
2991     { ISD::FMAXNUM,    MVT::f64,     2 },
2992     { ISD::FMAXNUM,    MVT::v2f64,   2 },
2993     { ISD::FMAXNUM,    MVT::v4f64,   2 },
2994     { ISD::FMAXNUM,    MVT::v8f64,   2 },
2995   };
2996   static const CostTblEntry XOPCostTbl[] = {
2997     { ISD::BITREVERSE, MVT::v4i64,   4 },
2998     { ISD::BITREVERSE, MVT::v8i32,   4 },
2999     { ISD::BITREVERSE, MVT::v16i16,  4 },
3000     { ISD::BITREVERSE, MVT::v32i8,   4 },
3001     { ISD::BITREVERSE, MVT::v2i64,   1 },
3002     { ISD::BITREVERSE, MVT::v4i32,   1 },
3003     { ISD::BITREVERSE, MVT::v8i16,   1 },
3004     { ISD::BITREVERSE, MVT::v16i8,   1 },
3005     { ISD::BITREVERSE, MVT::i64,     3 },
3006     { ISD::BITREVERSE, MVT::i32,     3 },
3007     { ISD::BITREVERSE, MVT::i16,     3 },
3008     { ISD::BITREVERSE, MVT::i8,      3 }
3009   };
3010   static const CostTblEntry AVX2CostTbl[] = {
3011     { ISD::ABS,        MVT::v4i64,   2 }, // VBLENDVPD(X,VPSUBQ(0,X),X)
3012     { ISD::ABS,        MVT::v8i32,   1 },
3013     { ISD::ABS,        MVT::v16i16,  1 },
3014     { ISD::ABS,        MVT::v32i8,   1 },
3015     { ISD::BITREVERSE, MVT::v2i64,   3 },
3016     { ISD::BITREVERSE, MVT::v4i64,   3 },
3017     { ISD::BITREVERSE, MVT::v4i32,   3 },
3018     { ISD::BITREVERSE, MVT::v8i32,   3 },
3019     { ISD::BITREVERSE, MVT::v8i16,   3 },
3020     { ISD::BITREVERSE, MVT::v16i16,  3 },
3021     { ISD::BITREVERSE, MVT::v16i8,   3 },
3022     { ISD::BITREVERSE, MVT::v32i8,   3 },
3023     { ISD::BSWAP,      MVT::v4i64,   1 },
3024     { ISD::BSWAP,      MVT::v8i32,   1 },
3025     { ISD::BSWAP,      MVT::v16i16,  1 },
3026     { ISD::CTLZ,       MVT::v2i64,   7 },
3027     { ISD::CTLZ,       MVT::v4i64,   7 },
3028     { ISD::CTLZ,       MVT::v4i32,   5 },
3029     { ISD::CTLZ,       MVT::v8i32,   5 },
3030     { ISD::CTLZ,       MVT::v8i16,   4 },
3031     { ISD::CTLZ,       MVT::v16i16,  4 },
3032     { ISD::CTLZ,       MVT::v16i8,   3 },
3033     { ISD::CTLZ,       MVT::v32i8,   3 },
3034     { ISD::CTPOP,      MVT::v2i64,   3 },
3035     { ISD::CTPOP,      MVT::v4i64,   3 },
3036     { ISD::CTPOP,      MVT::v4i32,   7 },
3037     { ISD::CTPOP,      MVT::v8i32,   7 },
3038     { ISD::CTPOP,      MVT::v8i16,   3 },
3039     { ISD::CTPOP,      MVT::v16i16,  3 },
3040     { ISD::CTPOP,      MVT::v16i8,   2 },
3041     { ISD::CTPOP,      MVT::v32i8,   2 },
3042     { ISD::CTTZ,       MVT::v2i64,   4 },
3043     { ISD::CTTZ,       MVT::v4i64,   4 },
3044     { ISD::CTTZ,       MVT::v4i32,   7 },
3045     { ISD::CTTZ,       MVT::v8i32,   7 },
3046     { ISD::CTTZ,       MVT::v8i16,   4 },
3047     { ISD::CTTZ,       MVT::v16i16,  4 },
3048     { ISD::CTTZ,       MVT::v16i8,   3 },
3049     { ISD::CTTZ,       MVT::v32i8,   3 },
3050     { ISD::SADDSAT,    MVT::v16i16,  1 },
3051     { ISD::SADDSAT,    MVT::v32i8,   1 },
3052     { ISD::SMAX,       MVT::v8i32,   1 },
3053     { ISD::SMAX,       MVT::v16i16,  1 },
3054     { ISD::SMAX,       MVT::v32i8,   1 },
3055     { ISD::SMIN,       MVT::v8i32,   1 },
3056     { ISD::SMIN,       MVT::v16i16,  1 },
3057     { ISD::SMIN,       MVT::v32i8,   1 },
3058     { ISD::SSUBSAT,    MVT::v16i16,  1 },
3059     { ISD::SSUBSAT,    MVT::v32i8,   1 },
3060     { ISD::UADDSAT,    MVT::v16i16,  1 },
3061     { ISD::UADDSAT,    MVT::v32i8,   1 },
3062     { ISD::UADDSAT,    MVT::v8i32,   3 }, // not + pminud + paddd
3063     { ISD::UMAX,       MVT::v8i32,   1 },
3064     { ISD::UMAX,       MVT::v16i16,  1 },
3065     { ISD::UMAX,       MVT::v32i8,   1 },
3066     { ISD::UMIN,       MVT::v8i32,   1 },
3067     { ISD::UMIN,       MVT::v16i16,  1 },
3068     { ISD::UMIN,       MVT::v32i8,   1 },
3069     { ISD::USUBSAT,    MVT::v16i16,  1 },
3070     { ISD::USUBSAT,    MVT::v32i8,   1 },
3071     { ISD::USUBSAT,    MVT::v8i32,   2 }, // pmaxud + psubd
3072     { ISD::FMAXNUM,    MVT::v8f32,   3 }, // MAXPS + CMPUNORDPS + BLENDVPS
3073     { ISD::FMAXNUM,    MVT::v4f64,   3 }, // MAXPD + CMPUNORDPD + BLENDVPD
3074     { ISD::FSQRT,      MVT::f32,     7 }, // Haswell from http://www.agner.org/
3075     { ISD::FSQRT,      MVT::v4f32,   7 }, // Haswell from http://www.agner.org/
3076     { ISD::FSQRT,      MVT::v8f32,  14 }, // Haswell from http://www.agner.org/
3077     { ISD::FSQRT,      MVT::f64,    14 }, // Haswell from http://www.agner.org/
3078     { ISD::FSQRT,      MVT::v2f64,  14 }, // Haswell from http://www.agner.org/
3079     { ISD::FSQRT,      MVT::v4f64,  28 }, // Haswell from http://www.agner.org/
3080   };
3081   static const CostTblEntry AVX1CostTbl[] = {
3082     { ISD::ABS,        MVT::v4i64,   5 }, // VBLENDVPD(X,VPSUBQ(0,X),X)
3083     { ISD::ABS,        MVT::v8i32,   3 },
3084     { ISD::ABS,        MVT::v16i16,  3 },
3085     { ISD::ABS,        MVT::v32i8,   3 },
3086     { ISD::BITREVERSE, MVT::v4i64,  12 }, // 2 x 128-bit Op + extract/insert
3087     { ISD::BITREVERSE, MVT::v8i32,  12 }, // 2 x 128-bit Op + extract/insert
3088     { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert
3089     { ISD::BITREVERSE, MVT::v32i8,  12 }, // 2 x 128-bit Op + extract/insert
3090     { ISD::BSWAP,      MVT::v4i64,   4 },
3091     { ISD::BSWAP,      MVT::v8i32,   4 },
3092     { ISD::BSWAP,      MVT::v16i16,  4 },
3093     { ISD::CTLZ,       MVT::v4i64,  48 }, // 2 x 128-bit Op + extract/insert
3094     { ISD::CTLZ,       MVT::v8i32,  38 }, // 2 x 128-bit Op + extract/insert
3095     { ISD::CTLZ,       MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert
3096     { ISD::CTLZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
3097     { ISD::CTPOP,      MVT::v4i64,  16 }, // 2 x 128-bit Op + extract/insert
3098     { ISD::CTPOP,      MVT::v8i32,  24 }, // 2 x 128-bit Op + extract/insert
3099     { ISD::CTPOP,      MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert
3100     { ISD::CTPOP,      MVT::v32i8,  14 }, // 2 x 128-bit Op + extract/insert
3101     { ISD::CTTZ,       MVT::v4i64,  22 }, // 2 x 128-bit Op + extract/insert
3102     { ISD::CTTZ,       MVT::v8i32,  30 }, // 2 x 128-bit Op + extract/insert
3103     { ISD::CTTZ,       MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert
3104     { ISD::CTTZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
3105     { ISD::SADDSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3106     { ISD::SADDSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3107     { ISD::SMAX,       MVT::v8i32,   4 }, // 2 x 128-bit Op + extract/insert
3108     { ISD::SMAX,       MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3109     { ISD::SMAX,       MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3110     { ISD::SMIN,       MVT::v8i32,   4 }, // 2 x 128-bit Op + extract/insert
3111     { ISD::SMIN,       MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3112     { ISD::SMIN,       MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3113     { ISD::SSUBSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3114     { ISD::SSUBSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3115     { ISD::UADDSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3116     { ISD::UADDSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3117     { ISD::UADDSAT,    MVT::v8i32,   8 }, // 2 x 128-bit Op + extract/insert
3118     { ISD::UMAX,       MVT::v8i32,   4 }, // 2 x 128-bit Op + extract/insert
3119     { ISD::UMAX,       MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3120     { ISD::UMAX,       MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3121     { ISD::UMIN,       MVT::v8i32,   4 }, // 2 x 128-bit Op + extract/insert
3122     { ISD::UMIN,       MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3123     { ISD::UMIN,       MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3124     { ISD::USUBSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
3125     { ISD::USUBSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
3126     { ISD::USUBSAT,    MVT::v8i32,   6 }, // 2 x 128-bit Op + extract/insert
3127     { ISD::FMAXNUM,    MVT::f32,     3 }, // MAXSS + CMPUNORDSS + BLENDVPS
3128     { ISD::FMAXNUM,    MVT::v4f32,   3 }, // MAXPS + CMPUNORDPS + BLENDVPS
3129     { ISD::FMAXNUM,    MVT::v8f32,   5 }, // MAXPS + CMPUNORDPS + BLENDVPS + ?
3130     { ISD::FMAXNUM,    MVT::f64,     3 }, // MAXSD + CMPUNORDSD + BLENDVPD
3131     { ISD::FMAXNUM,    MVT::v2f64,   3 }, // MAXPD + CMPUNORDPD + BLENDVPD
3132     { ISD::FMAXNUM,    MVT::v4f64,   5 }, // MAXPD + CMPUNORDPD + BLENDVPD + ?
3133     { ISD::FSQRT,      MVT::f32,    14 }, // SNB from http://www.agner.org/
3134     { ISD::FSQRT,      MVT::v4f32,  14 }, // SNB from http://www.agner.org/
3135     { ISD::FSQRT,      MVT::v8f32,  28 }, // SNB from http://www.agner.org/
3136     { ISD::FSQRT,      MVT::f64,    21 }, // SNB from http://www.agner.org/
3137     { ISD::FSQRT,      MVT::v2f64,  21 }, // SNB from http://www.agner.org/
3138     { ISD::FSQRT,      MVT::v4f64,  43 }, // SNB from http://www.agner.org/
3139   };
3140   static const CostTblEntry GLMCostTbl[] = {
3141     { ISD::FSQRT, MVT::f32,   19 }, // sqrtss
3142     { ISD::FSQRT, MVT::v4f32, 37 }, // sqrtps
3143     { ISD::FSQRT, MVT::f64,   34 }, // sqrtsd
3144     { ISD::FSQRT, MVT::v2f64, 67 }, // sqrtpd
3145   };
3146   static const CostTblEntry SLMCostTbl[] = {
3147     { ISD::FSQRT, MVT::f32,   20 }, // sqrtss
3148     { ISD::FSQRT, MVT::v4f32, 40 }, // sqrtps
3149     { ISD::FSQRT, MVT::f64,   35 }, // sqrtsd
3150     { ISD::FSQRT, MVT::v2f64, 70 }, // sqrtpd
3151   };
3152   static const CostTblEntry SSE42CostTbl[] = {
3153     { ISD::USUBSAT,    MVT::v4i32,   2 }, // pmaxud + psubd
3154     { ISD::UADDSAT,    MVT::v4i32,   3 }, // not + pminud + paddd
3155     { ISD::FSQRT,      MVT::f32,    18 }, // Nehalem from http://www.agner.org/
3156     { ISD::FSQRT,      MVT::v4f32,  18 }, // Nehalem from http://www.agner.org/
3157   };
3158   static const CostTblEntry SSE41CostTbl[] = {
3159     { ISD::ABS,        MVT::v2i64,   2 }, // BLENDVPD(X,PSUBQ(0,X),X)
3160     { ISD::SMAX,       MVT::v4i32,   1 },
3161     { ISD::SMAX,       MVT::v16i8,   1 },
3162     { ISD::SMIN,       MVT::v4i32,   1 },
3163     { ISD::SMIN,       MVT::v16i8,   1 },
3164     { ISD::UMAX,       MVT::v4i32,   1 },
3165     { ISD::UMAX,       MVT::v8i16,   1 },
3166     { ISD::UMIN,       MVT::v4i32,   1 },
3167     { ISD::UMIN,       MVT::v8i16,   1 },
3168   };
3169   static const CostTblEntry SSSE3CostTbl[] = {
3170     { ISD::ABS,        MVT::v4i32,   1 },
3171     { ISD::ABS,        MVT::v8i16,   1 },
3172     { ISD::ABS,        MVT::v16i8,   1 },
3173     { ISD::BITREVERSE, MVT::v2i64,   5 },
3174     { ISD::BITREVERSE, MVT::v4i32,   5 },
3175     { ISD::BITREVERSE, MVT::v8i16,   5 },
3176     { ISD::BITREVERSE, MVT::v16i8,   5 },
3177     { ISD::BSWAP,      MVT::v2i64,   1 },
3178     { ISD::BSWAP,      MVT::v4i32,   1 },
3179     { ISD::BSWAP,      MVT::v8i16,   1 },
3180     { ISD::CTLZ,       MVT::v2i64,  23 },
3181     { ISD::CTLZ,       MVT::v4i32,  18 },
3182     { ISD::CTLZ,       MVT::v8i16,  14 },
3183     { ISD::CTLZ,       MVT::v16i8,   9 },
3184     { ISD::CTPOP,      MVT::v2i64,   7 },
3185     { ISD::CTPOP,      MVT::v4i32,  11 },
3186     { ISD::CTPOP,      MVT::v8i16,   9 },
3187     { ISD::CTPOP,      MVT::v16i8,   6 },
3188     { ISD::CTTZ,       MVT::v2i64,  10 },
3189     { ISD::CTTZ,       MVT::v4i32,  14 },
3190     { ISD::CTTZ,       MVT::v8i16,  12 },
3191     { ISD::CTTZ,       MVT::v16i8,   9 }
3192   };
3193   static const CostTblEntry SSE2CostTbl[] = {
3194     { ISD::ABS,        MVT::v2i64,   4 },
3195     { ISD::ABS,        MVT::v4i32,   3 },
3196     { ISD::ABS,        MVT::v8i16,   2 },
3197     { ISD::ABS,        MVT::v16i8,   2 },
3198     { ISD::BITREVERSE, MVT::v2i64,  29 },
3199     { ISD::BITREVERSE, MVT::v4i32,  27 },
3200     { ISD::BITREVERSE, MVT::v8i16,  27 },
3201     { ISD::BITREVERSE, MVT::v16i8,  20 },
3202     { ISD::BSWAP,      MVT::v2i64,   7 },
3203     { ISD::BSWAP,      MVT::v4i32,   7 },
3204     { ISD::BSWAP,      MVT::v8i16,   7 },
3205     { ISD::CTLZ,       MVT::v2i64,  25 },
3206     { ISD::CTLZ,       MVT::v4i32,  26 },
3207     { ISD::CTLZ,       MVT::v8i16,  20 },
3208     { ISD::CTLZ,       MVT::v16i8,  17 },
3209     { ISD::CTPOP,      MVT::v2i64,  12 },
3210     { ISD::CTPOP,      MVT::v4i32,  15 },
3211     { ISD::CTPOP,      MVT::v8i16,  13 },
3212     { ISD::CTPOP,      MVT::v16i8,  10 },
3213     { ISD::CTTZ,       MVT::v2i64,  14 },
3214     { ISD::CTTZ,       MVT::v4i32,  18 },
3215     { ISD::CTTZ,       MVT::v8i16,  16 },
3216     { ISD::CTTZ,       MVT::v16i8,  13 },
3217     { ISD::SADDSAT,    MVT::v8i16,   1 },
3218     { ISD::SADDSAT,    MVT::v16i8,   1 },
3219     { ISD::SMAX,       MVT::v8i16,   1 },
3220     { ISD::SMIN,       MVT::v8i16,   1 },
3221     { ISD::SSUBSAT,    MVT::v8i16,   1 },
3222     { ISD::SSUBSAT,    MVT::v16i8,   1 },
3223     { ISD::UADDSAT,    MVT::v8i16,   1 },
3224     { ISD::UADDSAT,    MVT::v16i8,   1 },
3225     { ISD::UMAX,       MVT::v8i16,   2 },
3226     { ISD::UMAX,       MVT::v16i8,   1 },
3227     { ISD::UMIN,       MVT::v8i16,   2 },
3228     { ISD::UMIN,       MVT::v16i8,   1 },
3229     { ISD::USUBSAT,    MVT::v8i16,   1 },
3230     { ISD::USUBSAT,    MVT::v16i8,   1 },
3231     { ISD::FMAXNUM,    MVT::f64,     4 },
3232     { ISD::FMAXNUM,    MVT::v2f64,   4 },
3233     { ISD::FSQRT,      MVT::f64,    32 }, // Nehalem from http://www.agner.org/
3234     { ISD::FSQRT,      MVT::v2f64,  32 }, // Nehalem from http://www.agner.org/
3235   };
3236   static const CostTblEntry SSE1CostTbl[] = {
3237     { ISD::FMAXNUM,    MVT::f32,     4 },
3238     { ISD::FMAXNUM,    MVT::v4f32,   4 },
3239     { ISD::FSQRT,      MVT::f32,    28 }, // Pentium III from http://www.agner.org/
3240     { ISD::FSQRT,      MVT::v4f32,  56 }, // Pentium III from http://www.agner.org/
3241   };
3242   static const CostTblEntry BMI64CostTbl[] = { // 64-bit targets
3243     { ISD::CTTZ,       MVT::i64,     1 },
3244   };
3245   static const CostTblEntry BMI32CostTbl[] = { // 32 or 64-bit targets
3246     { ISD::CTTZ,       MVT::i32,     1 },
3247     { ISD::CTTZ,       MVT::i16,     1 },
3248     { ISD::CTTZ,       MVT::i8,      1 },
3249   };
3250   static const CostTblEntry LZCNT64CostTbl[] = { // 64-bit targets
3251     { ISD::CTLZ,       MVT::i64,     1 },
3252   };
3253   static const CostTblEntry LZCNT32CostTbl[] = { // 32 or 64-bit targets
3254     { ISD::CTLZ,       MVT::i32,     1 },
3255     { ISD::CTLZ,       MVT::i16,     1 },
3256     { ISD::CTLZ,       MVT::i8,      1 },
3257   };
3258   static const CostTblEntry POPCNT64CostTbl[] = { // 64-bit targets
3259     { ISD::CTPOP,      MVT::i64,     1 },
3260   };
3261   static const CostTblEntry POPCNT32CostTbl[] = { // 32 or 64-bit targets
3262     { ISD::CTPOP,      MVT::i32,     1 },
3263     { ISD::CTPOP,      MVT::i16,     1 },
3264     { ISD::CTPOP,      MVT::i8,      1 },
3265   };
3266   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
3267     { ISD::ABS,        MVT::i64,     2 }, // SUB+CMOV
3268     { ISD::BITREVERSE, MVT::i64,    14 },
3269     { ISD::BSWAP,      MVT::i64,     1 },
3270     { ISD::CTLZ,       MVT::i64,     4 }, // BSR+XOR or BSR+XOR+CMOV
3271     { ISD::CTTZ,       MVT::i64,     3 }, // TEST+BSF+CMOV/BRANCH
3272     { ISD::CTPOP,      MVT::i64,    10 },
3273     { ISD::SADDO,      MVT::i64,     1 },
3274     { ISD::UADDO,      MVT::i64,     1 },
3275     { ISD::UMULO,      MVT::i64,     2 }, // mulq + seto
3276   };
3277   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
3278     { ISD::ABS,        MVT::i32,     2 }, // SUB+CMOV
3279     { ISD::ABS,        MVT::i16,     2 }, // SUB+CMOV
3280     { ISD::BITREVERSE, MVT::i32,    14 },
3281     { ISD::BITREVERSE, MVT::i16,    14 },
3282     { ISD::BITREVERSE, MVT::i8,     11 },
3283     { ISD::BSWAP,      MVT::i32,     1 },
3284     { ISD::BSWAP,      MVT::i16,     1 }, // ROL
3285     { ISD::CTLZ,       MVT::i32,     4 }, // BSR+XOR or BSR+XOR+CMOV
3286     { ISD::CTLZ,       MVT::i16,     4 }, // BSR+XOR or BSR+XOR+CMOV
3287     { ISD::CTLZ,       MVT::i8,      4 }, // BSR+XOR or BSR+XOR+CMOV
3288     { ISD::CTTZ,       MVT::i32,     3 }, // TEST+BSF+CMOV/BRANCH
3289     { ISD::CTTZ,       MVT::i16,     3 }, // TEST+BSF+CMOV/BRANCH
3290     { ISD::CTTZ,       MVT::i8,      3 }, // TEST+BSF+CMOV/BRANCH
3291     { ISD::CTPOP,      MVT::i32,     8 },
3292     { ISD::CTPOP,      MVT::i16,     9 },
3293     { ISD::CTPOP,      MVT::i8,      7 },
3294     { ISD::SADDO,      MVT::i32,     1 },
3295     { ISD::SADDO,      MVT::i16,     1 },
3296     { ISD::SADDO,      MVT::i8,      1 },
3297     { ISD::UADDO,      MVT::i32,     1 },
3298     { ISD::UADDO,      MVT::i16,     1 },
3299     { ISD::UADDO,      MVT::i8,      1 },
3300     { ISD::UMULO,      MVT::i32,     2 }, // mul + seto
3301     { ISD::UMULO,      MVT::i16,     2 },
3302     { ISD::UMULO,      MVT::i8,      2 },
3303   };
3304 
3305   Type *RetTy = ICA.getReturnType();
3306   Type *OpTy = RetTy;
3307   Intrinsic::ID IID = ICA.getID();
3308   unsigned ISD = ISD::DELETED_NODE;
3309   switch (IID) {
3310   default:
3311     break;
3312   case Intrinsic::abs:
3313     ISD = ISD::ABS;
3314     break;
3315   case Intrinsic::bitreverse:
3316     ISD = ISD::BITREVERSE;
3317     break;
3318   case Intrinsic::bswap:
3319     ISD = ISD::BSWAP;
3320     break;
3321   case Intrinsic::ctlz:
3322     ISD = ISD::CTLZ;
3323     break;
3324   case Intrinsic::ctpop:
3325     ISD = ISD::CTPOP;
3326     break;
3327   case Intrinsic::cttz:
3328     ISD = ISD::CTTZ;
3329     break;
3330   case Intrinsic::maxnum:
3331   case Intrinsic::minnum:
3332     // FMINNUM has same costs so don't duplicate.
3333     ISD = ISD::FMAXNUM;
3334     break;
3335   case Intrinsic::sadd_sat:
3336     ISD = ISD::SADDSAT;
3337     break;
3338   case Intrinsic::smax:
3339     ISD = ISD::SMAX;
3340     break;
3341   case Intrinsic::smin:
3342     ISD = ISD::SMIN;
3343     break;
3344   case Intrinsic::ssub_sat:
3345     ISD = ISD::SSUBSAT;
3346     break;
3347   case Intrinsic::uadd_sat:
3348     ISD = ISD::UADDSAT;
3349     break;
3350   case Intrinsic::umax:
3351     ISD = ISD::UMAX;
3352     break;
3353   case Intrinsic::umin:
3354     ISD = ISD::UMIN;
3355     break;
3356   case Intrinsic::usub_sat:
3357     ISD = ISD::USUBSAT;
3358     break;
3359   case Intrinsic::sqrt:
3360     ISD = ISD::FSQRT;
3361     break;
3362   case Intrinsic::sadd_with_overflow:
3363   case Intrinsic::ssub_with_overflow:
3364     // SSUBO has same costs so don't duplicate.
3365     ISD = ISD::SADDO;
3366     OpTy = RetTy->getContainedType(0);
3367     break;
3368   case Intrinsic::uadd_with_overflow:
3369   case Intrinsic::usub_with_overflow:
3370     // USUBO has same costs so don't duplicate.
3371     ISD = ISD::UADDO;
3372     OpTy = RetTy->getContainedType(0);
3373     break;
3374   case Intrinsic::umul_with_overflow:
3375   case Intrinsic::smul_with_overflow:
3376     // SMULO has same costs so don't duplicate.
3377     ISD = ISD::UMULO;
3378     OpTy = RetTy->getContainedType(0);
3379     break;
3380   }
3381 
3382   if (ISD != ISD::DELETED_NODE) {
3383     // Legalize the type.
3384     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, OpTy);
3385     MVT MTy = LT.second;
3386 
3387     // Attempt to lookup cost.
3388     if (ISD == ISD::BITREVERSE && ST->hasGFNI() && ST->hasSSSE3() &&
3389         MTy.isVector()) {
3390       // With PSHUFB the code is very similar for all types. If we have integer
3391       // byte operations, we just need a GF2P8AFFINEQB for vXi8. For other types
3392       // we also need a PSHUFB.
3393       unsigned Cost = MTy.getVectorElementType() == MVT::i8 ? 1 : 2;
3394 
3395       // Without byte operations, we need twice as many GF2P8AFFINEQB and PSHUFB
3396       // instructions. We also need an extract and an insert.
3397       if (!(MTy.is128BitVector() || (ST->hasAVX2() && MTy.is256BitVector()) ||
3398             (ST->hasBWI() && MTy.is512BitVector())))
3399         Cost = Cost * 2 + 2;
3400 
3401       return LT.first * Cost;
3402     }
3403 
3404     auto adjustTableCost = [](const CostTblEntry &Entry,
3405                               InstructionCost LegalizationCost,
3406                               FastMathFlags FMF) {
3407       // If there are no NANs to deal with, then these are reduced to a
3408       // single MIN** or MAX** instruction instead of the MIN/CMP/SELECT that we
3409       // assume is used in the non-fast case.
3410       if (Entry.ISD == ISD::FMAXNUM || Entry.ISD == ISD::FMINNUM) {
3411         if (FMF.noNaNs())
3412           return LegalizationCost * 1;
3413       }
3414       return LegalizationCost * (int)Entry.Cost;
3415     };
3416 
3417     if (ST->useGLMDivSqrtCosts())
3418       if (const auto *Entry = CostTableLookup(GLMCostTbl, ISD, MTy))
3419         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3420 
3421     if (ST->useSLMArithCosts())
3422       if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy))
3423         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3424 
3425     if (ST->hasBITALG())
3426       if (const auto *Entry = CostTableLookup(AVX512BITALGCostTbl, ISD, MTy))
3427         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3428 
3429     if (ST->hasVPOPCNTDQ())
3430       if (const auto *Entry = CostTableLookup(AVX512VPOPCNTDQCostTbl, ISD, MTy))
3431         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3432 
3433     if (ST->hasCDI())
3434       if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy))
3435         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3436 
3437     if (ST->hasBWI())
3438       if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
3439         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3440 
3441     if (ST->hasAVX512())
3442       if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
3443         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3444 
3445     if (ST->hasXOP())
3446       if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
3447         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3448 
3449     if (ST->hasAVX2())
3450       if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
3451         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3452 
3453     if (ST->hasAVX())
3454       if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
3455         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3456 
3457     if (ST->hasSSE42())
3458       if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
3459         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3460 
3461     if (ST->hasSSE41())
3462       if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
3463         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3464 
3465     if (ST->hasSSSE3())
3466       if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy))
3467         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3468 
3469     if (ST->hasSSE2())
3470       if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
3471         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3472 
3473     if (ST->hasSSE1())
3474       if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
3475         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3476 
3477     if (ST->hasBMI()) {
3478       if (ST->is64Bit())
3479         if (const auto *Entry = CostTableLookup(BMI64CostTbl, ISD, MTy))
3480           return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3481 
3482       if (const auto *Entry = CostTableLookup(BMI32CostTbl, ISD, MTy))
3483         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3484     }
3485 
3486     if (ST->hasLZCNT()) {
3487       if (ST->is64Bit())
3488         if (const auto *Entry = CostTableLookup(LZCNT64CostTbl, ISD, MTy))
3489           return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3490 
3491       if (const auto *Entry = CostTableLookup(LZCNT32CostTbl, ISD, MTy))
3492         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3493     }
3494 
3495     if (ST->hasPOPCNT()) {
3496       if (ST->is64Bit())
3497         if (const auto *Entry = CostTableLookup(POPCNT64CostTbl, ISD, MTy))
3498           return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3499 
3500       if (const auto *Entry = CostTableLookup(POPCNT32CostTbl, ISD, MTy))
3501         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3502     }
3503 
3504     if (ISD == ISD::BSWAP && ST->hasMOVBE() && ST->hasFastMOVBE()) {
3505       if (const Instruction *II = ICA.getInst()) {
3506         if (II->hasOneUse() && isa<StoreInst>(II->user_back()))
3507           return TTI::TCC_Free;
3508         if (auto *LI = dyn_cast<LoadInst>(II->getOperand(0))) {
3509           if (LI->hasOneUse())
3510             return TTI::TCC_Free;
3511         }
3512       }
3513     }
3514 
3515     if (ST->is64Bit())
3516       if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
3517         return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3518 
3519     if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
3520       return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3521   }
3522 
3523   return BaseT::getIntrinsicInstrCost(ICA, CostKind);
3524 }
3525 
3526 InstructionCost
3527 X86TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
3528                                   TTI::TargetCostKind CostKind) {
3529   if (ICA.isTypeBasedOnly())
3530     return getTypeBasedIntrinsicInstrCost(ICA, CostKind);
3531 
3532   static const CostTblEntry AVX512BWCostTbl[] = {
3533     { ISD::ROTL,       MVT::v32i16,  2 },
3534     { ISD::ROTL,       MVT::v16i16,  2 },
3535     { ISD::ROTL,       MVT::v8i16,   2 },
3536     { ISD::ROTL,       MVT::v64i8,   5 },
3537     { ISD::ROTL,       MVT::v32i8,   5 },
3538     { ISD::ROTL,       MVT::v16i8,   5 },
3539     { ISD::ROTR,       MVT::v32i16,  2 },
3540     { ISD::ROTR,       MVT::v16i16,  2 },
3541     { ISD::ROTR,       MVT::v8i16,   2 },
3542     { ISD::ROTR,       MVT::v64i8,   5 },
3543     { ISD::ROTR,       MVT::v32i8,   5 },
3544     { ISD::ROTR,       MVT::v16i8,   5 }
3545   };
3546   static const CostTblEntry AVX512CostTbl[] = {
3547     { ISD::ROTL,       MVT::v8i64,   1 },
3548     { ISD::ROTL,       MVT::v4i64,   1 },
3549     { ISD::ROTL,       MVT::v2i64,   1 },
3550     { ISD::ROTL,       MVT::v16i32,  1 },
3551     { ISD::ROTL,       MVT::v8i32,   1 },
3552     { ISD::ROTL,       MVT::v4i32,   1 },
3553     { ISD::ROTR,       MVT::v8i64,   1 },
3554     { ISD::ROTR,       MVT::v4i64,   1 },
3555     { ISD::ROTR,       MVT::v2i64,   1 },
3556     { ISD::ROTR,       MVT::v16i32,  1 },
3557     { ISD::ROTR,       MVT::v8i32,   1 },
3558     { ISD::ROTR,       MVT::v4i32,   1 }
3559   };
3560   // XOP: ROTL = VPROT(X,Y), ROTR = VPROT(X,SUB(0,Y))
3561   static const CostTblEntry XOPCostTbl[] = {
3562     { ISD::ROTL,       MVT::v4i64,   4 },
3563     { ISD::ROTL,       MVT::v8i32,   4 },
3564     { ISD::ROTL,       MVT::v16i16,  4 },
3565     { ISD::ROTL,       MVT::v32i8,   4 },
3566     { ISD::ROTL,       MVT::v2i64,   1 },
3567     { ISD::ROTL,       MVT::v4i32,   1 },
3568     { ISD::ROTL,       MVT::v8i16,   1 },
3569     { ISD::ROTL,       MVT::v16i8,   1 },
3570     { ISD::ROTR,       MVT::v4i64,   6 },
3571     { ISD::ROTR,       MVT::v8i32,   6 },
3572     { ISD::ROTR,       MVT::v16i16,  6 },
3573     { ISD::ROTR,       MVT::v32i8,   6 },
3574     { ISD::ROTR,       MVT::v2i64,   2 },
3575     { ISD::ROTR,       MVT::v4i32,   2 },
3576     { ISD::ROTR,       MVT::v8i16,   2 },
3577     { ISD::ROTR,       MVT::v16i8,   2 }
3578   };
3579   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
3580     { ISD::ROTL,       MVT::i64,     1 },
3581     { ISD::ROTR,       MVT::i64,     1 },
3582     { ISD::FSHL,       MVT::i64,     4 }
3583   };
3584   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
3585     { ISD::ROTL,       MVT::i32,     1 },
3586     { ISD::ROTL,       MVT::i16,     1 },
3587     { ISD::ROTL,       MVT::i8,      1 },
3588     { ISD::ROTR,       MVT::i32,     1 },
3589     { ISD::ROTR,       MVT::i16,     1 },
3590     { ISD::ROTR,       MVT::i8,      1 },
3591     { ISD::FSHL,       MVT::i32,     4 },
3592     { ISD::FSHL,       MVT::i16,     4 },
3593     { ISD::FSHL,       MVT::i8,      4 }
3594   };
3595 
3596   Intrinsic::ID IID = ICA.getID();
3597   Type *RetTy = ICA.getReturnType();
3598   const SmallVectorImpl<const Value *> &Args = ICA.getArgs();
3599   unsigned ISD = ISD::DELETED_NODE;
3600   switch (IID) {
3601   default:
3602     break;
3603   case Intrinsic::fshl:
3604     ISD = ISD::FSHL;
3605     if (Args[0] == Args[1])
3606       ISD = ISD::ROTL;
3607     break;
3608   case Intrinsic::fshr:
3609     // FSHR has same costs so don't duplicate.
3610     ISD = ISD::FSHL;
3611     if (Args[0] == Args[1])
3612       ISD = ISD::ROTR;
3613     break;
3614   }
3615 
3616   if (ISD != ISD::DELETED_NODE) {
3617     // Legalize the type.
3618     std::pair<InstructionCost, MVT> LT =
3619         TLI->getTypeLegalizationCost(DL, RetTy);
3620     MVT MTy = LT.second;
3621 
3622     // Attempt to lookup cost.
3623     if (ST->hasBWI())
3624       if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
3625         return LT.first * Entry->Cost;
3626 
3627     if (ST->hasAVX512())
3628       if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
3629         return LT.first * Entry->Cost;
3630 
3631     if (ST->hasXOP())
3632       if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
3633         return LT.first * Entry->Cost;
3634 
3635     if (ST->is64Bit())
3636       if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
3637         return LT.first * Entry->Cost;
3638 
3639     if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
3640       return LT.first * Entry->Cost;
3641   }
3642 
3643   return BaseT::getIntrinsicInstrCost(ICA, CostKind);
3644 }
3645 
3646 InstructionCost X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
3647                                                unsigned Index) {
3648   static const CostTblEntry SLMCostTbl[] = {
3649      { ISD::EXTRACT_VECTOR_ELT,       MVT::i8,      4 },
3650      { ISD::EXTRACT_VECTOR_ELT,       MVT::i16,     4 },
3651      { ISD::EXTRACT_VECTOR_ELT,       MVT::i32,     4 },
3652      { ISD::EXTRACT_VECTOR_ELT,       MVT::i64,     7 }
3653    };
3654 
3655   assert(Val->isVectorTy() && "This must be a vector type");
3656   Type *ScalarType = Val->getScalarType();
3657   InstructionCost RegisterFileMoveCost = 0;
3658 
3659   // Non-immediate extraction/insertion can be handled as a sequence of
3660   // aliased loads+stores via the stack.
3661   if (Index == -1U && (Opcode == Instruction::ExtractElement ||
3662                        Opcode == Instruction::InsertElement)) {
3663     // TODO: On some SSE41+ targets, we expand to cmp+splat+select patterns:
3664     // inselt N0, N1, N2 --> select (SplatN2 == {0,1,2...}) ? SplatN1 : N0.
3665 
3666     // TODO: Move this to BasicTTIImpl.h? We'd need better gep + index handling.
3667     assert(isa<FixedVectorType>(Val) && "Fixed vector type expected");
3668     Align VecAlign = DL.getPrefTypeAlign(Val);
3669     Align SclAlign = DL.getPrefTypeAlign(ScalarType);
3670 
3671     // Extract - store vector to stack, load scalar.
3672     if (Opcode == Instruction::ExtractElement) {
3673       return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0,
3674                              TTI::TargetCostKind::TCK_RecipThroughput) +
3675              getMemoryOpCost(Instruction::Load, ScalarType, SclAlign, 0,
3676                              TTI::TargetCostKind::TCK_RecipThroughput);
3677     }
3678     // Insert - store vector to stack, store scalar, load vector.
3679     if (Opcode == Instruction::InsertElement) {
3680       return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0,
3681                              TTI::TargetCostKind::TCK_RecipThroughput) +
3682              getMemoryOpCost(Instruction::Store, ScalarType, SclAlign, 0,
3683                              TTI::TargetCostKind::TCK_RecipThroughput) +
3684              getMemoryOpCost(Instruction::Load, Val, VecAlign, 0,
3685                              TTI::TargetCostKind::TCK_RecipThroughput);
3686     }
3687   }
3688 
3689   if (Index != -1U && (Opcode == Instruction::ExtractElement ||
3690                        Opcode == Instruction::InsertElement)) {
3691     // Extraction of vXi1 elements are now efficiently handled by MOVMSK.
3692     if (Opcode == Instruction::ExtractElement &&
3693         ScalarType->getScalarSizeInBits() == 1 &&
3694         cast<FixedVectorType>(Val)->getNumElements() > 1)
3695       return 1;
3696 
3697     // Legalize the type.
3698     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Val);
3699 
3700     // This type is legalized to a scalar type.
3701     if (!LT.second.isVector())
3702       return 0;
3703 
3704     // The type may be split. Normalize the index to the new type.
3705     unsigned SizeInBits = LT.second.getSizeInBits();
3706     unsigned NumElts = LT.second.getVectorNumElements();
3707     unsigned SubNumElts = NumElts;
3708     Index = Index % NumElts;
3709 
3710     // For >128-bit vectors, we need to extract higher 128-bit subvectors.
3711     // For inserts, we also need to insert the subvector back.
3712     if (SizeInBits > 128) {
3713       assert((SizeInBits % 128) == 0 && "Illegal vector");
3714       unsigned NumSubVecs = SizeInBits / 128;
3715       SubNumElts = NumElts / NumSubVecs;
3716       if (SubNumElts <= Index) {
3717         RegisterFileMoveCost += (Opcode == Instruction::InsertElement ? 2 : 1);
3718         Index %= SubNumElts;
3719       }
3720     }
3721 
3722     if (Index == 0) {
3723       // Floating point scalars are already located in index #0.
3724       // Many insertions to #0 can fold away for scalar fp-ops, so let's assume
3725       // true for all.
3726       if (ScalarType->isFloatingPointTy())
3727         return RegisterFileMoveCost;
3728 
3729       // Assume movd/movq XMM -> GPR is relatively cheap on all targets.
3730       if (ScalarType->isIntegerTy() && Opcode == Instruction::ExtractElement)
3731         return 1 + RegisterFileMoveCost;
3732     }
3733 
3734     int ISD = TLI->InstructionOpcodeToISD(Opcode);
3735     assert(ISD && "Unexpected vector opcode");
3736     MVT MScalarTy = LT.second.getScalarType();
3737     if (ST->useSLMArithCosts())
3738       if (auto *Entry = CostTableLookup(SLMCostTbl, ISD, MScalarTy))
3739         return Entry->Cost + RegisterFileMoveCost;
3740 
3741     // Assume pinsr/pextr XMM <-> GPR is relatively cheap on all targets.
3742     if ((MScalarTy == MVT::i16 && ST->hasSSE2()) ||
3743         (MScalarTy.isInteger() && ST->hasSSE41()))
3744       return 1 + RegisterFileMoveCost;
3745 
3746     // Assume insertps is relatively cheap on all targets.
3747     if (MScalarTy == MVT::f32 && ST->hasSSE41() &&
3748         Opcode == Instruction::InsertElement)
3749       return 1 + RegisterFileMoveCost;
3750 
3751     // For extractions we just need to shuffle the element to index 0, which
3752     // should be very cheap (assume cost = 1). For insertions we need to shuffle
3753     // the elements to its destination. In both cases we must handle the
3754     // subvector move(s).
3755     // If the vector type is already less than 128-bits then don't reduce it.
3756     // TODO: Under what circumstances should we shuffle using the full width?
3757     InstructionCost ShuffleCost = 1;
3758     if (Opcode == Instruction::InsertElement) {
3759       auto *SubTy = cast<VectorType>(Val);
3760       EVT VT = TLI->getValueType(DL, Val);
3761       if (VT.getScalarType() != MScalarTy || VT.getSizeInBits() >= 128)
3762         SubTy = FixedVectorType::get(ScalarType, SubNumElts);
3763       ShuffleCost =
3764           getShuffleCost(TTI::SK_PermuteTwoSrc, SubTy, None, 0, SubTy);
3765     }
3766     int IntOrFpCost = ScalarType->isFloatingPointTy() ? 0 : 1;
3767     return ShuffleCost + IntOrFpCost + RegisterFileMoveCost;
3768   }
3769 
3770   // Add to the base cost if we know that the extracted element of a vector is
3771   // destined to be moved to and used in the integer register file.
3772   if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy())
3773     RegisterFileMoveCost += 1;
3774 
3775   return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost;
3776 }
3777 
3778 InstructionCost X86TTIImpl::getScalarizationOverhead(VectorType *Ty,
3779                                                      const APInt &DemandedElts,
3780                                                      bool Insert,
3781                                                      bool Extract) {
3782   assert(DemandedElts.getBitWidth() ==
3783              cast<FixedVectorType>(Ty)->getNumElements() &&
3784          "Vector size mismatch");
3785 
3786   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
3787   MVT MScalarTy = LT.second.getScalarType();
3788   unsigned SizeInBits = LT.second.getSizeInBits();
3789 
3790   InstructionCost Cost = 0;
3791 
3792   // For insertions, a ISD::BUILD_VECTOR style vector initialization can be much
3793   // cheaper than an accumulation of ISD::INSERT_VECTOR_ELT.
3794   if (Insert) {
3795     if ((MScalarTy == MVT::i16 && ST->hasSSE2()) ||
3796         (MScalarTy.isInteger() && ST->hasSSE41()) ||
3797         (MScalarTy == MVT::f32 && ST->hasSSE41())) {
3798       // For types we can insert directly, insertion into 128-bit sub vectors is
3799       // cheap, followed by a cheap chain of concatenations.
3800       if (SizeInBits <= 128) {
3801         Cost +=
3802             BaseT::getScalarizationOverhead(Ty, DemandedElts, Insert, false);
3803       } else {
3804         // In each 128-lane, if at least one index is demanded but not all
3805         // indices are demanded and this 128-lane is not the first 128-lane of
3806         // the legalized-vector, then this 128-lane needs a extracti128; If in
3807         // each 128-lane, there is at least one demanded index, this 128-lane
3808         // needs a inserti128.
3809 
3810         // The following cases will help you build a better understanding:
3811         // Assume we insert several elements into a v8i32 vector in avx2,
3812         // Case#1: inserting into 1th index needs vpinsrd + inserti128.
3813         // Case#2: inserting into 5th index needs extracti128 + vpinsrd +
3814         // inserti128.
3815         // Case#3: inserting into 4,5,6,7 index needs 4*vpinsrd + inserti128.
3816         const int CostValue = *LT.first.getValue();
3817         assert(CostValue >= 0 && "Negative cost!");
3818         unsigned Num128Lanes = SizeInBits / 128 * CostValue;
3819         unsigned NumElts = LT.second.getVectorNumElements() * CostValue;
3820         APInt WidenedDemandedElts = DemandedElts.zext(NumElts);
3821         unsigned Scale = NumElts / Num128Lanes;
3822         // We iterate each 128-lane, and check if we need a
3823         // extracti128/inserti128 for this 128-lane.
3824         for (unsigned I = 0; I < NumElts; I += Scale) {
3825           APInt Mask = WidenedDemandedElts.getBitsSet(NumElts, I, I + Scale);
3826           APInt MaskedDE = Mask & WidenedDemandedElts;
3827           unsigned Population = MaskedDE.countPopulation();
3828           Cost += (Population > 0 && Population != Scale &&
3829                    I % LT.second.getVectorNumElements() != 0);
3830           Cost += Population > 0;
3831         }
3832         Cost += DemandedElts.countPopulation();
3833 
3834         // For vXf32 cases, insertion into the 0'th index in each v4f32
3835         // 128-bit vector is free.
3836         // NOTE: This assumes legalization widens vXf32 vectors.
3837         if (MScalarTy == MVT::f32)
3838           for (unsigned i = 0, e = cast<FixedVectorType>(Ty)->getNumElements();
3839                i < e; i += 4)
3840             if (DemandedElts[i])
3841               Cost--;
3842       }
3843     } else if (LT.second.isVector()) {
3844       // Without fast insertion, we need to use MOVD/MOVQ to pass each demanded
3845       // integer element as a SCALAR_TO_VECTOR, then we build the vector as a
3846       // series of UNPCK followed by CONCAT_VECTORS - all of these can be
3847       // considered cheap.
3848       if (Ty->isIntOrIntVectorTy())
3849         Cost += DemandedElts.countPopulation();
3850 
3851       // Get the smaller of the legalized or original pow2-extended number of
3852       // vector elements, which represents the number of unpacks we'll end up
3853       // performing.
3854       unsigned NumElts = LT.second.getVectorNumElements();
3855       unsigned Pow2Elts =
3856           PowerOf2Ceil(cast<FixedVectorType>(Ty)->getNumElements());
3857       Cost += (std::min<unsigned>(NumElts, Pow2Elts) - 1) * LT.first;
3858     }
3859   }
3860 
3861   if (Extract) {
3862     // vXi1 can be efficiently extracted with MOVMSK.
3863     // TODO: AVX512 predicate mask handling.
3864     // NOTE: This doesn't work well for roundtrip scalarization.
3865     if (!Insert && Ty->getScalarSizeInBits() == 1 && !ST->hasAVX512()) {
3866       unsigned NumElts = cast<FixedVectorType>(Ty)->getNumElements();
3867       unsigned MaxElts = ST->hasAVX2() ? 32 : 16;
3868       unsigned MOVMSKCost = (NumElts + MaxElts - 1) / MaxElts;
3869       return MOVMSKCost;
3870     }
3871 
3872     if (LT.second.isVector()) {
3873       int CostValue = *LT.first.getValue();
3874       assert(CostValue >= 0 && "Negative cost!");
3875 
3876       unsigned NumElts = LT.second.getVectorNumElements() * CostValue;
3877       assert(NumElts >= DemandedElts.getBitWidth() &&
3878              "Vector has been legalized to smaller element count");
3879 
3880       // If we're extracting elements from a 128-bit subvector lane, we only need
3881       // to extract each lane once, not for every element.
3882       if (SizeInBits > 128) {
3883         assert((SizeInBits % 128) == 0 && "Illegal vector");
3884         unsigned NumLegal128Lanes = SizeInBits / 128;
3885         unsigned Num128Lanes = NumLegal128Lanes * CostValue;
3886         APInt WidenedDemandedElts = DemandedElts.zext(NumElts);
3887         unsigned Scale = NumElts / Num128Lanes;
3888 
3889         // Add cost for each demanded 128-bit subvector extraction.
3890         // Luckily this is a lot easier than for insertion.
3891         APInt DemandedUpper128Lanes =
3892             APIntOps::ScaleBitMask(WidenedDemandedElts, Num128Lanes);
3893         auto *Ty128 = FixedVectorType::get(Ty->getElementType(), Scale);
3894         for (unsigned I = 0; I != Num128Lanes; ++I)
3895           if (DemandedUpper128Lanes[I])
3896             Cost += getShuffleCost(TTI::SK_ExtractSubvector, Ty, None,
3897                                    I * Scale, Ty128);
3898 
3899         // Add all the demanded element extractions together, but adjust the
3900         // index to use the equivalent of the bottom 128 bit lane.
3901         for (unsigned I = 0; I != NumElts; ++I)
3902           if (WidenedDemandedElts[I]) {
3903             unsigned Idx = I % Scale;
3904             Cost += getVectorInstrCost(Instruction::ExtractElement, Ty, Idx);
3905           }
3906 
3907         return Cost;
3908       }
3909     }
3910 
3911     // Fallback to default extraction.
3912     Cost += BaseT::getScalarizationOverhead(Ty, DemandedElts, false, Extract);
3913   }
3914 
3915   return Cost;
3916 }
3917 
3918 InstructionCost
3919 X86TTIImpl::getReplicationShuffleCost(Type *EltTy, int ReplicationFactor,
3920                                       int VF, const APInt &DemandedDstElts,
3921                                       TTI::TargetCostKind CostKind) {
3922   const unsigned EltTyBits = DL.getTypeSizeInBits(EltTy);
3923   // We don't differentiate element types here, only element bit width.
3924   EltTy = IntegerType::getIntNTy(EltTy->getContext(), EltTyBits);
3925 
3926   auto bailout = [&]() {
3927     return BaseT::getReplicationShuffleCost(EltTy, ReplicationFactor, VF,
3928                                             DemandedDstElts, CostKind);
3929   };
3930 
3931   // For now, only deal with AVX512 cases.
3932   if (!ST->hasAVX512())
3933     return bailout();
3934 
3935   // Do we have a native shuffle for this element type, or should we promote?
3936   unsigned PromEltTyBits = EltTyBits;
3937   switch (EltTyBits) {
3938   case 32:
3939   case 64:
3940     break; // AVX512F.
3941   case 16:
3942     if (!ST->hasBWI())
3943       PromEltTyBits = 32; // promote to i32, AVX512F.
3944     break;                // AVX512BW
3945   case 8:
3946     if (!ST->hasVBMI())
3947       PromEltTyBits = 32; // promote to i32, AVX512F.
3948     break;                // AVX512VBMI
3949   case 1:
3950     // There is no support for shuffling i1 elements. We *must* promote.
3951     if (ST->hasBWI()) {
3952       if (ST->hasVBMI())
3953         PromEltTyBits = 8; // promote to i8, AVX512VBMI.
3954       else
3955         PromEltTyBits = 16; // promote to i16, AVX512BW.
3956       break;
3957     }
3958     if (ST->hasDQI()) {
3959       PromEltTyBits = 32; // promote to i32, AVX512F.
3960       break;
3961     }
3962     return bailout();
3963   default:
3964     return bailout();
3965   }
3966   auto *PromEltTy = IntegerType::getIntNTy(EltTy->getContext(), PromEltTyBits);
3967 
3968   auto *SrcVecTy = FixedVectorType::get(EltTy, VF);
3969   auto *PromSrcVecTy = FixedVectorType::get(PromEltTy, VF);
3970 
3971   int NumDstElements = VF * ReplicationFactor;
3972   auto *PromDstVecTy = FixedVectorType::get(PromEltTy, NumDstElements);
3973   auto *DstVecTy = FixedVectorType::get(EltTy, NumDstElements);
3974 
3975   // Legalize the types.
3976   MVT LegalSrcVecTy = TLI->getTypeLegalizationCost(DL, SrcVecTy).second;
3977   MVT LegalPromSrcVecTy = TLI->getTypeLegalizationCost(DL, PromSrcVecTy).second;
3978   MVT LegalPromDstVecTy = TLI->getTypeLegalizationCost(DL, PromDstVecTy).second;
3979   MVT LegalDstVecTy = TLI->getTypeLegalizationCost(DL, DstVecTy).second;
3980   // They should have legalized into vector types.
3981   if (!LegalSrcVecTy.isVector() || !LegalPromSrcVecTy.isVector() ||
3982       !LegalPromDstVecTy.isVector() || !LegalDstVecTy.isVector())
3983     return bailout();
3984 
3985   if (PromEltTyBits != EltTyBits) {
3986     // If we have to perform the shuffle with wider elt type than our data type,
3987     // then we will first need to anyext (we don't care about the new bits)
3988     // the source elements, and then truncate Dst elements.
3989     InstructionCost PromotionCost;
3990     PromotionCost += getCastInstrCost(
3991         Instruction::SExt, /*Dst=*/PromSrcVecTy, /*Src=*/SrcVecTy,
3992         TargetTransformInfo::CastContextHint::None, CostKind);
3993     PromotionCost +=
3994         getCastInstrCost(Instruction::Trunc, /*Dst=*/DstVecTy,
3995                          /*Src=*/PromDstVecTy,
3996                          TargetTransformInfo::CastContextHint::None, CostKind);
3997     return PromotionCost + getReplicationShuffleCost(PromEltTy,
3998                                                      ReplicationFactor, VF,
3999                                                      DemandedDstElts, CostKind);
4000   }
4001 
4002   assert(LegalSrcVecTy.getScalarSizeInBits() == EltTyBits &&
4003          LegalSrcVecTy.getScalarType() == LegalDstVecTy.getScalarType() &&
4004          "We expect that the legalization doesn't affect the element width, "
4005          "doesn't coalesce/split elements.");
4006 
4007   unsigned NumEltsPerDstVec = LegalDstVecTy.getVectorNumElements();
4008   unsigned NumDstVectors =
4009       divideCeil(DstVecTy->getNumElements(), NumEltsPerDstVec);
4010 
4011   auto *SingleDstVecTy = FixedVectorType::get(EltTy, NumEltsPerDstVec);
4012 
4013   // Not all the produced Dst elements may be demanded. In our case,
4014   // given that a single Dst vector is formed by a single shuffle,
4015   // if all elements that will form a single Dst vector aren't demanded,
4016   // then we won't need to do that shuffle, so adjust the cost accordingly.
4017   APInt DemandedDstVectors = APIntOps::ScaleBitMask(
4018       DemandedDstElts.zext(NumDstVectors * NumEltsPerDstVec), NumDstVectors);
4019   unsigned NumDstVectorsDemanded = DemandedDstVectors.countPopulation();
4020 
4021   InstructionCost SingleShuffleCost =
4022       getShuffleCost(TTI::SK_PermuteSingleSrc, SingleDstVecTy,
4023                      /*Mask=*/None, /*Index=*/0, /*SubTp=*/nullptr);
4024   return NumDstVectorsDemanded * SingleShuffleCost;
4025 }
4026 
4027 InstructionCost X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
4028                                             MaybeAlign Alignment,
4029                                             unsigned AddressSpace,
4030                                             TTI::TargetCostKind CostKind,
4031                                             const Instruction *I) {
4032   // TODO: Handle other cost kinds.
4033   if (CostKind != TTI::TCK_RecipThroughput) {
4034     if (auto *SI = dyn_cast_or_null<StoreInst>(I)) {
4035       // Store instruction with index and scale costs 2 Uops.
4036       // Check the preceding GEP to identify non-const indices.
4037       if (auto *GEP = dyn_cast<GetElementPtrInst>(SI->getPointerOperand())) {
4038         if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); }))
4039           return TTI::TCC_Basic * 2;
4040       }
4041     }
4042     return TTI::TCC_Basic;
4043   }
4044 
4045   assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
4046          "Invalid Opcode");
4047   // Type legalization can't handle structs
4048   if (TLI->getValueType(DL, Src, true) == MVT::Other)
4049     return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
4050                                   CostKind);
4051 
4052   // Legalize the type.
4053   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
4054 
4055   auto *VTy = dyn_cast<FixedVectorType>(Src);
4056 
4057   // Handle the simple case of non-vectors.
4058   // NOTE: this assumes that legalization never creates vector from scalars!
4059   if (!VTy || !LT.second.isVector())
4060     // Each load/store unit costs 1.
4061     return LT.first * 1;
4062 
4063   bool IsLoad = Opcode == Instruction::Load;
4064 
4065   Type *EltTy = VTy->getElementType();
4066 
4067   const int EltTyBits = DL.getTypeSizeInBits(EltTy);
4068 
4069   InstructionCost Cost = 0;
4070 
4071   // Source of truth: how many elements were there in the original IR vector?
4072   const unsigned SrcNumElt = VTy->getNumElements();
4073 
4074   // How far have we gotten?
4075   int NumEltRemaining = SrcNumElt;
4076   // Note that we intentionally capture by-reference, NumEltRemaining changes.
4077   auto NumEltDone = [&]() { return SrcNumElt - NumEltRemaining; };
4078 
4079   const int MaxLegalOpSizeBytes = divideCeil(LT.second.getSizeInBits(), 8);
4080 
4081   // Note that even if we can store 64 bits of an XMM, we still operate on XMM.
4082   const unsigned XMMBits = 128;
4083   if (XMMBits % EltTyBits != 0)
4084     // Vector size must be a multiple of the element size. I.e. no padding.
4085     return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
4086                                   CostKind);
4087   const int NumEltPerXMM = XMMBits / EltTyBits;
4088 
4089   auto *XMMVecTy = FixedVectorType::get(EltTy, NumEltPerXMM);
4090 
4091   for (int CurrOpSizeBytes = MaxLegalOpSizeBytes, SubVecEltsLeft = 0;
4092        NumEltRemaining > 0; CurrOpSizeBytes /= 2) {
4093     // How many elements would a single op deal with at once?
4094     if ((8 * CurrOpSizeBytes) % EltTyBits != 0)
4095       // Vector size must be a multiple of the element size. I.e. no padding.
4096       return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
4097                                     CostKind);
4098     int CurrNumEltPerOp = (8 * CurrOpSizeBytes) / EltTyBits;
4099 
4100     assert(CurrOpSizeBytes > 0 && CurrNumEltPerOp > 0 && "How'd we get here?");
4101     assert((((NumEltRemaining * EltTyBits) < (2 * 8 * CurrOpSizeBytes)) ||
4102             (CurrOpSizeBytes == MaxLegalOpSizeBytes)) &&
4103            "Unless we haven't halved the op size yet, "
4104            "we have less than two op's sized units of work left.");
4105 
4106     auto *CurrVecTy = CurrNumEltPerOp > NumEltPerXMM
4107                           ? FixedVectorType::get(EltTy, CurrNumEltPerOp)
4108                           : XMMVecTy;
4109 
4110     assert(CurrVecTy->getNumElements() % CurrNumEltPerOp == 0 &&
4111            "After halving sizes, the vector elt count is no longer a multiple "
4112            "of number of elements per operation?");
4113     auto *CoalescedVecTy =
4114         CurrNumEltPerOp == 1
4115             ? CurrVecTy
4116             : FixedVectorType::get(
4117                   IntegerType::get(Src->getContext(),
4118                                    EltTyBits * CurrNumEltPerOp),
4119                   CurrVecTy->getNumElements() / CurrNumEltPerOp);
4120     assert(DL.getTypeSizeInBits(CoalescedVecTy) ==
4121                DL.getTypeSizeInBits(CurrVecTy) &&
4122            "coalesciing elements doesn't change vector width.");
4123 
4124     while (NumEltRemaining > 0) {
4125       assert(SubVecEltsLeft >= 0 && "Subreg element count overconsumtion?");
4126 
4127       // Can we use this vector size, as per the remaining element count?
4128       // Iff the vector is naturally aligned, we can do a wide load regardless.
4129       if (NumEltRemaining < CurrNumEltPerOp &&
4130           (!IsLoad || Alignment.valueOrOne() < CurrOpSizeBytes) &&
4131           CurrOpSizeBytes != 1)
4132         break; // Try smalled vector size.
4133 
4134       bool Is0thSubVec = (NumEltDone() % LT.second.getVectorNumElements()) == 0;
4135 
4136       // If we have fully processed the previous reg, we need to replenish it.
4137       if (SubVecEltsLeft == 0) {
4138         SubVecEltsLeft += CurrVecTy->getNumElements();
4139         // And that's free only for the 0'th subvector of a legalized vector.
4140         if (!Is0thSubVec)
4141           Cost += getShuffleCost(IsLoad ? TTI::ShuffleKind::SK_InsertSubvector
4142                                         : TTI::ShuffleKind::SK_ExtractSubvector,
4143                                  VTy, None, NumEltDone(), CurrVecTy);
4144       }
4145 
4146       // While we can directly load/store ZMM, YMM, and 64-bit halves of XMM,
4147       // for smaller widths (32/16/8) we have to insert/extract them separately.
4148       // Again, it's free for the 0'th subreg (if op is 32/64 bit wide,
4149       // but let's pretend that it is also true for 16/8 bit wide ops...)
4150       if (CurrOpSizeBytes <= 32 / 8 && !Is0thSubVec) {
4151         int NumEltDoneInCurrXMM = NumEltDone() % NumEltPerXMM;
4152         assert(NumEltDoneInCurrXMM % CurrNumEltPerOp == 0 && "");
4153         int CoalescedVecEltIdx = NumEltDoneInCurrXMM / CurrNumEltPerOp;
4154         APInt DemandedElts =
4155             APInt::getBitsSet(CoalescedVecTy->getNumElements(),
4156                               CoalescedVecEltIdx, CoalescedVecEltIdx + 1);
4157         assert(DemandedElts.countPopulation() == 1 && "Inserting single value");
4158         Cost += getScalarizationOverhead(CoalescedVecTy, DemandedElts, IsLoad,
4159                                          !IsLoad);
4160       }
4161 
4162       // This isn't exactly right. We're using slow unaligned 32-byte accesses
4163       // as a proxy for a double-pumped AVX memory interface such as on
4164       // Sandybridge.
4165       if (CurrOpSizeBytes == 32 && ST->isUnalignedMem32Slow())
4166         Cost += 2;
4167       else
4168         Cost += 1;
4169 
4170       SubVecEltsLeft -= CurrNumEltPerOp;
4171       NumEltRemaining -= CurrNumEltPerOp;
4172       Alignment = commonAlignment(Alignment.valueOrOne(), CurrOpSizeBytes);
4173     }
4174   }
4175 
4176   assert(NumEltRemaining <= 0 && "Should have processed all the elements.");
4177 
4178   return Cost;
4179 }
4180 
4181 InstructionCost
4182 X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy, Align Alignment,
4183                                   unsigned AddressSpace,
4184                                   TTI::TargetCostKind CostKind) {
4185   bool IsLoad = (Instruction::Load == Opcode);
4186   bool IsStore = (Instruction::Store == Opcode);
4187 
4188   auto *SrcVTy = dyn_cast<FixedVectorType>(SrcTy);
4189   if (!SrcVTy)
4190     // To calculate scalar take the regular cost, without mask
4191     return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace, CostKind);
4192 
4193   unsigned NumElem = SrcVTy->getNumElements();
4194   auto *MaskTy =
4195       FixedVectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem);
4196   if ((IsLoad && !isLegalMaskedLoad(SrcVTy, Alignment)) ||
4197       (IsStore && !isLegalMaskedStore(SrcVTy, Alignment))) {
4198     // Scalarization
4199     APInt DemandedElts = APInt::getAllOnes(NumElem);
4200     InstructionCost MaskSplitCost =
4201         getScalarizationOverhead(MaskTy, DemandedElts, false, true);
4202     InstructionCost ScalarCompareCost = getCmpSelInstrCost(
4203         Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr,
4204         CmpInst::BAD_ICMP_PREDICATE, CostKind);
4205     InstructionCost BranchCost = getCFInstrCost(Instruction::Br, CostKind);
4206     InstructionCost MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost);
4207     InstructionCost ValueSplitCost =
4208         getScalarizationOverhead(SrcVTy, DemandedElts, IsLoad, IsStore);
4209     InstructionCost MemopCost =
4210         NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
4211                                          Alignment, AddressSpace, CostKind);
4212     return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost;
4213   }
4214 
4215   // Legalize the type.
4216   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy);
4217   auto VT = TLI->getValueType(DL, SrcVTy);
4218   InstructionCost Cost = 0;
4219   if (VT.isSimple() && LT.second != VT.getSimpleVT() &&
4220       LT.second.getVectorNumElements() == NumElem)
4221     // Promotion requires extend/truncate for data and a shuffle for mask.
4222     Cost += getShuffleCost(TTI::SK_PermuteTwoSrc, SrcVTy, None, 0, nullptr) +
4223             getShuffleCost(TTI::SK_PermuteTwoSrc, MaskTy, None, 0, nullptr);
4224 
4225   else if (LT.first * LT.second.getVectorNumElements() > NumElem) {
4226     auto *NewMaskTy = FixedVectorType::get(MaskTy->getElementType(),
4227                                            LT.second.getVectorNumElements());
4228     // Expanding requires fill mask with zeroes
4229     Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, None, 0, MaskTy);
4230   }
4231 
4232   // Pre-AVX512 - each maskmov load costs 2 + store costs ~8.
4233   if (!ST->hasAVX512())
4234     return Cost + LT.first * (IsLoad ? 2 : 8);
4235 
4236   // AVX-512 masked load/store is cheapper
4237   return Cost + LT.first;
4238 }
4239 
4240 InstructionCost X86TTIImpl::getAddressComputationCost(Type *Ty,
4241                                                       ScalarEvolution *SE,
4242                                                       const SCEV *Ptr) {
4243   // Address computations in vectorized code with non-consecutive addresses will
4244   // likely result in more instructions compared to scalar code where the
4245   // computation can more often be merged into the index mode. The resulting
4246   // extra micro-ops can significantly decrease throughput.
4247   const unsigned NumVectorInstToHideOverhead = 10;
4248 
4249   // Cost modeling of Strided Access Computation is hidden by the indexing
4250   // modes of X86 regardless of the stride value. We dont believe that there
4251   // is a difference between constant strided access in gerenal and constant
4252   // strided value which is less than or equal to 64.
4253   // Even in the case of (loop invariant) stride whose value is not known at
4254   // compile time, the address computation will not incur more than one extra
4255   // ADD instruction.
4256   if (Ty->isVectorTy() && SE && !ST->hasAVX2()) {
4257     // TODO: AVX2 is the current cut-off because we don't have correct
4258     //       interleaving costs for prior ISA's.
4259     if (!BaseT::isStridedAccess(Ptr))
4260       return NumVectorInstToHideOverhead;
4261     if (!BaseT::getConstantStrideStep(SE, Ptr))
4262       return 1;
4263   }
4264 
4265   return BaseT::getAddressComputationCost(Ty, SE, Ptr);
4266 }
4267 
4268 InstructionCost
4269 X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy,
4270                                        Optional<FastMathFlags> FMF,
4271                                        TTI::TargetCostKind CostKind) {
4272   if (TTI::requiresOrderedReduction(FMF))
4273     return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
4274 
4275   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
4276   // and make it as the cost.
4277 
4278   static const CostTblEntry SLMCostTblNoPairWise[] = {
4279     { ISD::FADD,  MVT::v2f64,   3 },
4280     { ISD::ADD,   MVT::v2i64,   5 },
4281   };
4282 
4283   static const CostTblEntry SSE2CostTblNoPairWise[] = {
4284     { ISD::FADD,  MVT::v2f64,   2 },
4285     { ISD::FADD,  MVT::v2f32,   2 },
4286     { ISD::FADD,  MVT::v4f32,   4 },
4287     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
4288     { ISD::ADD,   MVT::v2i32,   2 }, // FIXME: chosen to be less than v4i32
4289     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.3".
4290     { ISD::ADD,   MVT::v2i16,   2 },      // The data reported by the IACA tool is "4.3".
4291     { ISD::ADD,   MVT::v4i16,   3 },      // The data reported by the IACA tool is "4.3".
4292     { ISD::ADD,   MVT::v8i16,   4 },      // The data reported by the IACA tool is "4.3".
4293     { ISD::ADD,   MVT::v2i8,    2 },
4294     { ISD::ADD,   MVT::v4i8,    2 },
4295     { ISD::ADD,   MVT::v8i8,    2 },
4296     { ISD::ADD,   MVT::v16i8,   3 },
4297   };
4298 
4299   static const CostTblEntry AVX1CostTblNoPairWise[] = {
4300     { ISD::FADD,  MVT::v4f64,   3 },
4301     { ISD::FADD,  MVT::v4f32,   3 },
4302     { ISD::FADD,  MVT::v8f32,   4 },
4303     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
4304     { ISD::ADD,   MVT::v4i64,   3 },
4305     { ISD::ADD,   MVT::v8i32,   5 },
4306     { ISD::ADD,   MVT::v16i16,  5 },
4307     { ISD::ADD,   MVT::v32i8,   4 },
4308   };
4309 
4310   int ISD = TLI->InstructionOpcodeToISD(Opcode);
4311   assert(ISD && "Invalid opcode");
4312 
4313   // Before legalizing the type, give a chance to look up illegal narrow types
4314   // in the table.
4315   // FIXME: Is there a better way to do this?
4316   EVT VT = TLI->getValueType(DL, ValTy);
4317   if (VT.isSimple()) {
4318     MVT MTy = VT.getSimpleVT();
4319     if (ST->useSLMArithCosts())
4320       if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy))
4321         return Entry->Cost;
4322 
4323     if (ST->hasAVX())
4324       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
4325         return Entry->Cost;
4326 
4327     if (ST->hasSSE2())
4328       if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
4329         return Entry->Cost;
4330   }
4331 
4332   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
4333 
4334   MVT MTy = LT.second;
4335 
4336   auto *ValVTy = cast<FixedVectorType>(ValTy);
4337 
4338   // Special case: vXi8 mul reductions are performed as vXi16.
4339   if (ISD == ISD::MUL && MTy.getScalarType() == MVT::i8) {
4340     auto *WideSclTy = IntegerType::get(ValVTy->getContext(), 16);
4341     auto *WideVecTy = FixedVectorType::get(WideSclTy, ValVTy->getNumElements());
4342     return getCastInstrCost(Instruction::ZExt, WideVecTy, ValTy,
4343                             TargetTransformInfo::CastContextHint::None,
4344                             CostKind) +
4345            getArithmeticReductionCost(Opcode, WideVecTy, FMF, CostKind);
4346   }
4347 
4348   InstructionCost ArithmeticCost = 0;
4349   if (LT.first != 1 && MTy.isVector() &&
4350       MTy.getVectorNumElements() < ValVTy->getNumElements()) {
4351     // Type needs to be split. We need LT.first - 1 arithmetic ops.
4352     auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(),
4353                                             MTy.getVectorNumElements());
4354     ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind);
4355     ArithmeticCost *= LT.first - 1;
4356   }
4357 
4358   if (ST->useSLMArithCosts())
4359     if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy))
4360       return ArithmeticCost + Entry->Cost;
4361 
4362   if (ST->hasAVX())
4363     if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
4364       return ArithmeticCost + Entry->Cost;
4365 
4366   if (ST->hasSSE2())
4367     if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
4368       return ArithmeticCost + Entry->Cost;
4369 
4370   // FIXME: These assume a naive kshift+binop lowering, which is probably
4371   // conservative in most cases.
4372   static const CostTblEntry AVX512BoolReduction[] = {
4373     { ISD::AND,  MVT::v2i1,   3 },
4374     { ISD::AND,  MVT::v4i1,   5 },
4375     { ISD::AND,  MVT::v8i1,   7 },
4376     { ISD::AND,  MVT::v16i1,  9 },
4377     { ISD::AND,  MVT::v32i1, 11 },
4378     { ISD::AND,  MVT::v64i1, 13 },
4379     { ISD::OR,   MVT::v2i1,   3 },
4380     { ISD::OR,   MVT::v4i1,   5 },
4381     { ISD::OR,   MVT::v8i1,   7 },
4382     { ISD::OR,   MVT::v16i1,  9 },
4383     { ISD::OR,   MVT::v32i1, 11 },
4384     { ISD::OR,   MVT::v64i1, 13 },
4385   };
4386 
4387   static const CostTblEntry AVX2BoolReduction[] = {
4388     { ISD::AND,  MVT::v16i16,  2 }, // vpmovmskb + cmp
4389     { ISD::AND,  MVT::v32i8,   2 }, // vpmovmskb + cmp
4390     { ISD::OR,   MVT::v16i16,  2 }, // vpmovmskb + cmp
4391     { ISD::OR,   MVT::v32i8,   2 }, // vpmovmskb + cmp
4392   };
4393 
4394   static const CostTblEntry AVX1BoolReduction[] = {
4395     { ISD::AND,  MVT::v4i64,   2 }, // vmovmskpd + cmp
4396     { ISD::AND,  MVT::v8i32,   2 }, // vmovmskps + cmp
4397     { ISD::AND,  MVT::v16i16,  4 }, // vextractf128 + vpand + vpmovmskb + cmp
4398     { ISD::AND,  MVT::v32i8,   4 }, // vextractf128 + vpand + vpmovmskb + cmp
4399     { ISD::OR,   MVT::v4i64,   2 }, // vmovmskpd + cmp
4400     { ISD::OR,   MVT::v8i32,   2 }, // vmovmskps + cmp
4401     { ISD::OR,   MVT::v16i16,  4 }, // vextractf128 + vpor + vpmovmskb + cmp
4402     { ISD::OR,   MVT::v32i8,   4 }, // vextractf128 + vpor + vpmovmskb + cmp
4403   };
4404 
4405   static const CostTblEntry SSE2BoolReduction[] = {
4406     { ISD::AND,  MVT::v2i64,   2 }, // movmskpd + cmp
4407     { ISD::AND,  MVT::v4i32,   2 }, // movmskps + cmp
4408     { ISD::AND,  MVT::v8i16,   2 }, // pmovmskb + cmp
4409     { ISD::AND,  MVT::v16i8,   2 }, // pmovmskb + cmp
4410     { ISD::OR,   MVT::v2i64,   2 }, // movmskpd + cmp
4411     { ISD::OR,   MVT::v4i32,   2 }, // movmskps + cmp
4412     { ISD::OR,   MVT::v8i16,   2 }, // pmovmskb + cmp
4413     { ISD::OR,   MVT::v16i8,   2 }, // pmovmskb + cmp
4414   };
4415 
4416   // Handle bool allof/anyof patterns.
4417   if (ValVTy->getElementType()->isIntegerTy(1)) {
4418     InstructionCost ArithmeticCost = 0;
4419     if (LT.first != 1 && MTy.isVector() &&
4420         MTy.getVectorNumElements() < ValVTy->getNumElements()) {
4421       // Type needs to be split. We need LT.first - 1 arithmetic ops.
4422       auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(),
4423                                               MTy.getVectorNumElements());
4424       ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind);
4425       ArithmeticCost *= LT.first - 1;
4426     }
4427 
4428     if (ST->hasAVX512())
4429       if (const auto *Entry = CostTableLookup(AVX512BoolReduction, ISD, MTy))
4430         return ArithmeticCost + Entry->Cost;
4431     if (ST->hasAVX2())
4432       if (const auto *Entry = CostTableLookup(AVX2BoolReduction, ISD, MTy))
4433         return ArithmeticCost + Entry->Cost;
4434     if (ST->hasAVX())
4435       if (const auto *Entry = CostTableLookup(AVX1BoolReduction, ISD, MTy))
4436         return ArithmeticCost + Entry->Cost;
4437     if (ST->hasSSE2())
4438       if (const auto *Entry = CostTableLookup(SSE2BoolReduction, ISD, MTy))
4439         return ArithmeticCost + Entry->Cost;
4440 
4441     return BaseT::getArithmeticReductionCost(Opcode, ValVTy, FMF, CostKind);
4442   }
4443 
4444   unsigned NumVecElts = ValVTy->getNumElements();
4445   unsigned ScalarSize = ValVTy->getScalarSizeInBits();
4446 
4447   // Special case power of 2 reductions where the scalar type isn't changed
4448   // by type legalization.
4449   if (!isPowerOf2_32(NumVecElts) || ScalarSize != MTy.getScalarSizeInBits())
4450     return BaseT::getArithmeticReductionCost(Opcode, ValVTy, FMF, CostKind);
4451 
4452   InstructionCost ReductionCost = 0;
4453 
4454   auto *Ty = ValVTy;
4455   if (LT.first != 1 && MTy.isVector() &&
4456       MTy.getVectorNumElements() < ValVTy->getNumElements()) {
4457     // Type needs to be split. We need LT.first - 1 arithmetic ops.
4458     Ty = FixedVectorType::get(ValVTy->getElementType(),
4459                               MTy.getVectorNumElements());
4460     ReductionCost = getArithmeticInstrCost(Opcode, Ty, CostKind);
4461     ReductionCost *= LT.first - 1;
4462     NumVecElts = MTy.getVectorNumElements();
4463   }
4464 
4465   // Now handle reduction with the legal type, taking into account size changes
4466   // at each level.
4467   while (NumVecElts > 1) {
4468     // Determine the size of the remaining vector we need to reduce.
4469     unsigned Size = NumVecElts * ScalarSize;
4470     NumVecElts /= 2;
4471     // If we're reducing from 256/512 bits, use an extract_subvector.
4472     if (Size > 128) {
4473       auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts);
4474       ReductionCost +=
4475           getShuffleCost(TTI::SK_ExtractSubvector, Ty, None, NumVecElts, SubTy);
4476       Ty = SubTy;
4477     } else if (Size == 128) {
4478       // Reducing from 128 bits is a permute of v2f64/v2i64.
4479       FixedVectorType *ShufTy;
4480       if (ValVTy->isFloatingPointTy())
4481         ShufTy =
4482             FixedVectorType::get(Type::getDoubleTy(ValVTy->getContext()), 2);
4483       else
4484         ShufTy =
4485             FixedVectorType::get(Type::getInt64Ty(ValVTy->getContext()), 2);
4486       ReductionCost +=
4487           getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4488     } else if (Size == 64) {
4489       // Reducing from 64 bits is a shuffle of v4f32/v4i32.
4490       FixedVectorType *ShufTy;
4491       if (ValVTy->isFloatingPointTy())
4492         ShufTy =
4493             FixedVectorType::get(Type::getFloatTy(ValVTy->getContext()), 4);
4494       else
4495         ShufTy =
4496             FixedVectorType::get(Type::getInt32Ty(ValVTy->getContext()), 4);
4497       ReductionCost +=
4498           getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4499     } else {
4500       // Reducing from smaller size is a shift by immediate.
4501       auto *ShiftTy = FixedVectorType::get(
4502           Type::getIntNTy(ValVTy->getContext(), Size), 128 / Size);
4503       ReductionCost += getArithmeticInstrCost(
4504           Instruction::LShr, ShiftTy, CostKind,
4505           TargetTransformInfo::OK_AnyValue,
4506           TargetTransformInfo::OK_UniformConstantValue,
4507           TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
4508     }
4509 
4510     // Add the arithmetic op for this level.
4511     ReductionCost += getArithmeticInstrCost(Opcode, Ty, CostKind);
4512   }
4513 
4514   // Add the final extract element to the cost.
4515   return ReductionCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
4516 }
4517 
4518 InstructionCost X86TTIImpl::getMinMaxCost(Type *Ty, Type *CondTy,
4519                                           bool IsUnsigned) {
4520   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
4521 
4522   MVT MTy = LT.second;
4523 
4524   int ISD;
4525   if (Ty->isIntOrIntVectorTy()) {
4526     ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN;
4527   } else {
4528     assert(Ty->isFPOrFPVectorTy() &&
4529            "Expected float point or integer vector type.");
4530     ISD = ISD::FMINNUM;
4531   }
4532 
4533   static const CostTblEntry SSE1CostTbl[] = {
4534     {ISD::FMINNUM, MVT::v4f32, 1},
4535   };
4536 
4537   static const CostTblEntry SSE2CostTbl[] = {
4538     {ISD::FMINNUM, MVT::v2f64, 1},
4539     {ISD::SMIN,    MVT::v8i16, 1},
4540     {ISD::UMIN,    MVT::v16i8, 1},
4541   };
4542 
4543   static const CostTblEntry SSE41CostTbl[] = {
4544     {ISD::SMIN,    MVT::v4i32, 1},
4545     {ISD::UMIN,    MVT::v4i32, 1},
4546     {ISD::UMIN,    MVT::v8i16, 1},
4547     {ISD::SMIN,    MVT::v16i8, 1},
4548   };
4549 
4550   static const CostTblEntry SSE42CostTbl[] = {
4551     {ISD::UMIN,    MVT::v2i64, 3}, // xor+pcmpgtq+blendvpd
4552   };
4553 
4554   static const CostTblEntry AVX1CostTbl[] = {
4555     {ISD::FMINNUM, MVT::v8f32,  1},
4556     {ISD::FMINNUM, MVT::v4f64,  1},
4557     {ISD::SMIN,    MVT::v8i32,  3},
4558     {ISD::UMIN,    MVT::v8i32,  3},
4559     {ISD::SMIN,    MVT::v16i16, 3},
4560     {ISD::UMIN,    MVT::v16i16, 3},
4561     {ISD::SMIN,    MVT::v32i8,  3},
4562     {ISD::UMIN,    MVT::v32i8,  3},
4563   };
4564 
4565   static const CostTblEntry AVX2CostTbl[] = {
4566     {ISD::SMIN,    MVT::v8i32,  1},
4567     {ISD::UMIN,    MVT::v8i32,  1},
4568     {ISD::SMIN,    MVT::v16i16, 1},
4569     {ISD::UMIN,    MVT::v16i16, 1},
4570     {ISD::SMIN,    MVT::v32i8,  1},
4571     {ISD::UMIN,    MVT::v32i8,  1},
4572   };
4573 
4574   static const CostTblEntry AVX512CostTbl[] = {
4575     {ISD::FMINNUM, MVT::v16f32, 1},
4576     {ISD::FMINNUM, MVT::v8f64,  1},
4577     {ISD::SMIN,    MVT::v2i64,  1},
4578     {ISD::UMIN,    MVT::v2i64,  1},
4579     {ISD::SMIN,    MVT::v4i64,  1},
4580     {ISD::UMIN,    MVT::v4i64,  1},
4581     {ISD::SMIN,    MVT::v8i64,  1},
4582     {ISD::UMIN,    MVT::v8i64,  1},
4583     {ISD::SMIN,    MVT::v16i32, 1},
4584     {ISD::UMIN,    MVT::v16i32, 1},
4585   };
4586 
4587   static const CostTblEntry AVX512BWCostTbl[] = {
4588     {ISD::SMIN,    MVT::v32i16, 1},
4589     {ISD::UMIN,    MVT::v32i16, 1},
4590     {ISD::SMIN,    MVT::v64i8,  1},
4591     {ISD::UMIN,    MVT::v64i8,  1},
4592   };
4593 
4594   // If we have a native MIN/MAX instruction for this type, use it.
4595   if (ST->hasBWI())
4596     if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
4597       return LT.first * Entry->Cost;
4598 
4599   if (ST->hasAVX512())
4600     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
4601       return LT.first * Entry->Cost;
4602 
4603   if (ST->hasAVX2())
4604     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
4605       return LT.first * Entry->Cost;
4606 
4607   if (ST->hasAVX())
4608     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
4609       return LT.first * Entry->Cost;
4610 
4611   if (ST->hasSSE42())
4612     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
4613       return LT.first * Entry->Cost;
4614 
4615   if (ST->hasSSE41())
4616     if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
4617       return LT.first * Entry->Cost;
4618 
4619   if (ST->hasSSE2())
4620     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
4621       return LT.first * Entry->Cost;
4622 
4623   if (ST->hasSSE1())
4624     if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
4625       return LT.first * Entry->Cost;
4626 
4627   unsigned CmpOpcode;
4628   if (Ty->isFPOrFPVectorTy()) {
4629     CmpOpcode = Instruction::FCmp;
4630   } else {
4631     assert(Ty->isIntOrIntVectorTy() &&
4632            "expecting floating point or integer type for min/max reduction");
4633     CmpOpcode = Instruction::ICmp;
4634   }
4635 
4636   TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
4637   // Otherwise fall back to cmp+select.
4638   InstructionCost Result =
4639       getCmpSelInstrCost(CmpOpcode, Ty, CondTy, CmpInst::BAD_ICMP_PREDICATE,
4640                          CostKind) +
4641       getCmpSelInstrCost(Instruction::Select, Ty, CondTy,
4642                          CmpInst::BAD_ICMP_PREDICATE, CostKind);
4643   return Result;
4644 }
4645 
4646 InstructionCost
4647 X86TTIImpl::getMinMaxReductionCost(VectorType *ValTy, VectorType *CondTy,
4648                                    bool IsUnsigned,
4649                                    TTI::TargetCostKind CostKind) {
4650   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
4651 
4652   MVT MTy = LT.second;
4653 
4654   int ISD;
4655   if (ValTy->isIntOrIntVectorTy()) {
4656     ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN;
4657   } else {
4658     assert(ValTy->isFPOrFPVectorTy() &&
4659            "Expected float point or integer vector type.");
4660     ISD = ISD::FMINNUM;
4661   }
4662 
4663   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
4664   // and make it as the cost.
4665 
4666   static const CostTblEntry SSE2CostTblNoPairWise[] = {
4667       {ISD::UMIN, MVT::v2i16, 5}, // need pxors to use pminsw/pmaxsw
4668       {ISD::UMIN, MVT::v4i16, 7}, // need pxors to use pminsw/pmaxsw
4669       {ISD::UMIN, MVT::v8i16, 9}, // need pxors to use pminsw/pmaxsw
4670   };
4671 
4672   static const CostTblEntry SSE41CostTblNoPairWise[] = {
4673       {ISD::SMIN, MVT::v2i16, 3}, // same as sse2
4674       {ISD::SMIN, MVT::v4i16, 5}, // same as sse2
4675       {ISD::UMIN, MVT::v2i16, 5}, // same as sse2
4676       {ISD::UMIN, MVT::v4i16, 7}, // same as sse2
4677       {ISD::SMIN, MVT::v8i16, 4}, // phminposuw+xor
4678       {ISD::UMIN, MVT::v8i16, 4}, // FIXME: umin is cheaper than umax
4679       {ISD::SMIN, MVT::v2i8,  3}, // pminsb
4680       {ISD::SMIN, MVT::v4i8,  5}, // pminsb
4681       {ISD::SMIN, MVT::v8i8,  7}, // pminsb
4682       {ISD::SMIN, MVT::v16i8, 6},
4683       {ISD::UMIN, MVT::v2i8,  3}, // same as sse2
4684       {ISD::UMIN, MVT::v4i8,  5}, // same as sse2
4685       {ISD::UMIN, MVT::v8i8,  7}, // same as sse2
4686       {ISD::UMIN, MVT::v16i8, 6}, // FIXME: umin is cheaper than umax
4687   };
4688 
4689   static const CostTblEntry AVX1CostTblNoPairWise[] = {
4690       {ISD::SMIN, MVT::v16i16, 6},
4691       {ISD::UMIN, MVT::v16i16, 6}, // FIXME: umin is cheaper than umax
4692       {ISD::SMIN, MVT::v32i8, 8},
4693       {ISD::UMIN, MVT::v32i8, 8},
4694   };
4695 
4696   static const CostTblEntry AVX512BWCostTblNoPairWise[] = {
4697       {ISD::SMIN, MVT::v32i16, 8},
4698       {ISD::UMIN, MVT::v32i16, 8}, // FIXME: umin is cheaper than umax
4699       {ISD::SMIN, MVT::v64i8, 10},
4700       {ISD::UMIN, MVT::v64i8, 10},
4701   };
4702 
4703   // Before legalizing the type, give a chance to look up illegal narrow types
4704   // in the table.
4705   // FIXME: Is there a better way to do this?
4706   EVT VT = TLI->getValueType(DL, ValTy);
4707   if (VT.isSimple()) {
4708     MVT MTy = VT.getSimpleVT();
4709     if (ST->hasBWI())
4710       if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy))
4711         return Entry->Cost;
4712 
4713     if (ST->hasAVX())
4714       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
4715         return Entry->Cost;
4716 
4717     if (ST->hasSSE41())
4718       if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy))
4719         return Entry->Cost;
4720 
4721     if (ST->hasSSE2())
4722       if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
4723         return Entry->Cost;
4724   }
4725 
4726   auto *ValVTy = cast<FixedVectorType>(ValTy);
4727   unsigned NumVecElts = ValVTy->getNumElements();
4728 
4729   auto *Ty = ValVTy;
4730   InstructionCost MinMaxCost = 0;
4731   if (LT.first != 1 && MTy.isVector() &&
4732       MTy.getVectorNumElements() < ValVTy->getNumElements()) {
4733     // Type needs to be split. We need LT.first - 1 operations ops.
4734     Ty = FixedVectorType::get(ValVTy->getElementType(),
4735                               MTy.getVectorNumElements());
4736     auto *SubCondTy = FixedVectorType::get(CondTy->getElementType(),
4737                                            MTy.getVectorNumElements());
4738     MinMaxCost = getMinMaxCost(Ty, SubCondTy, IsUnsigned);
4739     MinMaxCost *= LT.first - 1;
4740     NumVecElts = MTy.getVectorNumElements();
4741   }
4742 
4743   if (ST->hasBWI())
4744     if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy))
4745       return MinMaxCost + Entry->Cost;
4746 
4747   if (ST->hasAVX())
4748     if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
4749       return MinMaxCost + Entry->Cost;
4750 
4751   if (ST->hasSSE41())
4752     if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy))
4753       return MinMaxCost + Entry->Cost;
4754 
4755   if (ST->hasSSE2())
4756     if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
4757       return MinMaxCost + Entry->Cost;
4758 
4759   unsigned ScalarSize = ValTy->getScalarSizeInBits();
4760 
4761   // Special case power of 2 reductions where the scalar type isn't changed
4762   // by type legalization.
4763   if (!isPowerOf2_32(ValVTy->getNumElements()) ||
4764       ScalarSize != MTy.getScalarSizeInBits())
4765     return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsUnsigned, CostKind);
4766 
4767   // Now handle reduction with the legal type, taking into account size changes
4768   // at each level.
4769   while (NumVecElts > 1) {
4770     // Determine the size of the remaining vector we need to reduce.
4771     unsigned Size = NumVecElts * ScalarSize;
4772     NumVecElts /= 2;
4773     // If we're reducing from 256/512 bits, use an extract_subvector.
4774     if (Size > 128) {
4775       auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts);
4776       MinMaxCost +=
4777           getShuffleCost(TTI::SK_ExtractSubvector, Ty, None, NumVecElts, SubTy);
4778       Ty = SubTy;
4779     } else if (Size == 128) {
4780       // Reducing from 128 bits is a permute of v2f64/v2i64.
4781       VectorType *ShufTy;
4782       if (ValTy->isFloatingPointTy())
4783         ShufTy =
4784             FixedVectorType::get(Type::getDoubleTy(ValTy->getContext()), 2);
4785       else
4786         ShufTy = FixedVectorType::get(Type::getInt64Ty(ValTy->getContext()), 2);
4787       MinMaxCost +=
4788           getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4789     } else if (Size == 64) {
4790       // Reducing from 64 bits is a shuffle of v4f32/v4i32.
4791       FixedVectorType *ShufTy;
4792       if (ValTy->isFloatingPointTy())
4793         ShufTy = FixedVectorType::get(Type::getFloatTy(ValTy->getContext()), 4);
4794       else
4795         ShufTy = FixedVectorType::get(Type::getInt32Ty(ValTy->getContext()), 4);
4796       MinMaxCost +=
4797           getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4798     } else {
4799       // Reducing from smaller size is a shift by immediate.
4800       auto *ShiftTy = FixedVectorType::get(
4801           Type::getIntNTy(ValTy->getContext(), Size), 128 / Size);
4802       MinMaxCost += getArithmeticInstrCost(
4803           Instruction::LShr, ShiftTy, TTI::TCK_RecipThroughput,
4804           TargetTransformInfo::OK_AnyValue,
4805           TargetTransformInfo::OK_UniformConstantValue,
4806           TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
4807     }
4808 
4809     // Add the arithmetic op for this level.
4810     auto *SubCondTy =
4811         FixedVectorType::get(CondTy->getElementType(), Ty->getNumElements());
4812     MinMaxCost += getMinMaxCost(Ty, SubCondTy, IsUnsigned);
4813   }
4814 
4815   // Add the final extract element to the cost.
4816   return MinMaxCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
4817 }
4818 
4819 /// Calculate the cost of materializing a 64-bit value. This helper
4820 /// method might only calculate a fraction of a larger immediate. Therefore it
4821 /// is valid to return a cost of ZERO.
4822 InstructionCost X86TTIImpl::getIntImmCost(int64_t Val) {
4823   if (Val == 0)
4824     return TTI::TCC_Free;
4825 
4826   if (isInt<32>(Val))
4827     return TTI::TCC_Basic;
4828 
4829   return 2 * TTI::TCC_Basic;
4830 }
4831 
4832 InstructionCost X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
4833                                           TTI::TargetCostKind CostKind) {
4834   assert(Ty->isIntegerTy());
4835 
4836   unsigned BitSize = Ty->getPrimitiveSizeInBits();
4837   if (BitSize == 0)
4838     return ~0U;
4839 
4840   // Never hoist constants larger than 128bit, because this might lead to
4841   // incorrect code generation or assertions in codegen.
4842   // Fixme: Create a cost model for types larger than i128 once the codegen
4843   // issues have been fixed.
4844   if (BitSize > 128)
4845     return TTI::TCC_Free;
4846 
4847   if (Imm == 0)
4848     return TTI::TCC_Free;
4849 
4850   // Sign-extend all constants to a multiple of 64-bit.
4851   APInt ImmVal = Imm;
4852   if (BitSize % 64 != 0)
4853     ImmVal = Imm.sext(alignTo(BitSize, 64));
4854 
4855   // Split the constant into 64-bit chunks and calculate the cost for each
4856   // chunk.
4857   InstructionCost Cost = 0;
4858   for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
4859     APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
4860     int64_t Val = Tmp.getSExtValue();
4861     Cost += getIntImmCost(Val);
4862   }
4863   // We need at least one instruction to materialize the constant.
4864   return std::max<InstructionCost>(1, Cost);
4865 }
4866 
4867 InstructionCost X86TTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
4868                                               const APInt &Imm, Type *Ty,
4869                                               TTI::TargetCostKind CostKind,
4870                                               Instruction *Inst) {
4871   assert(Ty->isIntegerTy());
4872 
4873   unsigned BitSize = Ty->getPrimitiveSizeInBits();
4874   // There is no cost model for constants with a bit size of 0. Return TCC_Free
4875   // here, so that constant hoisting will ignore this constant.
4876   if (BitSize == 0)
4877     return TTI::TCC_Free;
4878 
4879   unsigned ImmIdx = ~0U;
4880   switch (Opcode) {
4881   default:
4882     return TTI::TCC_Free;
4883   case Instruction::GetElementPtr:
4884     // Always hoist the base address of a GetElementPtr. This prevents the
4885     // creation of new constants for every base constant that gets constant
4886     // folded with the offset.
4887     if (Idx == 0)
4888       return 2 * TTI::TCC_Basic;
4889     return TTI::TCC_Free;
4890   case Instruction::Store:
4891     ImmIdx = 0;
4892     break;
4893   case Instruction::ICmp:
4894     // This is an imperfect hack to prevent constant hoisting of
4895     // compares that might be trying to check if a 64-bit value fits in
4896     // 32-bits. The backend can optimize these cases using a right shift by 32.
4897     // Ideally we would check the compare predicate here. There also other
4898     // similar immediates the backend can use shifts for.
4899     if (Idx == 1 && Imm.getBitWidth() == 64) {
4900       uint64_t ImmVal = Imm.getZExtValue();
4901       if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff)
4902         return TTI::TCC_Free;
4903     }
4904     ImmIdx = 1;
4905     break;
4906   case Instruction::And:
4907     // We support 64-bit ANDs with immediates with 32-bits of leading zeroes
4908     // by using a 32-bit operation with implicit zero extension. Detect such
4909     // immediates here as the normal path expects bit 31 to be sign extended.
4910     if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue()))
4911       return TTI::TCC_Free;
4912     ImmIdx = 1;
4913     break;
4914   case Instruction::Add:
4915   case Instruction::Sub:
4916     // For add/sub, we can use the opposite instruction for INT32_MIN.
4917     if (Idx == 1 && Imm.getBitWidth() == 64 && Imm.getZExtValue() == 0x80000000)
4918       return TTI::TCC_Free;
4919     ImmIdx = 1;
4920     break;
4921   case Instruction::UDiv:
4922   case Instruction::SDiv:
4923   case Instruction::URem:
4924   case Instruction::SRem:
4925     // Division by constant is typically expanded later into a different
4926     // instruction sequence. This completely changes the constants.
4927     // Report them as "free" to stop ConstantHoist from marking them as opaque.
4928     return TTI::TCC_Free;
4929   case Instruction::Mul:
4930   case Instruction::Or:
4931   case Instruction::Xor:
4932     ImmIdx = 1;
4933     break;
4934   // Always return TCC_Free for the shift value of a shift instruction.
4935   case Instruction::Shl:
4936   case Instruction::LShr:
4937   case Instruction::AShr:
4938     if (Idx == 1)
4939       return TTI::TCC_Free;
4940     break;
4941   case Instruction::Trunc:
4942   case Instruction::ZExt:
4943   case Instruction::SExt:
4944   case Instruction::IntToPtr:
4945   case Instruction::PtrToInt:
4946   case Instruction::BitCast:
4947   case Instruction::PHI:
4948   case Instruction::Call:
4949   case Instruction::Select:
4950   case Instruction::Ret:
4951   case Instruction::Load:
4952     break;
4953   }
4954 
4955   if (Idx == ImmIdx) {
4956     int NumConstants = divideCeil(BitSize, 64);
4957     InstructionCost Cost = X86TTIImpl::getIntImmCost(Imm, Ty, CostKind);
4958     return (Cost <= NumConstants * TTI::TCC_Basic)
4959                ? static_cast<int>(TTI::TCC_Free)
4960                : Cost;
4961   }
4962 
4963   return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind);
4964 }
4965 
4966 InstructionCost X86TTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
4967                                                 const APInt &Imm, Type *Ty,
4968                                                 TTI::TargetCostKind CostKind) {
4969   assert(Ty->isIntegerTy());
4970 
4971   unsigned BitSize = Ty->getPrimitiveSizeInBits();
4972   // There is no cost model for constants with a bit size of 0. Return TCC_Free
4973   // here, so that constant hoisting will ignore this constant.
4974   if (BitSize == 0)
4975     return TTI::TCC_Free;
4976 
4977   switch (IID) {
4978   default:
4979     return TTI::TCC_Free;
4980   case Intrinsic::sadd_with_overflow:
4981   case Intrinsic::uadd_with_overflow:
4982   case Intrinsic::ssub_with_overflow:
4983   case Intrinsic::usub_with_overflow:
4984   case Intrinsic::smul_with_overflow:
4985   case Intrinsic::umul_with_overflow:
4986     if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
4987       return TTI::TCC_Free;
4988     break;
4989   case Intrinsic::experimental_stackmap:
4990     if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
4991       return TTI::TCC_Free;
4992     break;
4993   case Intrinsic::experimental_patchpoint_void:
4994   case Intrinsic::experimental_patchpoint_i64:
4995     if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
4996       return TTI::TCC_Free;
4997     break;
4998   }
4999   return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind);
5000 }
5001 
5002 InstructionCost X86TTIImpl::getCFInstrCost(unsigned Opcode,
5003                                            TTI::TargetCostKind CostKind,
5004                                            const Instruction *I) {
5005   if (CostKind != TTI::TCK_RecipThroughput)
5006     return Opcode == Instruction::PHI ? 0 : 1;
5007   // Branches are assumed to be predicted.
5008   return 0;
5009 }
5010 
5011 int X86TTIImpl::getGatherOverhead() const {
5012   // Some CPUs have more overhead for gather. The specified overhead is relative
5013   // to the Load operation. "2" is the number provided by Intel architects. This
5014   // parameter is used for cost estimation of Gather Op and comparison with
5015   // other alternatives.
5016   // TODO: Remove the explicit hasAVX512()?, That would mean we would only
5017   // enable gather with a -march.
5018   if (ST->hasAVX512() || (ST->hasAVX2() && ST->hasFastGather()))
5019     return 2;
5020 
5021   return 1024;
5022 }
5023 
5024 int X86TTIImpl::getScatterOverhead() const {
5025   if (ST->hasAVX512())
5026     return 2;
5027 
5028   return 1024;
5029 }
5030 
5031 // Return an average cost of Gather / Scatter instruction, maybe improved later.
5032 // FIXME: Add TargetCostKind support.
5033 InstructionCost X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy,
5034                                             const Value *Ptr, Align Alignment,
5035                                             unsigned AddressSpace) {
5036 
5037   assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost");
5038   unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements();
5039 
5040   // Try to reduce index size from 64 bit (default for GEP)
5041   // to 32. It is essential for VF 16. If the index can't be reduced to 32, the
5042   // operation will use 16 x 64 indices which do not fit in a zmm and needs
5043   // to split. Also check that the base pointer is the same for all lanes,
5044   // and that there's at most one variable index.
5045   auto getIndexSizeInBits = [](const Value *Ptr, const DataLayout &DL) {
5046     unsigned IndexSize = DL.getPointerSizeInBits();
5047     const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
5048     if (IndexSize < 64 || !GEP)
5049       return IndexSize;
5050 
5051     unsigned NumOfVarIndices = 0;
5052     const Value *Ptrs = GEP->getPointerOperand();
5053     if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs))
5054       return IndexSize;
5055     for (unsigned i = 1; i < GEP->getNumOperands(); ++i) {
5056       if (isa<Constant>(GEP->getOperand(i)))
5057         continue;
5058       Type *IndxTy = GEP->getOperand(i)->getType();
5059       if (auto *IndexVTy = dyn_cast<VectorType>(IndxTy))
5060         IndxTy = IndexVTy->getElementType();
5061       if ((IndxTy->getPrimitiveSizeInBits() == 64 &&
5062           !isa<SExtInst>(GEP->getOperand(i))) ||
5063          ++NumOfVarIndices > 1)
5064         return IndexSize; // 64
5065     }
5066     return (unsigned)32;
5067   };
5068 
5069   // Trying to reduce IndexSize to 32 bits for vector 16.
5070   // By default the IndexSize is equal to pointer size.
5071   unsigned IndexSize = (ST->hasAVX512() && VF >= 16)
5072                            ? getIndexSizeInBits(Ptr, DL)
5073                            : DL.getPointerSizeInBits();
5074 
5075   auto *IndexVTy = FixedVectorType::get(
5076       IntegerType::get(SrcVTy->getContext(), IndexSize), VF);
5077   std::pair<InstructionCost, MVT> IdxsLT =
5078       TLI->getTypeLegalizationCost(DL, IndexVTy);
5079   std::pair<InstructionCost, MVT> SrcLT =
5080       TLI->getTypeLegalizationCost(DL, SrcVTy);
5081   InstructionCost::CostType SplitFactor =
5082       *std::max(IdxsLT.first, SrcLT.first).getValue();
5083   if (SplitFactor > 1) {
5084     // Handle splitting of vector of pointers
5085     auto *SplitSrcTy =
5086         FixedVectorType::get(SrcVTy->getScalarType(), VF / SplitFactor);
5087     return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment,
5088                                          AddressSpace);
5089   }
5090 
5091   // The gather / scatter cost is given by Intel architects. It is a rough
5092   // number since we are looking at one instruction in a time.
5093   const int GSOverhead = (Opcode == Instruction::Load)
5094                              ? getGatherOverhead()
5095                              : getScatterOverhead();
5096   return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
5097                                            MaybeAlign(Alignment), AddressSpace,
5098                                            TTI::TCK_RecipThroughput);
5099 }
5100 
5101 /// Return the cost of full scalarization of gather / scatter operation.
5102 ///
5103 /// Opcode - Load or Store instruction.
5104 /// SrcVTy - The type of the data vector that should be gathered or scattered.
5105 /// VariableMask - The mask is non-constant at compile time.
5106 /// Alignment - Alignment for one element.
5107 /// AddressSpace - pointer[s] address space.
5108 ///
5109 /// FIXME: Add TargetCostKind support.
5110 InstructionCost X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy,
5111                                             bool VariableMask, Align Alignment,
5112                                             unsigned AddressSpace) {
5113   Type *ScalarTy = SrcVTy->getScalarType();
5114   unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements();
5115   APInt DemandedElts = APInt::getAllOnes(VF);
5116   TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
5117 
5118   InstructionCost MaskUnpackCost = 0;
5119   if (VariableMask) {
5120     auto *MaskTy =
5121         FixedVectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF);
5122     MaskUnpackCost = getScalarizationOverhead(
5123         MaskTy, DemandedElts, /*Insert=*/false, /*Extract=*/true);
5124     InstructionCost ScalarCompareCost = getCmpSelInstrCost(
5125         Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()), nullptr,
5126         CmpInst::BAD_ICMP_PREDICATE, CostKind);
5127     InstructionCost BranchCost = getCFInstrCost(Instruction::Br, CostKind);
5128     MaskUnpackCost += VF * (BranchCost + ScalarCompareCost);
5129   }
5130 
5131   InstructionCost AddressUnpackCost = getScalarizationOverhead(
5132       FixedVectorType::get(ScalarTy->getPointerTo(), VF), DemandedElts,
5133       /*Insert=*/false, /*Extract=*/true);
5134 
5135   // The cost of the scalar loads/stores.
5136   InstructionCost MemoryOpCost =
5137       VF * getMemoryOpCost(Opcode, ScalarTy, MaybeAlign(Alignment),
5138                            AddressSpace, CostKind);
5139 
5140   // The cost of forming the vector from loaded scalars/
5141   // scalarizing the vector to perform scalar stores.
5142   InstructionCost InsertExtractCost =
5143       getScalarizationOverhead(cast<FixedVectorType>(SrcVTy), DemandedElts,
5144                                /*Insert=*/Opcode == Instruction::Load,
5145                                /*Extract=*/Opcode == Instruction::Store);
5146 
5147   return AddressUnpackCost + MemoryOpCost + MaskUnpackCost + InsertExtractCost;
5148 }
5149 
5150 /// Calculate the cost of Gather / Scatter operation
5151 InstructionCost X86TTIImpl::getGatherScatterOpCost(
5152     unsigned Opcode, Type *SrcVTy, const Value *Ptr, bool VariableMask,
5153     Align Alignment, TTI::TargetCostKind CostKind,
5154     const Instruction *I = nullptr) {
5155   if (CostKind != TTI::TCK_RecipThroughput) {
5156     if ((Opcode == Instruction::Load &&
5157          isLegalMaskedGather(SrcVTy, Align(Alignment)) &&
5158          !forceScalarizeMaskedGather(cast<VectorType>(SrcVTy),
5159                                      Align(Alignment))) ||
5160         (Opcode == Instruction::Store &&
5161          isLegalMaskedScatter(SrcVTy, Align(Alignment)) &&
5162          !forceScalarizeMaskedScatter(cast<VectorType>(SrcVTy),
5163                                       Align(Alignment))))
5164       return 1;
5165     return BaseT::getGatherScatterOpCost(Opcode, SrcVTy, Ptr, VariableMask,
5166                                          Alignment, CostKind, I);
5167   }
5168 
5169   assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter");
5170   PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
5171   if (!PtrTy && Ptr->getType()->isVectorTy())
5172     PtrTy = dyn_cast<PointerType>(
5173         cast<VectorType>(Ptr->getType())->getElementType());
5174   assert(PtrTy && "Unexpected type for Ptr argument");
5175   unsigned AddressSpace = PtrTy->getAddressSpace();
5176 
5177   if ((Opcode == Instruction::Load &&
5178        (!isLegalMaskedGather(SrcVTy, Align(Alignment)) ||
5179         forceScalarizeMaskedGather(cast<VectorType>(SrcVTy),
5180                                    Align(Alignment)))) ||
5181       (Opcode == Instruction::Store &&
5182        (!isLegalMaskedScatter(SrcVTy, Align(Alignment)) ||
5183         forceScalarizeMaskedScatter(cast<VectorType>(SrcVTy),
5184                                     Align(Alignment)))))
5185     return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment,
5186                            AddressSpace);
5187 
5188   return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace);
5189 }
5190 
5191 bool X86TTIImpl::isLSRCostLess(const TargetTransformInfo::LSRCost &C1,
5192                                const TargetTransformInfo::LSRCost &C2) {
5193     // X86 specific here are "instruction number 1st priority".
5194     return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost,
5195                     C1.NumIVMuls, C1.NumBaseAdds,
5196                     C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
5197            std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost,
5198                     C2.NumIVMuls, C2.NumBaseAdds,
5199                     C2.ScaleCost, C2.ImmCost, C2.SetupCost);
5200 }
5201 
5202 bool X86TTIImpl::canMacroFuseCmp() {
5203   return ST->hasMacroFusion() || ST->hasBranchFusion();
5204 }
5205 
5206 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy, Align Alignment) {
5207   if (!ST->hasAVX())
5208     return false;
5209 
5210   // The backend can't handle a single element vector.
5211   if (isa<VectorType>(DataTy) &&
5212       cast<FixedVectorType>(DataTy)->getNumElements() == 1)
5213     return false;
5214   Type *ScalarTy = DataTy->getScalarType();
5215 
5216   if (ScalarTy->isPointerTy())
5217     return true;
5218 
5219   if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
5220     return true;
5221 
5222   if (ScalarTy->isHalfTy() && ST->hasBWI() && ST->hasFP16())
5223     return true;
5224 
5225   if (!ScalarTy->isIntegerTy())
5226     return false;
5227 
5228   unsigned IntWidth = ScalarTy->getIntegerBitWidth();
5229   return IntWidth == 32 || IntWidth == 64 ||
5230          ((IntWidth == 8 || IntWidth == 16) && ST->hasBWI());
5231 }
5232 
5233 bool X86TTIImpl::isLegalMaskedStore(Type *DataType, Align Alignment) {
5234   return isLegalMaskedLoad(DataType, Alignment);
5235 }
5236 
5237 bool X86TTIImpl::isLegalNTLoad(Type *DataType, Align Alignment) {
5238   unsigned DataSize = DL.getTypeStoreSize(DataType);
5239   // The only supported nontemporal loads are for aligned vectors of 16 or 32
5240   // bytes.  Note that 32-byte nontemporal vector loads are supported by AVX2
5241   // (the equivalent stores only require AVX).
5242   if (Alignment >= DataSize && (DataSize == 16 || DataSize == 32))
5243     return DataSize == 16 ?  ST->hasSSE1() : ST->hasAVX2();
5244 
5245   return false;
5246 }
5247 
5248 bool X86TTIImpl::isLegalNTStore(Type *DataType, Align Alignment) {
5249   unsigned DataSize = DL.getTypeStoreSize(DataType);
5250 
5251   // SSE4A supports nontemporal stores of float and double at arbitrary
5252   // alignment.
5253   if (ST->hasSSE4A() && (DataType->isFloatTy() || DataType->isDoubleTy()))
5254     return true;
5255 
5256   // Besides the SSE4A subtarget exception above, only aligned stores are
5257   // available nontemporaly on any other subtarget.  And only stores with a size
5258   // of 4..32 bytes (powers of 2, only) are permitted.
5259   if (Alignment < DataSize || DataSize < 4 || DataSize > 32 ||
5260       !isPowerOf2_32(DataSize))
5261     return false;
5262 
5263   // 32-byte vector nontemporal stores are supported by AVX (the equivalent
5264   // loads require AVX2).
5265   if (DataSize == 32)
5266     return ST->hasAVX();
5267   if (DataSize == 16)
5268     return ST->hasSSE1();
5269   return true;
5270 }
5271 
5272 bool X86TTIImpl::isLegalBroadcastLoad(Type *ElementTy,
5273                                       ElementCount NumElements) const {
5274   // movddup
5275   return ST->hasSSE3() && !NumElements.isScalable() &&
5276          NumElements.getFixedValue() == 2 &&
5277          ElementTy == Type::getDoubleTy(ElementTy->getContext());
5278 }
5279 
5280 bool X86TTIImpl::isLegalMaskedExpandLoad(Type *DataTy) {
5281   if (!isa<VectorType>(DataTy))
5282     return false;
5283 
5284   if (!ST->hasAVX512())
5285     return false;
5286 
5287   // The backend can't handle a single element vector.
5288   if (cast<FixedVectorType>(DataTy)->getNumElements() == 1)
5289     return false;
5290 
5291   Type *ScalarTy = cast<VectorType>(DataTy)->getElementType();
5292 
5293   if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
5294     return true;
5295 
5296   if (!ScalarTy->isIntegerTy())
5297     return false;
5298 
5299   unsigned IntWidth = ScalarTy->getIntegerBitWidth();
5300   return IntWidth == 32 || IntWidth == 64 ||
5301          ((IntWidth == 8 || IntWidth == 16) && ST->hasVBMI2());
5302 }
5303 
5304 bool X86TTIImpl::isLegalMaskedCompressStore(Type *DataTy) {
5305   return isLegalMaskedExpandLoad(DataTy);
5306 }
5307 
5308 bool X86TTIImpl::supportsGather() const {
5309   // Some CPUs have better gather performance than others.
5310   // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only
5311   // enable gather with a -march.
5312   return ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2());
5313 }
5314 
5315 bool X86TTIImpl::forceScalarizeMaskedGather(VectorType *VTy, Align Alignment) {
5316   // Gather / Scatter for vector 2 is not profitable on KNL / SKX
5317   // Vector-4 of gather/scatter instruction does not exist on KNL. We can extend
5318   // it to 8 elements, but zeroing upper bits of the mask vector will add more
5319   // instructions. Right now we give the scalar cost of vector-4 for KNL. TODO:
5320   // Check, maybe the gather/scatter instruction is better in the VariableMask
5321   // case.
5322   unsigned NumElts = cast<FixedVectorType>(VTy)->getNumElements();
5323   return NumElts == 1 ||
5324          (ST->hasAVX512() && (NumElts == 2 || (NumElts == 4 && !ST->hasVLX())));
5325 }
5326 
5327 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy, Align Alignment) {
5328   if (!supportsGather())
5329     return false;
5330   Type *ScalarTy = DataTy->getScalarType();
5331   if (ScalarTy->isPointerTy())
5332     return true;
5333 
5334   if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
5335     return true;
5336 
5337   if (!ScalarTy->isIntegerTy())
5338     return false;
5339 
5340   unsigned IntWidth = ScalarTy->getIntegerBitWidth();
5341   return IntWidth == 32 || IntWidth == 64;
5342 }
5343 
5344 bool X86TTIImpl::isLegalAltInstr(VectorType *VecTy, unsigned Opcode0,
5345                                  unsigned Opcode1,
5346                                  const SmallBitVector &OpcodeMask) const {
5347   // ADDSUBPS  4xf32 SSE3
5348   // VADDSUBPS 4xf32 AVX
5349   // VADDSUBPS 8xf32 AVX2
5350   // ADDSUBPD  2xf64 SSE3
5351   // VADDSUBPD 2xf64 AVX
5352   // VADDSUBPD 4xf64 AVX2
5353 
5354   unsigned NumElements = cast<FixedVectorType>(VecTy)->getNumElements();
5355   assert(OpcodeMask.size() == NumElements && "Mask and VecTy are incompatible");
5356   if (!isPowerOf2_32(NumElements))
5357     return false;
5358   // Check the opcode pattern. We apply the mask on the opcode arguments and
5359   // then check if it is what we expect.
5360   for (int Lane : seq<int>(0, NumElements)) {
5361     unsigned Opc = OpcodeMask.test(Lane) ? Opcode1 : Opcode0;
5362     // We expect FSub for even lanes and FAdd for odd lanes.
5363     if (Lane % 2 == 0 && Opc != Instruction::FSub)
5364       return false;
5365     if (Lane % 2 == 1 && Opc != Instruction::FAdd)
5366       return false;
5367   }
5368   // Now check that the pattern is supported by the target ISA.
5369   Type *ElemTy = cast<VectorType>(VecTy)->getElementType();
5370   if (ElemTy->isFloatTy())
5371     return ST->hasSSE3() && NumElements % 4 == 0;
5372   if (ElemTy->isDoubleTy())
5373     return ST->hasSSE3() && NumElements % 2 == 0;
5374   return false;
5375 }
5376 
5377 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType, Align Alignment) {
5378   // AVX2 doesn't support scatter
5379   if (!ST->hasAVX512())
5380     return false;
5381   return isLegalMaskedGather(DataType, Alignment);
5382 }
5383 
5384 bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) {
5385   EVT VT = TLI->getValueType(DL, DataType);
5386   return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT);
5387 }
5388 
5389 bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
5390   return false;
5391 }
5392 
5393 bool X86TTIImpl::areInlineCompatible(const Function *Caller,
5394                                      const Function *Callee) const {
5395   const TargetMachine &TM = getTLI()->getTargetMachine();
5396 
5397   // Work this as a subsetting of subtarget features.
5398   const FeatureBitset &CallerBits =
5399       TM.getSubtargetImpl(*Caller)->getFeatureBits();
5400   const FeatureBitset &CalleeBits =
5401       TM.getSubtargetImpl(*Callee)->getFeatureBits();
5402 
5403   // Check whether features are the same (apart from the ignore list).
5404   FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList;
5405   FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList;
5406   if (RealCallerBits == RealCalleeBits)
5407     return true;
5408 
5409   // If the features are a subset, we need to additionally check for calls
5410   // that may become ABI-incompatible as a result of inlining.
5411   if ((RealCallerBits & RealCalleeBits) != RealCalleeBits)
5412     return false;
5413 
5414   for (const Instruction &I : instructions(Callee)) {
5415     if (const auto *CB = dyn_cast<CallBase>(&I)) {
5416       SmallVector<Type *, 8> Types;
5417       for (Value *Arg : CB->args())
5418         Types.push_back(Arg->getType());
5419       if (!CB->getType()->isVoidTy())
5420         Types.push_back(CB->getType());
5421 
5422       // Simple types are always ABI compatible.
5423       auto IsSimpleTy = [](Type *Ty) {
5424         return !Ty->isVectorTy() && !Ty->isAggregateType();
5425       };
5426       if (all_of(Types, IsSimpleTy))
5427         continue;
5428 
5429       if (Function *NestedCallee = CB->getCalledFunction()) {
5430         // Assume that intrinsics are always ABI compatible.
5431         if (NestedCallee->isIntrinsic())
5432           continue;
5433 
5434         // Do a precise compatibility check.
5435         if (!areTypesABICompatible(Caller, NestedCallee, Types))
5436           return false;
5437       } else {
5438         // We don't know the target features of the callee,
5439         // assume it is incompatible.
5440         return false;
5441       }
5442     }
5443   }
5444   return true;
5445 }
5446 
5447 bool X86TTIImpl::areTypesABICompatible(const Function *Caller,
5448                                        const Function *Callee,
5449                                        const ArrayRef<Type *> &Types) const {
5450   if (!BaseT::areTypesABICompatible(Caller, Callee, Types))
5451     return false;
5452 
5453   // If we get here, we know the target features match. If one function
5454   // considers 512-bit vectors legal and the other does not, consider them
5455   // incompatible.
5456   const TargetMachine &TM = getTLI()->getTargetMachine();
5457 
5458   if (TM.getSubtarget<X86Subtarget>(*Caller).useAVX512Regs() ==
5459       TM.getSubtarget<X86Subtarget>(*Callee).useAVX512Regs())
5460     return true;
5461 
5462   // Consider the arguments compatible if they aren't vectors or aggregates.
5463   // FIXME: Look at the size of vectors.
5464   // FIXME: Look at the element types of aggregates to see if there are vectors.
5465   return llvm::none_of(Types,
5466       [](Type *T) { return T->isVectorTy() || T->isAggregateType(); });
5467 }
5468 
5469 X86TTIImpl::TTI::MemCmpExpansionOptions
5470 X86TTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
5471   TTI::MemCmpExpansionOptions Options;
5472   Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize);
5473   Options.NumLoadsPerBlock = 2;
5474   // All GPR and vector loads can be unaligned.
5475   Options.AllowOverlappingLoads = true;
5476   if (IsZeroCmp) {
5477     // Only enable vector loads for equality comparison. Right now the vector
5478     // version is not as fast for three way compare (see #33329).
5479     const unsigned PreferredWidth = ST->getPreferVectorWidth();
5480     if (PreferredWidth >= 512 && ST->hasAVX512()) Options.LoadSizes.push_back(64);
5481     if (PreferredWidth >= 256 && ST->hasAVX()) Options.LoadSizes.push_back(32);
5482     if (PreferredWidth >= 128 && ST->hasSSE2()) Options.LoadSizes.push_back(16);
5483   }
5484   if (ST->is64Bit()) {
5485     Options.LoadSizes.push_back(8);
5486   }
5487   Options.LoadSizes.push_back(4);
5488   Options.LoadSizes.push_back(2);
5489   Options.LoadSizes.push_back(1);
5490   return Options;
5491 }
5492 
5493 bool X86TTIImpl::prefersVectorizedAddressing() const {
5494   return supportsGather();
5495 }
5496 
5497 bool X86TTIImpl::supportsEfficientVectorElementLoadStore() const {
5498   return false;
5499 }
5500 
5501 bool X86TTIImpl::enableInterleavedAccessVectorization() {
5502   // TODO: We expect this to be beneficial regardless of arch,
5503   // but there are currently some unexplained performance artifacts on Atom.
5504   // As a temporary solution, disable on Atom.
5505   return !(ST->isAtom());
5506 }
5507 
5508 // Get estimation for interleaved load/store operations and strided load.
5509 // \p Indices contains indices for strided load.
5510 // \p Factor - the factor of interleaving.
5511 // AVX-512 provides 3-src shuffles that significantly reduces the cost.
5512 InstructionCost X86TTIImpl::getInterleavedMemoryOpCostAVX512(
5513     unsigned Opcode, FixedVectorType *VecTy, unsigned Factor,
5514     ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace,
5515     TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) {
5516   // VecTy for interleave memop is <VF*Factor x Elt>.
5517   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
5518   // VecTy = <12 x i32>.
5519 
5520   // Calculate the number of memory operations (NumOfMemOps), required
5521   // for load/store the VecTy.
5522   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
5523   unsigned VecTySize = DL.getTypeStoreSize(VecTy);
5524   unsigned LegalVTSize = LegalVT.getStoreSize();
5525   unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize;
5526 
5527   // Get the cost of one memory operation.
5528   auto *SingleMemOpTy = FixedVectorType::get(VecTy->getElementType(),
5529                                              LegalVT.getVectorNumElements());
5530   InstructionCost MemOpCost;
5531   bool UseMaskedMemOp = UseMaskForCond || UseMaskForGaps;
5532   if (UseMaskedMemOp)
5533     MemOpCost = getMaskedMemoryOpCost(Opcode, SingleMemOpTy, Alignment,
5534                                       AddressSpace, CostKind);
5535   else
5536     MemOpCost = getMemoryOpCost(Opcode, SingleMemOpTy, MaybeAlign(Alignment),
5537                                 AddressSpace, CostKind);
5538 
5539   unsigned VF = VecTy->getNumElements() / Factor;
5540   MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF);
5541 
5542   InstructionCost MaskCost;
5543   if (UseMaskedMemOp) {
5544     APInt DemandedLoadStoreElts = APInt::getZero(VecTy->getNumElements());
5545     for (unsigned Index : Indices) {
5546       assert(Index < Factor && "Invalid index for interleaved memory op");
5547       for (unsigned Elm = 0; Elm < VF; Elm++)
5548         DemandedLoadStoreElts.setBit(Index + Elm * Factor);
5549     }
5550 
5551     Type *I1Type = Type::getInt1Ty(VecTy->getContext());
5552 
5553     MaskCost = getReplicationShuffleCost(
5554         I1Type, Factor, VF,
5555         UseMaskForGaps ? DemandedLoadStoreElts
5556                        : APInt::getAllOnes(VecTy->getNumElements()),
5557         CostKind);
5558 
5559     // The Gaps mask is invariant and created outside the loop, therefore the
5560     // cost of creating it is not accounted for here. However if we have both
5561     // a MaskForGaps and some other mask that guards the execution of the
5562     // memory access, we need to account for the cost of And-ing the two masks
5563     // inside the loop.
5564     if (UseMaskForGaps) {
5565       auto *MaskVT = FixedVectorType::get(I1Type, VecTy->getNumElements());
5566       MaskCost += getArithmeticInstrCost(BinaryOperator::And, MaskVT, CostKind);
5567     }
5568   }
5569 
5570   if (Opcode == Instruction::Load) {
5571     // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl)
5572     // contain the cost of the optimized shuffle sequence that the
5573     // X86InterleavedAccess pass will generate.
5574     // The cost of loads and stores are computed separately from the table.
5575 
5576     // X86InterleavedAccess support only the following interleaved-access group.
5577     static const CostTblEntry AVX512InterleavedLoadTbl[] = {
5578         {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8
5579         {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8
5580         {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8
5581     };
5582 
5583     if (const auto *Entry =
5584             CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT))
5585       return MaskCost + NumOfMemOps * MemOpCost + Entry->Cost;
5586     //If an entry does not exist, fallback to the default implementation.
5587 
5588     // Kind of shuffle depends on number of loaded values.
5589     // If we load the entire data in one register, we can use a 1-src shuffle.
5590     // Otherwise, we'll merge 2 sources in each operation.
5591     TTI::ShuffleKind ShuffleKind =
5592         (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc;
5593 
5594     InstructionCost ShuffleCost =
5595         getShuffleCost(ShuffleKind, SingleMemOpTy, None, 0, nullptr);
5596 
5597     unsigned NumOfLoadsInInterleaveGrp =
5598         Indices.size() ? Indices.size() : Factor;
5599     auto *ResultTy = FixedVectorType::get(VecTy->getElementType(),
5600                                           VecTy->getNumElements() / Factor);
5601     InstructionCost NumOfResults =
5602         getTLI()->getTypeLegalizationCost(DL, ResultTy).first *
5603         NumOfLoadsInInterleaveGrp;
5604 
5605     // About a half of the loads may be folded in shuffles when we have only
5606     // one result. If we have more than one result, or the loads are masked,
5607     // we do not fold loads at all.
5608     unsigned NumOfUnfoldedLoads =
5609         UseMaskedMemOp || NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2;
5610 
5611     // Get a number of shuffle operations per result.
5612     unsigned NumOfShufflesPerResult =
5613         std::max((unsigned)1, (unsigned)(NumOfMemOps - 1));
5614 
5615     // The SK_MergeTwoSrc shuffle clobbers one of src operands.
5616     // When we have more than one destination, we need additional instructions
5617     // to keep sources.
5618     InstructionCost NumOfMoves = 0;
5619     if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc)
5620       NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2;
5621 
5622     InstructionCost Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost +
5623                            MaskCost + NumOfUnfoldedLoads * MemOpCost +
5624                            NumOfMoves;
5625 
5626     return Cost;
5627   }
5628 
5629   // Store.
5630   assert(Opcode == Instruction::Store &&
5631          "Expected Store Instruction at this  point");
5632   // X86InterleavedAccess support only the following interleaved-access group.
5633   static const CostTblEntry AVX512InterleavedStoreTbl[] = {
5634       {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store)
5635       {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store)
5636       {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store)
5637 
5638       {4, MVT::v8i8, 10},  // interleave 4 x 8i8  into 32i8  (and store)
5639       {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8  (and store)
5640       {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store)
5641       {4, MVT::v64i8, 24}  // interleave 4 x 32i8 into 256i8 (and store)
5642   };
5643 
5644   if (const auto *Entry =
5645           CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT))
5646     return MaskCost + NumOfMemOps * MemOpCost + Entry->Cost;
5647   //If an entry does not exist, fallback to the default implementation.
5648 
5649   // There is no strided stores meanwhile. And store can't be folded in
5650   // shuffle.
5651   unsigned NumOfSources = Factor; // The number of values to be merged.
5652   InstructionCost ShuffleCost =
5653       getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, None, 0, nullptr);
5654   unsigned NumOfShufflesPerStore = NumOfSources - 1;
5655 
5656   // The SK_MergeTwoSrc shuffle clobbers one of src operands.
5657   // We need additional instructions to keep sources.
5658   unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2;
5659   InstructionCost Cost =
5660       MaskCost +
5661       NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) +
5662       NumOfMoves;
5663   return Cost;
5664 }
5665 
5666 InstructionCost X86TTIImpl::getInterleavedMemoryOpCost(
5667     unsigned Opcode, Type *BaseTy, unsigned Factor, ArrayRef<unsigned> Indices,
5668     Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
5669     bool UseMaskForCond, bool UseMaskForGaps) {
5670   auto *VecTy = cast<FixedVectorType>(BaseTy);
5671 
5672   auto isSupportedOnAVX512 = [&](Type *VecTy, bool HasBW) {
5673     Type *EltTy = cast<VectorType>(VecTy)->getElementType();
5674     if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) ||
5675         EltTy->isIntegerTy(32) || EltTy->isPointerTy())
5676       return true;
5677     if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8) ||
5678         (!ST->useSoftFloat() && ST->hasFP16() && EltTy->isHalfTy()))
5679       return HasBW;
5680     return false;
5681   };
5682   if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI()))
5683     return getInterleavedMemoryOpCostAVX512(
5684         Opcode, VecTy, Factor, Indices, Alignment,
5685         AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps);
5686 
5687   if (UseMaskForCond || UseMaskForGaps)
5688     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5689                                              Alignment, AddressSpace, CostKind,
5690                                              UseMaskForCond, UseMaskForGaps);
5691 
5692   // Get estimation for interleaved load/store operations for SSE-AVX2.
5693   // As opposed to AVX-512, SSE-AVX2 do not have generic shuffles that allow
5694   // computing the cost using a generic formula as a function of generic
5695   // shuffles. We therefore use a lookup table instead, filled according to
5696   // the instruction sequences that codegen currently generates.
5697 
5698   // VecTy for interleave memop is <VF*Factor x Elt>.
5699   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
5700   // VecTy = <12 x i32>.
5701   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
5702 
5703   // This function can be called with VecTy=<6xi128>, Factor=3, in which case
5704   // the VF=2, while v2i128 is an unsupported MVT vector type
5705   // (see MachineValueType.h::getVectorVT()).
5706   if (!LegalVT.isVector())
5707     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5708                                              Alignment, AddressSpace, CostKind);
5709 
5710   unsigned VF = VecTy->getNumElements() / Factor;
5711   Type *ScalarTy = VecTy->getElementType();
5712   // Deduplicate entries, model floats/pointers as appropriately-sized integers.
5713   if (!ScalarTy->isIntegerTy())
5714     ScalarTy =
5715         Type::getIntNTy(ScalarTy->getContext(), DL.getTypeSizeInBits(ScalarTy));
5716 
5717   // Get the cost of all the memory operations.
5718   // FIXME: discount dead loads.
5719   InstructionCost MemOpCosts = getMemoryOpCost(
5720       Opcode, VecTy, MaybeAlign(Alignment), AddressSpace, CostKind);
5721 
5722   auto *VT = FixedVectorType::get(ScalarTy, VF);
5723   EVT ETy = TLI->getValueType(DL, VT);
5724   if (!ETy.isSimple())
5725     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5726                                              Alignment, AddressSpace, CostKind);
5727 
5728   // TODO: Complete for other data-types and strides.
5729   // Each combination of Stride, element bit width and VF results in a different
5730   // sequence; The cost tables are therefore accessed with:
5731   // Factor (stride) and VectorType=VFxiN.
5732   // The Cost accounts only for the shuffle sequence;
5733   // The cost of the loads/stores is accounted for separately.
5734   //
5735   static const CostTblEntry AVX2InterleavedLoadTbl[] = {
5736       {2, MVT::v2i8, 2},  // (load 4i8 and) deinterleave into 2 x 2i8
5737       {2, MVT::v4i8, 2},  // (load 8i8 and) deinterleave into 2 x 4i8
5738       {2, MVT::v8i8, 2},  // (load 16i8 and) deinterleave into 2 x 8i8
5739       {2, MVT::v16i8, 4}, // (load 32i8 and) deinterleave into 2 x 16i8
5740       {2, MVT::v32i8, 6}, // (load 64i8 and) deinterleave into 2 x 32i8
5741 
5742       {2, MVT::v8i16, 6},   // (load 16i16 and) deinterleave into 2 x 8i16
5743       {2, MVT::v16i16, 9},  // (load 32i16 and) deinterleave into 2 x 16i16
5744       {2, MVT::v32i16, 18}, // (load 64i16 and) deinterleave into 2 x 32i16
5745 
5746       {2, MVT::v8i32, 4},   // (load 16i32 and) deinterleave into 2 x 8i32
5747       {2, MVT::v16i32, 8},  // (load 32i32 and) deinterleave into 2 x 16i32
5748       {2, MVT::v32i32, 16}, // (load 64i32 and) deinterleave into 2 x 32i32
5749 
5750       {2, MVT::v4i64, 4},   // (load 8i64 and) deinterleave into 2 x 4i64
5751       {2, MVT::v8i64, 8},   // (load 16i64 and) deinterleave into 2 x 8i64
5752       {2, MVT::v16i64, 16}, // (load 32i64 and) deinterleave into 2 x 16i64
5753       {2, MVT::v32i64, 32}, // (load 64i64 and) deinterleave into 2 x 32i64
5754 
5755       {3, MVT::v2i8, 3},   // (load 6i8 and) deinterleave into 3 x 2i8
5756       {3, MVT::v4i8, 3},   // (load 12i8 and) deinterleave into 3 x 4i8
5757       {3, MVT::v8i8, 6},   // (load 24i8 and) deinterleave into 3 x 8i8
5758       {3, MVT::v16i8, 11}, // (load 48i8 and) deinterleave into 3 x 16i8
5759       {3, MVT::v32i8, 14}, // (load 96i8 and) deinterleave into 3 x 32i8
5760 
5761       {3, MVT::v2i16, 5},   // (load 6i16 and) deinterleave into 3 x 2i16
5762       {3, MVT::v4i16, 7},   // (load 12i16 and) deinterleave into 3 x 4i16
5763       {3, MVT::v8i16, 9},   // (load 24i16 and) deinterleave into 3 x 8i16
5764       {3, MVT::v16i16, 28}, // (load 48i16 and) deinterleave into 3 x 16i16
5765       {3, MVT::v32i16, 56}, // (load 96i16 and) deinterleave into 3 x 32i16
5766 
5767       {3, MVT::v2i32, 3},   // (load 6i32 and) deinterleave into 3 x 2i32
5768       {3, MVT::v4i32, 3},   // (load 12i32 and) deinterleave into 3 x 4i32
5769       {3, MVT::v8i32, 7},   // (load 24i32 and) deinterleave into 3 x 8i32
5770       {3, MVT::v16i32, 14}, // (load 48i32 and) deinterleave into 3 x 16i32
5771       {3, MVT::v32i32, 32}, // (load 96i32 and) deinterleave into 3 x 32i32
5772 
5773       {3, MVT::v2i64, 1},   // (load 6i64 and) deinterleave into 3 x 2i64
5774       {3, MVT::v4i64, 5},   // (load 12i64 and) deinterleave into 3 x 4i64
5775       {3, MVT::v8i64, 10},  // (load 24i64 and) deinterleave into 3 x 8i64
5776       {3, MVT::v16i64, 20}, // (load 48i64 and) deinterleave into 3 x 16i64
5777 
5778       {4, MVT::v2i8, 4},   // (load 8i8 and) deinterleave into 4 x 2i8
5779       {4, MVT::v4i8, 4},   // (load 16i8 and) deinterleave into 4 x 4i8
5780       {4, MVT::v8i8, 12},  // (load 32i8 and) deinterleave into 4 x 8i8
5781       {4, MVT::v16i8, 24}, // (load 64i8 and) deinterleave into 4 x 16i8
5782       {4, MVT::v32i8, 56}, // (load 128i8 and) deinterleave into 4 x 32i8
5783 
5784       {4, MVT::v2i16, 6},    // (load 8i16 and) deinterleave into 4 x 2i16
5785       {4, MVT::v4i16, 17},   // (load 16i16 and) deinterleave into 4 x 4i16
5786       {4, MVT::v8i16, 33},   // (load 32i16 and) deinterleave into 4 x 8i16
5787       {4, MVT::v16i16, 75},  // (load 64i16 and) deinterleave into 4 x 16i16
5788       {4, MVT::v32i16, 150}, // (load 128i16 and) deinterleave into 4 x 32i16
5789 
5790       {4, MVT::v2i32, 4},   // (load 8i32 and) deinterleave into 4 x 2i32
5791       {4, MVT::v4i32, 8},   // (load 16i32 and) deinterleave into 4 x 4i32
5792       {4, MVT::v8i32, 16},  // (load 32i32 and) deinterleave into 4 x 8i32
5793       {4, MVT::v16i32, 32}, // (load 64i32 and) deinterleave into 4 x 16i32
5794       {4, MVT::v32i32, 68}, // (load 128i32 and) deinterleave into 4 x 32i32
5795 
5796       {4, MVT::v2i64, 6},  // (load 8i64 and) deinterleave into 4 x 2i64
5797       {4, MVT::v4i64, 8},  // (load 16i64 and) deinterleave into 4 x 4i64
5798       {4, MVT::v8i64, 20}, // (load 32i64 and) deinterleave into 4 x 8i64
5799       {4, MVT::v16i64, 40}, // (load 64i64 and) deinterleave into 4 x 16i64
5800 
5801       {6, MVT::v2i8, 6},   // (load 12i8 and) deinterleave into 6 x 2i8
5802       {6, MVT::v4i8, 14},  // (load 24i8 and) deinterleave into 6 x 4i8
5803       {6, MVT::v8i8, 18},  // (load 48i8 and) deinterleave into 6 x 8i8
5804       {6, MVT::v16i8, 43}, // (load 96i8 and) deinterleave into 6 x 16i8
5805       {6, MVT::v32i8, 82}, // (load 192i8 and) deinterleave into 6 x 32i8
5806 
5807       {6, MVT::v2i16, 13},   // (load 12i16 and) deinterleave into 6 x 2i16
5808       {6, MVT::v4i16, 9},    // (load 24i16 and) deinterleave into 6 x 4i16
5809       {6, MVT::v8i16, 39},   // (load 48i16 and) deinterleave into 6 x 8i16
5810       {6, MVT::v16i16, 106}, // (load 96i16 and) deinterleave into 6 x 16i16
5811       {6, MVT::v32i16, 212}, // (load 192i16 and) deinterleave into 6 x 32i16
5812 
5813       {6, MVT::v2i32, 6},   // (load 12i32 and) deinterleave into 6 x 2i32
5814       {6, MVT::v4i32, 15},  // (load 24i32 and) deinterleave into 6 x 4i32
5815       {6, MVT::v8i32, 31},  // (load 48i32 and) deinterleave into 6 x 8i32
5816       {6, MVT::v16i32, 64}, // (load 96i32 and) deinterleave into 6 x 16i32
5817 
5818       {6, MVT::v2i64, 6},  // (load 12i64 and) deinterleave into 6 x 2i64
5819       {6, MVT::v4i64, 18}, // (load 24i64 and) deinterleave into 6 x 4i64
5820       {6, MVT::v8i64, 36}, // (load 48i64 and) deinterleave into 6 x 8i64
5821 
5822       {8, MVT::v8i32, 40} // (load 64i32 and) deinterleave into 8 x 8i32
5823   };
5824 
5825   static const CostTblEntry SSSE3InterleavedLoadTbl[] = {
5826       {2, MVT::v4i16, 2},   // (load 8i16 and) deinterleave into 2 x 4i16
5827   };
5828 
5829   static const CostTblEntry SSE2InterleavedLoadTbl[] = {
5830       {2, MVT::v2i16, 2},   // (load 4i16 and) deinterleave into 2 x 2i16
5831       {2, MVT::v4i16, 7},   // (load 8i16 and) deinterleave into 2 x 4i16
5832 
5833       {2, MVT::v2i32, 2},   // (load 4i32 and) deinterleave into 2 x 2i32
5834       {2, MVT::v4i32, 2},   // (load 8i32 and) deinterleave into 2 x 4i32
5835 
5836       {2, MVT::v2i64, 2},   // (load 4i64 and) deinterleave into 2 x 2i64
5837   };
5838 
5839   static const CostTblEntry AVX2InterleavedStoreTbl[] = {
5840       {2, MVT::v16i8, 3}, // interleave 2 x 16i8 into 32i8 (and store)
5841       {2, MVT::v32i8, 4}, // interleave 2 x 32i8 into 64i8 (and store)
5842 
5843       {2, MVT::v8i16, 3},  // interleave 2 x 8i16 into 16i16 (and store)
5844       {2, MVT::v16i16, 4}, // interleave 2 x 16i16 into 32i16 (and store)
5845       {2, MVT::v32i16, 8}, // interleave 2 x 32i16 into 64i16 (and store)
5846 
5847       {2, MVT::v4i32, 2},   // interleave 2 x 4i32 into 8i32 (and store)
5848       {2, MVT::v8i32, 4},   // interleave 2 x 8i32 into 16i32 (and store)
5849       {2, MVT::v16i32, 8},  // interleave 2 x 16i32 into 32i32 (and store)
5850       {2, MVT::v32i32, 16}, // interleave 2 x 32i32 into 64i32 (and store)
5851 
5852       {2, MVT::v2i64, 2},   // interleave 2 x 2i64 into 4i64 (and store)
5853       {2, MVT::v4i64, 4},   // interleave 2 x 4i64 into 8i64 (and store)
5854       {2, MVT::v8i64, 8},   // interleave 2 x 8i64 into 16i64 (and store)
5855       {2, MVT::v16i64, 16}, // interleave 2 x 16i64 into 32i64 (and store)
5856       {2, MVT::v32i64, 32}, // interleave 2 x 32i64 into 64i64 (and store)
5857 
5858       {3, MVT::v2i8, 4},   // interleave 3 x 2i8 into 6i8 (and store)
5859       {3, MVT::v4i8, 4},   // interleave 3 x 4i8 into 12i8 (and store)
5860       {3, MVT::v8i8, 6},   // interleave 3 x 8i8 into 24i8 (and store)
5861       {3, MVT::v16i8, 11}, // interleave 3 x 16i8 into 48i8 (and store)
5862       {3, MVT::v32i8, 13}, // interleave 3 x 32i8 into 96i8 (and store)
5863 
5864       {3, MVT::v2i16, 4},   // interleave 3 x 2i16 into 6i16 (and store)
5865       {3, MVT::v4i16, 6},   // interleave 3 x 4i16 into 12i16 (and store)
5866       {3, MVT::v8i16, 12},  // interleave 3 x 8i16 into 24i16 (and store)
5867       {3, MVT::v16i16, 27}, // interleave 3 x 16i16 into 48i16 (and store)
5868       {3, MVT::v32i16, 54}, // interleave 3 x 32i16 into 96i16 (and store)
5869 
5870       {3, MVT::v2i32, 4},   // interleave 3 x 2i32 into 6i32 (and store)
5871       {3, MVT::v4i32, 5},   // interleave 3 x 4i32 into 12i32 (and store)
5872       {3, MVT::v8i32, 11},  // interleave 3 x 8i32 into 24i32 (and store)
5873       {3, MVT::v16i32, 22}, // interleave 3 x 16i32 into 48i32 (and store)
5874       {3, MVT::v32i32, 48}, // interleave 3 x 32i32 into 96i32 (and store)
5875 
5876       {3, MVT::v2i64, 4},   // interleave 3 x 2i64 into 6i64 (and store)
5877       {3, MVT::v4i64, 6},   // interleave 3 x 4i64 into 12i64 (and store)
5878       {3, MVT::v8i64, 12},  // interleave 3 x 8i64 into 24i64 (and store)
5879       {3, MVT::v16i64, 24}, // interleave 3 x 16i64 into 48i64 (and store)
5880 
5881       {4, MVT::v2i8, 4},   // interleave 4 x 2i8 into 8i8 (and store)
5882       {4, MVT::v4i8, 4},   // interleave 4 x 4i8 into 16i8 (and store)
5883       {4, MVT::v8i8, 4},   // interleave 4 x 8i8 into 32i8 (and store)
5884       {4, MVT::v16i8, 8},  // interleave 4 x 16i8 into 64i8 (and store)
5885       {4, MVT::v32i8, 12}, // interleave 4 x 32i8 into 128i8 (and store)
5886 
5887       {4, MVT::v2i16, 2},   // interleave 4 x 2i16 into 8i16 (and store)
5888       {4, MVT::v4i16, 6},   // interleave 4 x 4i16 into 16i16 (and store)
5889       {4, MVT::v8i16, 10},  // interleave 4 x 8i16 into 32i16 (and store)
5890       {4, MVT::v16i16, 32}, // interleave 4 x 16i16 into 64i16 (and store)
5891       {4, MVT::v32i16, 64}, // interleave 4 x 32i16 into 128i16 (and store)
5892 
5893       {4, MVT::v2i32, 5},   // interleave 4 x 2i32 into 8i32 (and store)
5894       {4, MVT::v4i32, 6},   // interleave 4 x 4i32 into 16i32 (and store)
5895       {4, MVT::v8i32, 16},  // interleave 4 x 8i32 into 32i32 (and store)
5896       {4, MVT::v16i32, 32}, // interleave 4 x 16i32 into 64i32 (and store)
5897       {4, MVT::v32i32, 64}, // interleave 4 x 32i32 into 128i32 (and store)
5898 
5899       {4, MVT::v2i64, 6},  // interleave 4 x 2i64 into 8i64 (and store)
5900       {4, MVT::v4i64, 8},  // interleave 4 x 4i64 into 16i64 (and store)
5901       {4, MVT::v8i64, 20}, // interleave 4 x 8i64 into 32i64 (and store)
5902       {4, MVT::v16i64, 40}, // interleave 4 x 16i64 into 64i64 (and store)
5903 
5904       {6, MVT::v2i8, 7},   // interleave 6 x 2i8 into 12i8 (and store)
5905       {6, MVT::v4i8, 9},   // interleave 6 x 4i8 into 24i8 (and store)
5906       {6, MVT::v8i8, 16},  // interleave 6 x 8i8 into 48i8 (and store)
5907       {6, MVT::v16i8, 27}, // interleave 6 x 16i8 into 96i8 (and store)
5908       {6, MVT::v32i8, 90}, // interleave 6 x 32i8 into 192i8 (and store)
5909 
5910       {6, MVT::v2i16, 10},  // interleave 6 x 2i16 into 12i16 (and store)
5911       {6, MVT::v4i16, 15},  // interleave 6 x 4i16 into 24i16 (and store)
5912       {6, MVT::v8i16, 21},  // interleave 6 x 8i16 into 48i16 (and store)
5913       {6, MVT::v16i16, 58}, // interleave 6 x 16i16 into 96i16 (and store)
5914       {6, MVT::v32i16, 90}, // interleave 6 x 32i16 into 192i16 (and store)
5915 
5916       {6, MVT::v2i32, 9},   // interleave 6 x 2i32 into 12i32 (and store)
5917       {6, MVT::v4i32, 12},  // interleave 6 x 4i32 into 24i32 (and store)
5918       {6, MVT::v8i32, 33},  // interleave 6 x 8i32 into 48i32 (and store)
5919       {6, MVT::v16i32, 66}, // interleave 6 x 16i32 into 96i32 (and store)
5920 
5921       {6, MVT::v2i64, 8},  // interleave 6 x 2i64 into 12i64 (and store)
5922       {6, MVT::v4i64, 15}, // interleave 6 x 4i64 into 24i64 (and store)
5923       {6, MVT::v8i64, 30}, // interleave 6 x 8i64 into 48i64 (and store)
5924   };
5925 
5926   static const CostTblEntry SSE2InterleavedStoreTbl[] = {
5927       {2, MVT::v2i8, 1},   // interleave 2 x 2i8 into 4i8 (and store)
5928       {2, MVT::v4i8, 1},   // interleave 2 x 4i8 into 8i8 (and store)
5929       {2, MVT::v8i8, 1},   // interleave 2 x 8i8 into 16i8 (and store)
5930 
5931       {2, MVT::v2i16, 1},  // interleave 2 x 2i16 into 4i16 (and store)
5932       {2, MVT::v4i16, 1},  // interleave 2 x 4i16 into 8i16 (and store)
5933 
5934       {2, MVT::v2i32, 1},  // interleave 2 x 2i32 into 4i32 (and store)
5935   };
5936 
5937   if (Opcode == Instruction::Load) {
5938     auto GetDiscountedCost = [Factor, NumMembers = Indices.size(),
5939                               MemOpCosts](const CostTblEntry *Entry) {
5940       // NOTE: this is just an approximation!
5941       //       It can over/under -estimate the cost!
5942       return MemOpCosts + divideCeil(NumMembers * Entry->Cost, Factor);
5943     };
5944 
5945     if (ST->hasAVX2())
5946       if (const auto *Entry = CostTableLookup(AVX2InterleavedLoadTbl, Factor,
5947                                               ETy.getSimpleVT()))
5948         return GetDiscountedCost(Entry);
5949 
5950     if (ST->hasSSSE3())
5951       if (const auto *Entry = CostTableLookup(SSSE3InterleavedLoadTbl, Factor,
5952                                               ETy.getSimpleVT()))
5953         return GetDiscountedCost(Entry);
5954 
5955     if (ST->hasSSE2())
5956       if (const auto *Entry = CostTableLookup(SSE2InterleavedLoadTbl, Factor,
5957                                               ETy.getSimpleVT()))
5958         return GetDiscountedCost(Entry);
5959   } else {
5960     assert(Opcode == Instruction::Store &&
5961            "Expected Store Instruction at this point");
5962     assert((!Indices.size() || Indices.size() == Factor) &&
5963            "Interleaved store only supports fully-interleaved groups.");
5964     if (ST->hasAVX2())
5965       if (const auto *Entry = CostTableLookup(AVX2InterleavedStoreTbl, Factor,
5966                                               ETy.getSimpleVT()))
5967         return MemOpCosts + Entry->Cost;
5968 
5969     if (ST->hasSSE2())
5970       if (const auto *Entry = CostTableLookup(SSE2InterleavedStoreTbl, Factor,
5971                                               ETy.getSimpleVT()))
5972         return MemOpCosts + Entry->Cost;
5973   }
5974 
5975   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5976                                            Alignment, AddressSpace, CostKind,
5977                                            UseMaskForCond, UseMaskForGaps);
5978 }
5979