xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86TargetTransformInfo.cpp (revision 62cfcf62f627e5093fb37026a6d8c98e4d2ef04c)
1 //===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 /// \file
9 /// This file implements a TargetTransformInfo analysis pass specific to the
10 /// X86 target machine. It uses the target's detailed information to provide
11 /// more precise answers to certain TTI queries, while letting the target
12 /// independent and default TTI implementations handle the rest.
13 ///
14 //===----------------------------------------------------------------------===//
15 /// About Cost Model numbers used below it's necessary to say the following:
16 /// the numbers correspond to some "generic" X86 CPU instead of usage of
17 /// concrete CPU model. Usually the numbers correspond to CPU where the feature
18 /// apeared at the first time. For example, if we do Subtarget.hasSSE42() in
19 /// the lookups below the cost is based on Nehalem as that was the first CPU
20 /// to support that feature level and thus has most likely the worst case cost.
21 /// Some examples of other technologies/CPUs:
22 ///   SSE 3   - Pentium4 / Athlon64
23 ///   SSE 4.1 - Penryn
24 ///   SSE 4.2 - Nehalem
25 ///   AVX     - Sandy Bridge
26 ///   AVX2    - Haswell
27 ///   AVX-512 - Xeon Phi / Skylake
28 /// And some examples of instruction target dependent costs (latency)
29 ///                   divss     sqrtss          rsqrtss
30 ///   AMD K7            11-16     19              3
31 ///   Piledriver        9-24      13-15           5
32 ///   Jaguar            14        16              2
33 ///   Pentium II,III    18        30              2
34 ///   Nehalem           7-14      7-18            3
35 ///   Haswell           10-13     11              5
36 /// TODO: Develop and implement  the target dependent cost model and
37 /// specialize cost numbers for different Cost Model Targets such as throughput,
38 /// code size, latency and uop count.
39 //===----------------------------------------------------------------------===//
40 
41 #include "X86TargetTransformInfo.h"
42 #include "llvm/Analysis/TargetTransformInfo.h"
43 #include "llvm/CodeGen/BasicTTIImpl.h"
44 #include "llvm/CodeGen/CostTable.h"
45 #include "llvm/CodeGen/TargetLowering.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/Support/Debug.h"
48 
49 using namespace llvm;
50 
51 #define DEBUG_TYPE "x86tti"
52 
53 //===----------------------------------------------------------------------===//
54 //
55 // X86 cost model.
56 //
57 //===----------------------------------------------------------------------===//
58 
59 TargetTransformInfo::PopcntSupportKind
60 X86TTIImpl::getPopcntSupport(unsigned TyWidth) {
61   assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
62   // TODO: Currently the __builtin_popcount() implementation using SSE3
63   //   instructions is inefficient. Once the problem is fixed, we should
64   //   call ST->hasSSE3() instead of ST->hasPOPCNT().
65   return ST->hasPOPCNT() ? TTI::PSK_FastHardware : TTI::PSK_Software;
66 }
67 
68 llvm::Optional<unsigned> X86TTIImpl::getCacheSize(
69   TargetTransformInfo::CacheLevel Level) const {
70   switch (Level) {
71   case TargetTransformInfo::CacheLevel::L1D:
72     //   - Penryn
73     //   - Nehalem
74     //   - Westmere
75     //   - Sandy Bridge
76     //   - Ivy Bridge
77     //   - Haswell
78     //   - Broadwell
79     //   - Skylake
80     //   - Kabylake
81     return 32 * 1024;  //  32 KByte
82   case TargetTransformInfo::CacheLevel::L2D:
83     //   - Penryn
84     //   - Nehalem
85     //   - Westmere
86     //   - Sandy Bridge
87     //   - Ivy Bridge
88     //   - Haswell
89     //   - Broadwell
90     //   - Skylake
91     //   - Kabylake
92     return 256 * 1024; // 256 KByte
93   }
94 
95   llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
96 }
97 
98 llvm::Optional<unsigned> X86TTIImpl::getCacheAssociativity(
99   TargetTransformInfo::CacheLevel Level) const {
100   //   - Penryn
101   //   - Nehalem
102   //   - Westmere
103   //   - Sandy Bridge
104   //   - Ivy Bridge
105   //   - Haswell
106   //   - Broadwell
107   //   - Skylake
108   //   - Kabylake
109   switch (Level) {
110   case TargetTransformInfo::CacheLevel::L1D:
111     LLVM_FALLTHROUGH;
112   case TargetTransformInfo::CacheLevel::L2D:
113     return 8;
114   }
115 
116   llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
117 }
118 
119 unsigned X86TTIImpl::getNumberOfRegisters(unsigned ClassID) const {
120   bool Vector = (ClassID == 1);
121   if (Vector && !ST->hasSSE1())
122     return 0;
123 
124   if (ST->is64Bit()) {
125     if (Vector && ST->hasAVX512())
126       return 32;
127     return 16;
128   }
129   return 8;
130 }
131 
132 unsigned X86TTIImpl::getRegisterBitWidth(bool Vector) const {
133   unsigned PreferVectorWidth = ST->getPreferVectorWidth();
134   if (Vector) {
135     if (ST->hasAVX512() && PreferVectorWidth >= 512)
136       return 512;
137     if (ST->hasAVX() && PreferVectorWidth >= 256)
138       return 256;
139     if (ST->hasSSE1() && PreferVectorWidth >= 128)
140       return 128;
141     return 0;
142   }
143 
144   if (ST->is64Bit())
145     return 64;
146 
147   return 32;
148 }
149 
150 unsigned X86TTIImpl::getLoadStoreVecRegBitWidth(unsigned) const {
151   return getRegisterBitWidth(true);
152 }
153 
154 unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) {
155   // If the loop will not be vectorized, don't interleave the loop.
156   // Let regular unroll to unroll the loop, which saves the overflow
157   // check and memory check cost.
158   if (VF == 1)
159     return 1;
160 
161   if (ST->isAtom())
162     return 1;
163 
164   // Sandybridge and Haswell have multiple execution ports and pipelined
165   // vector units.
166   if (ST->hasAVX())
167     return 4;
168 
169   return 2;
170 }
171 
172 int X86TTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
173                                        TTI::OperandValueKind Op1Info,
174                                        TTI::OperandValueKind Op2Info,
175                                        TTI::OperandValueProperties Opd1PropInfo,
176                                        TTI::OperandValueProperties Opd2PropInfo,
177                                        ArrayRef<const Value *> Args,
178                                        const Instruction *CxtI) {
179   // Legalize the type.
180   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
181 
182   int ISD = TLI->InstructionOpcodeToISD(Opcode);
183   assert(ISD && "Invalid opcode");
184 
185   static const CostTblEntry GLMCostTable[] = {
186     { ISD::FDIV,  MVT::f32,   18 }, // divss
187     { ISD::FDIV,  MVT::v4f32, 35 }, // divps
188     { ISD::FDIV,  MVT::f64,   33 }, // divsd
189     { ISD::FDIV,  MVT::v2f64, 65 }, // divpd
190   };
191 
192   if (ST->useGLMDivSqrtCosts())
193     if (const auto *Entry = CostTableLookup(GLMCostTable, ISD,
194                                             LT.second))
195       return LT.first * Entry->Cost;
196 
197   static const CostTblEntry SLMCostTable[] = {
198     { ISD::MUL,   MVT::v4i32, 11 }, // pmulld
199     { ISD::MUL,   MVT::v8i16, 2  }, // pmullw
200     { ISD::MUL,   MVT::v16i8, 14 }, // extend/pmullw/trunc sequence.
201     { ISD::FMUL,  MVT::f64,   2  }, // mulsd
202     { ISD::FMUL,  MVT::v2f64, 4  }, // mulpd
203     { ISD::FMUL,  MVT::v4f32, 2  }, // mulps
204     { ISD::FDIV,  MVT::f32,   17 }, // divss
205     { ISD::FDIV,  MVT::v4f32, 39 }, // divps
206     { ISD::FDIV,  MVT::f64,   32 }, // divsd
207     { ISD::FDIV,  MVT::v2f64, 69 }, // divpd
208     { ISD::FADD,  MVT::v2f64, 2  }, // addpd
209     { ISD::FSUB,  MVT::v2f64, 2  }, // subpd
210     // v2i64/v4i64 mul is custom lowered as a series of long:
211     // multiplies(3), shifts(3) and adds(2)
212     // slm muldq version throughput is 2 and addq throughput 4
213     // thus: 3X2 (muldq throughput) + 3X1 (shift throughput) +
214     //       3X4 (addq throughput) = 17
215     { ISD::MUL,   MVT::v2i64, 17 },
216     // slm addq\subq throughput is 4
217     { ISD::ADD,   MVT::v2i64, 4  },
218     { ISD::SUB,   MVT::v2i64, 4  },
219   };
220 
221   if (ST->isSLM()) {
222     if (Args.size() == 2 && ISD == ISD::MUL && LT.second == MVT::v4i32) {
223       // Check if the operands can be shrinked into a smaller datatype.
224       bool Op1Signed = false;
225       unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed);
226       bool Op2Signed = false;
227       unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed);
228 
229       bool signedMode = Op1Signed | Op2Signed;
230       unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize);
231 
232       if (OpMinSize <= 7)
233         return LT.first * 3; // pmullw/sext
234       if (!signedMode && OpMinSize <= 8)
235         return LT.first * 3; // pmullw/zext
236       if (OpMinSize <= 15)
237         return LT.first * 5; // pmullw/pmulhw/pshuf
238       if (!signedMode && OpMinSize <= 16)
239         return LT.first * 5; // pmullw/pmulhw/pshuf
240     }
241 
242     if (const auto *Entry = CostTableLookup(SLMCostTable, ISD,
243                                             LT.second)) {
244       return LT.first * Entry->Cost;
245     }
246   }
247 
248   if ((ISD == ISD::SDIV || ISD == ISD::SREM || ISD == ISD::UDIV ||
249        ISD == ISD::UREM) &&
250       (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
251        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
252       Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
253     if (ISD == ISD::SDIV || ISD == ISD::SREM) {
254       // On X86, vector signed division by constants power-of-two are
255       // normally expanded to the sequence SRA + SRL + ADD + SRA.
256       // The OperandValue properties may not be the same as that of the previous
257       // operation; conservatively assume OP_None.
258       int Cost =
259           2 * getArithmeticInstrCost(Instruction::AShr, Ty, Op1Info, Op2Info,
260                                      TargetTransformInfo::OP_None,
261                                      TargetTransformInfo::OP_None);
262       Cost += getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info,
263                                      TargetTransformInfo::OP_None,
264                                      TargetTransformInfo::OP_None);
265       Cost += getArithmeticInstrCost(Instruction::Add, Ty, Op1Info, Op2Info,
266                                      TargetTransformInfo::OP_None,
267                                      TargetTransformInfo::OP_None);
268 
269       if (ISD == ISD::SREM) {
270         // For SREM: (X % C) is the equivalent of (X - (X/C)*C)
271         Cost += getArithmeticInstrCost(Instruction::Mul, Ty, Op1Info, Op2Info);
272         Cost += getArithmeticInstrCost(Instruction::Sub, Ty, Op1Info, Op2Info);
273       }
274 
275       return Cost;
276     }
277 
278     // Vector unsigned division/remainder will be simplified to shifts/masks.
279     if (ISD == ISD::UDIV)
280       return getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info,
281                                     TargetTransformInfo::OP_None,
282                                     TargetTransformInfo::OP_None);
283 
284     else // UREM
285       return getArithmeticInstrCost(Instruction::And, Ty, Op1Info, Op2Info,
286                                     TargetTransformInfo::OP_None,
287                                     TargetTransformInfo::OP_None);
288   }
289 
290   static const CostTblEntry AVX512BWUniformConstCostTable[] = {
291     { ISD::SHL,  MVT::v64i8,   2 }, // psllw + pand.
292     { ISD::SRL,  MVT::v64i8,   2 }, // psrlw + pand.
293     { ISD::SRA,  MVT::v64i8,   4 }, // psrlw, pand, pxor, psubb.
294   };
295 
296   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
297       ST->hasBWI()) {
298     if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD,
299                                             LT.second))
300       return LT.first * Entry->Cost;
301   }
302 
303   static const CostTblEntry AVX512UniformConstCostTable[] = {
304     { ISD::SRA,  MVT::v2i64,   1 },
305     { ISD::SRA,  MVT::v4i64,   1 },
306     { ISD::SRA,  MVT::v8i64,   1 },
307   };
308 
309   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
310       ST->hasAVX512()) {
311     if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD,
312                                             LT.second))
313       return LT.first * Entry->Cost;
314   }
315 
316   static const CostTblEntry AVX2UniformConstCostTable[] = {
317     { ISD::SHL,  MVT::v32i8,   2 }, // psllw + pand.
318     { ISD::SRL,  MVT::v32i8,   2 }, // psrlw + pand.
319     { ISD::SRA,  MVT::v32i8,   4 }, // psrlw, pand, pxor, psubb.
320 
321     { ISD::SRA,  MVT::v4i64,   4 }, // 2 x psrad + shuffle.
322   };
323 
324   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
325       ST->hasAVX2()) {
326     if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD,
327                                             LT.second))
328       return LT.first * Entry->Cost;
329   }
330 
331   static const CostTblEntry SSE2UniformConstCostTable[] = {
332     { ISD::SHL,  MVT::v16i8,     2 }, // psllw + pand.
333     { ISD::SRL,  MVT::v16i8,     2 }, // psrlw + pand.
334     { ISD::SRA,  MVT::v16i8,     4 }, // psrlw, pand, pxor, psubb.
335 
336     { ISD::SHL,  MVT::v32i8,   4+2 }, // 2*(psllw + pand) + split.
337     { ISD::SRL,  MVT::v32i8,   4+2 }, // 2*(psrlw + pand) + split.
338     { ISD::SRA,  MVT::v32i8,   8+2 }, // 2*(psrlw, pand, pxor, psubb) + split.
339   };
340 
341   // XOP has faster vXi8 shifts.
342   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
343       ST->hasSSE2() && !ST->hasXOP()) {
344     if (const auto *Entry =
345             CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second))
346       return LT.first * Entry->Cost;
347   }
348 
349   static const CostTblEntry AVX512BWConstCostTable[] = {
350     { ISD::SDIV, MVT::v64i8,  14 }, // 2*ext+2*pmulhw sequence
351     { ISD::SREM, MVT::v64i8,  16 }, // 2*ext+2*pmulhw+mul+sub sequence
352     { ISD::UDIV, MVT::v64i8,  14 }, // 2*ext+2*pmulhw sequence
353     { ISD::UREM, MVT::v64i8,  16 }, // 2*ext+2*pmulhw+mul+sub sequence
354     { ISD::SDIV, MVT::v32i16,  6 }, // vpmulhw sequence
355     { ISD::SREM, MVT::v32i16,  8 }, // vpmulhw+mul+sub sequence
356     { ISD::UDIV, MVT::v32i16,  6 }, // vpmulhuw sequence
357     { ISD::UREM, MVT::v32i16,  8 }, // vpmulhuw+mul+sub sequence
358   };
359 
360   if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
361        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
362       ST->hasBWI()) {
363     if (const auto *Entry =
364             CostTableLookup(AVX512BWConstCostTable, ISD, LT.second))
365       return LT.first * Entry->Cost;
366   }
367 
368   static const CostTblEntry AVX512ConstCostTable[] = {
369     { ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence
370     { ISD::SREM, MVT::v16i32, 17 }, // vpmuldq+mul+sub sequence
371     { ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence
372     { ISD::UREM, MVT::v16i32, 17 }, // vpmuludq+mul+sub sequence
373   };
374 
375   if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
376        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
377       ST->hasAVX512()) {
378     if (const auto *Entry =
379             CostTableLookup(AVX512ConstCostTable, ISD, LT.second))
380       return LT.first * Entry->Cost;
381   }
382 
383   static const CostTblEntry AVX2ConstCostTable[] = {
384     { ISD::SDIV, MVT::v32i8,  14 }, // 2*ext+2*pmulhw sequence
385     { ISD::SREM, MVT::v32i8,  16 }, // 2*ext+2*pmulhw+mul+sub sequence
386     { ISD::UDIV, MVT::v32i8,  14 }, // 2*ext+2*pmulhw sequence
387     { ISD::UREM, MVT::v32i8,  16 }, // 2*ext+2*pmulhw+mul+sub sequence
388     { ISD::SDIV, MVT::v16i16,  6 }, // vpmulhw sequence
389     { ISD::SREM, MVT::v16i16,  8 }, // vpmulhw+mul+sub sequence
390     { ISD::UDIV, MVT::v16i16,  6 }, // vpmulhuw sequence
391     { ISD::UREM, MVT::v16i16,  8 }, // vpmulhuw+mul+sub sequence
392     { ISD::SDIV, MVT::v8i32,  15 }, // vpmuldq sequence
393     { ISD::SREM, MVT::v8i32,  19 }, // vpmuldq+mul+sub sequence
394     { ISD::UDIV, MVT::v8i32,  15 }, // vpmuludq sequence
395     { ISD::UREM, MVT::v8i32,  19 }, // vpmuludq+mul+sub sequence
396   };
397 
398   if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
399        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
400       ST->hasAVX2()) {
401     if (const auto *Entry = CostTableLookup(AVX2ConstCostTable, ISD, LT.second))
402       return LT.first * Entry->Cost;
403   }
404 
405   static const CostTblEntry SSE2ConstCostTable[] = {
406     { ISD::SDIV, MVT::v32i8,  28+2 }, // 4*ext+4*pmulhw sequence + split.
407     { ISD::SREM, MVT::v32i8,  32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split.
408     { ISD::SDIV, MVT::v16i8,    14 }, // 2*ext+2*pmulhw sequence
409     { ISD::SREM, MVT::v16i8,    16 }, // 2*ext+2*pmulhw+mul+sub sequence
410     { ISD::UDIV, MVT::v32i8,  28+2 }, // 4*ext+4*pmulhw sequence + split.
411     { ISD::UREM, MVT::v32i8,  32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split.
412     { ISD::UDIV, MVT::v16i8,    14 }, // 2*ext+2*pmulhw sequence
413     { ISD::UREM, MVT::v16i8,    16 }, // 2*ext+2*pmulhw+mul+sub sequence
414     { ISD::SDIV, MVT::v16i16, 12+2 }, // 2*pmulhw sequence + split.
415     { ISD::SREM, MVT::v16i16, 16+2 }, // 2*pmulhw+mul+sub sequence + split.
416     { ISD::SDIV, MVT::v8i16,     6 }, // pmulhw sequence
417     { ISD::SREM, MVT::v8i16,     8 }, // pmulhw+mul+sub sequence
418     { ISD::UDIV, MVT::v16i16, 12+2 }, // 2*pmulhuw sequence + split.
419     { ISD::UREM, MVT::v16i16, 16+2 }, // 2*pmulhuw+mul+sub sequence + split.
420     { ISD::UDIV, MVT::v8i16,     6 }, // pmulhuw sequence
421     { ISD::UREM, MVT::v8i16,     8 }, // pmulhuw+mul+sub sequence
422     { ISD::SDIV, MVT::v8i32,  38+2 }, // 2*pmuludq sequence + split.
423     { ISD::SREM, MVT::v8i32,  48+2 }, // 2*pmuludq+mul+sub sequence + split.
424     { ISD::SDIV, MVT::v4i32,    19 }, // pmuludq sequence
425     { ISD::SREM, MVT::v4i32,    24 }, // pmuludq+mul+sub sequence
426     { ISD::UDIV, MVT::v8i32,  30+2 }, // 2*pmuludq sequence + split.
427     { ISD::UREM, MVT::v8i32,  40+2 }, // 2*pmuludq+mul+sub sequence + split.
428     { ISD::UDIV, MVT::v4i32,    15 }, // pmuludq sequence
429     { ISD::UREM, MVT::v4i32,    20 }, // pmuludq+mul+sub sequence
430   };
431 
432   if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
433        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
434       ST->hasSSE2()) {
435     // pmuldq sequence.
436     if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX())
437       return LT.first * 32;
438     if (ISD == ISD::SREM && LT.second == MVT::v8i32 && ST->hasAVX())
439       return LT.first * 38;
440     if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
441       return LT.first * 15;
442     if (ISD == ISD::SREM && LT.second == MVT::v4i32 && ST->hasSSE41())
443       return LT.first * 20;
444 
445     if (const auto *Entry = CostTableLookup(SSE2ConstCostTable, ISD, LT.second))
446       return LT.first * Entry->Cost;
447   }
448 
449   static const CostTblEntry AVX2UniformCostTable[] = {
450     // Uniform splats are cheaper for the following instructions.
451     { ISD::SHL,  MVT::v16i16, 1 }, // psllw.
452     { ISD::SRL,  MVT::v16i16, 1 }, // psrlw.
453     { ISD::SRA,  MVT::v16i16, 1 }, // psraw.
454   };
455 
456   if (ST->hasAVX2() &&
457       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
458        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
459     if (const auto *Entry =
460             CostTableLookup(AVX2UniformCostTable, ISD, LT.second))
461       return LT.first * Entry->Cost;
462   }
463 
464   static const CostTblEntry SSE2UniformCostTable[] = {
465     // Uniform splats are cheaper for the following instructions.
466     { ISD::SHL,  MVT::v8i16,  1 }, // psllw.
467     { ISD::SHL,  MVT::v4i32,  1 }, // pslld
468     { ISD::SHL,  MVT::v2i64,  1 }, // psllq.
469 
470     { ISD::SRL,  MVT::v8i16,  1 }, // psrlw.
471     { ISD::SRL,  MVT::v4i32,  1 }, // psrld.
472     { ISD::SRL,  MVT::v2i64,  1 }, // psrlq.
473 
474     { ISD::SRA,  MVT::v8i16,  1 }, // psraw.
475     { ISD::SRA,  MVT::v4i32,  1 }, // psrad.
476   };
477 
478   if (ST->hasSSE2() &&
479       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
480        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
481     if (const auto *Entry =
482             CostTableLookup(SSE2UniformCostTable, ISD, LT.second))
483       return LT.first * Entry->Cost;
484   }
485 
486   static const CostTblEntry AVX512DQCostTable[] = {
487     { ISD::MUL,  MVT::v2i64, 1 },
488     { ISD::MUL,  MVT::v4i64, 1 },
489     { ISD::MUL,  MVT::v8i64, 1 }
490   };
491 
492   // Look for AVX512DQ lowering tricks for custom cases.
493   if (ST->hasDQI())
494     if (const auto *Entry = CostTableLookup(AVX512DQCostTable, ISD, LT.second))
495       return LT.first * Entry->Cost;
496 
497   static const CostTblEntry AVX512BWCostTable[] = {
498     { ISD::SHL,   MVT::v8i16,      1 }, // vpsllvw
499     { ISD::SRL,   MVT::v8i16,      1 }, // vpsrlvw
500     { ISD::SRA,   MVT::v8i16,      1 }, // vpsravw
501 
502     { ISD::SHL,   MVT::v16i16,     1 }, // vpsllvw
503     { ISD::SRL,   MVT::v16i16,     1 }, // vpsrlvw
504     { ISD::SRA,   MVT::v16i16,     1 }, // vpsravw
505 
506     { ISD::SHL,   MVT::v32i16,     1 }, // vpsllvw
507     { ISD::SRL,   MVT::v32i16,     1 }, // vpsrlvw
508     { ISD::SRA,   MVT::v32i16,     1 }, // vpsravw
509 
510     { ISD::SHL,   MVT::v64i8,     11 }, // vpblendvb sequence.
511     { ISD::SRL,   MVT::v64i8,     11 }, // vpblendvb sequence.
512     { ISD::SRA,   MVT::v64i8,     24 }, // vpblendvb sequence.
513 
514     { ISD::MUL,   MVT::v64i8,     11 }, // extend/pmullw/trunc sequence.
515     { ISD::MUL,   MVT::v32i8,      4 }, // extend/pmullw/trunc sequence.
516     { ISD::MUL,   MVT::v16i8,      4 }, // extend/pmullw/trunc sequence.
517   };
518 
519   // Look for AVX512BW lowering tricks for custom cases.
520   if (ST->hasBWI())
521     if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD, LT.second))
522       return LT.first * Entry->Cost;
523 
524   static const CostTblEntry AVX512CostTable[] = {
525     { ISD::SHL,     MVT::v16i32,     1 },
526     { ISD::SRL,     MVT::v16i32,     1 },
527     { ISD::SRA,     MVT::v16i32,     1 },
528 
529     { ISD::SHL,     MVT::v8i64,      1 },
530     { ISD::SRL,     MVT::v8i64,      1 },
531 
532     { ISD::SRA,     MVT::v2i64,      1 },
533     { ISD::SRA,     MVT::v4i64,      1 },
534     { ISD::SRA,     MVT::v8i64,      1 },
535 
536     { ISD::MUL,     MVT::v32i8,     13 }, // extend/pmullw/trunc sequence.
537     { ISD::MUL,     MVT::v16i8,      5 }, // extend/pmullw/trunc sequence.
538     { ISD::MUL,     MVT::v16i32,     1 }, // pmulld (Skylake from agner.org)
539     { ISD::MUL,     MVT::v8i32,      1 }, // pmulld (Skylake from agner.org)
540     { ISD::MUL,     MVT::v4i32,      1 }, // pmulld (Skylake from agner.org)
541     { ISD::MUL,     MVT::v8i64,      8 }, // 3*pmuludq/3*shift/2*add
542 
543     { ISD::FADD,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
544     { ISD::FSUB,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
545     { ISD::FMUL,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
546 
547     { ISD::FADD,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
548     { ISD::FSUB,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
549     { ISD::FMUL,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
550   };
551 
552   if (ST->hasAVX512())
553     if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second))
554       return LT.first * Entry->Cost;
555 
556   static const CostTblEntry AVX2ShiftCostTable[] = {
557     // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
558     // customize them to detect the cases where shift amount is a scalar one.
559     { ISD::SHL,     MVT::v4i32,    1 },
560     { ISD::SRL,     MVT::v4i32,    1 },
561     { ISD::SRA,     MVT::v4i32,    1 },
562     { ISD::SHL,     MVT::v8i32,    1 },
563     { ISD::SRL,     MVT::v8i32,    1 },
564     { ISD::SRA,     MVT::v8i32,    1 },
565     { ISD::SHL,     MVT::v2i64,    1 },
566     { ISD::SRL,     MVT::v2i64,    1 },
567     { ISD::SHL,     MVT::v4i64,    1 },
568     { ISD::SRL,     MVT::v4i64,    1 },
569   };
570 
571   // Look for AVX2 lowering tricks.
572   if (ST->hasAVX2()) {
573     if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
574         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
575          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
576       // On AVX2, a packed v16i16 shift left by a constant build_vector
577       // is lowered into a vector multiply (vpmullw).
578       return getArithmeticInstrCost(Instruction::Mul, Ty, Op1Info, Op2Info,
579                                     TargetTransformInfo::OP_None,
580                                     TargetTransformInfo::OP_None);
581 
582     if (const auto *Entry = CostTableLookup(AVX2ShiftCostTable, ISD, LT.second))
583       return LT.first * Entry->Cost;
584   }
585 
586   static const CostTblEntry XOPShiftCostTable[] = {
587     // 128bit shifts take 1cy, but right shifts require negation beforehand.
588     { ISD::SHL,     MVT::v16i8,    1 },
589     { ISD::SRL,     MVT::v16i8,    2 },
590     { ISD::SRA,     MVT::v16i8,    2 },
591     { ISD::SHL,     MVT::v8i16,    1 },
592     { ISD::SRL,     MVT::v8i16,    2 },
593     { ISD::SRA,     MVT::v8i16,    2 },
594     { ISD::SHL,     MVT::v4i32,    1 },
595     { ISD::SRL,     MVT::v4i32,    2 },
596     { ISD::SRA,     MVT::v4i32,    2 },
597     { ISD::SHL,     MVT::v2i64,    1 },
598     { ISD::SRL,     MVT::v2i64,    2 },
599     { ISD::SRA,     MVT::v2i64,    2 },
600     // 256bit shifts require splitting if AVX2 didn't catch them above.
601     { ISD::SHL,     MVT::v32i8,  2+2 },
602     { ISD::SRL,     MVT::v32i8,  4+2 },
603     { ISD::SRA,     MVT::v32i8,  4+2 },
604     { ISD::SHL,     MVT::v16i16, 2+2 },
605     { ISD::SRL,     MVT::v16i16, 4+2 },
606     { ISD::SRA,     MVT::v16i16, 4+2 },
607     { ISD::SHL,     MVT::v8i32,  2+2 },
608     { ISD::SRL,     MVT::v8i32,  4+2 },
609     { ISD::SRA,     MVT::v8i32,  4+2 },
610     { ISD::SHL,     MVT::v4i64,  2+2 },
611     { ISD::SRL,     MVT::v4i64,  4+2 },
612     { ISD::SRA,     MVT::v4i64,  4+2 },
613   };
614 
615   // Look for XOP lowering tricks.
616   if (ST->hasXOP()) {
617     // If the right shift is constant then we'll fold the negation so
618     // it's as cheap as a left shift.
619     int ShiftISD = ISD;
620     if ((ShiftISD == ISD::SRL || ShiftISD == ISD::SRA) &&
621         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
622          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
623       ShiftISD = ISD::SHL;
624     if (const auto *Entry =
625             CostTableLookup(XOPShiftCostTable, ShiftISD, LT.second))
626       return LT.first * Entry->Cost;
627   }
628 
629   static const CostTblEntry SSE2UniformShiftCostTable[] = {
630     // Uniform splats are cheaper for the following instructions.
631     { ISD::SHL,  MVT::v16i16, 2+2 }, // 2*psllw + split.
632     { ISD::SHL,  MVT::v8i32,  2+2 }, // 2*pslld + split.
633     { ISD::SHL,  MVT::v4i64,  2+2 }, // 2*psllq + split.
634 
635     { ISD::SRL,  MVT::v16i16, 2+2 }, // 2*psrlw + split.
636     { ISD::SRL,  MVT::v8i32,  2+2 }, // 2*psrld + split.
637     { ISD::SRL,  MVT::v4i64,  2+2 }, // 2*psrlq + split.
638 
639     { ISD::SRA,  MVT::v16i16, 2+2 }, // 2*psraw + split.
640     { ISD::SRA,  MVT::v8i32,  2+2 }, // 2*psrad + split.
641     { ISD::SRA,  MVT::v2i64,    4 }, // 2*psrad + shuffle.
642     { ISD::SRA,  MVT::v4i64,  8+2 }, // 2*(2*psrad + shuffle) + split.
643   };
644 
645   if (ST->hasSSE2() &&
646       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
647        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
648 
649     // Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table.
650     if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2())
651       return LT.first * 4; // 2*psrad + shuffle.
652 
653     if (const auto *Entry =
654             CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second))
655       return LT.first * Entry->Cost;
656   }
657 
658   if (ISD == ISD::SHL &&
659       Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
660     MVT VT = LT.second;
661     // Vector shift left by non uniform constant can be lowered
662     // into vector multiply.
663     if (((VT == MVT::v8i16 || VT == MVT::v4i32) && ST->hasSSE2()) ||
664         ((VT == MVT::v16i16 || VT == MVT::v8i32) && ST->hasAVX()))
665       ISD = ISD::MUL;
666   }
667 
668   static const CostTblEntry AVX2CostTable[] = {
669     { ISD::SHL,  MVT::v32i8,     11 }, // vpblendvb sequence.
670     { ISD::SHL,  MVT::v16i16,    10 }, // extend/vpsrlvd/pack sequence.
671 
672     { ISD::SRL,  MVT::v32i8,     11 }, // vpblendvb sequence.
673     { ISD::SRL,  MVT::v16i16,    10 }, // extend/vpsrlvd/pack sequence.
674 
675     { ISD::SRA,  MVT::v32i8,     24 }, // vpblendvb sequence.
676     { ISD::SRA,  MVT::v16i16,    10 }, // extend/vpsravd/pack sequence.
677     { ISD::SRA,  MVT::v2i64,      4 }, // srl/xor/sub sequence.
678     { ISD::SRA,  MVT::v4i64,      4 }, // srl/xor/sub sequence.
679 
680     { ISD::SUB,  MVT::v32i8,      1 }, // psubb
681     { ISD::ADD,  MVT::v32i8,      1 }, // paddb
682     { ISD::SUB,  MVT::v16i16,     1 }, // psubw
683     { ISD::ADD,  MVT::v16i16,     1 }, // paddw
684     { ISD::SUB,  MVT::v8i32,      1 }, // psubd
685     { ISD::ADD,  MVT::v8i32,      1 }, // paddd
686     { ISD::SUB,  MVT::v4i64,      1 }, // psubq
687     { ISD::ADD,  MVT::v4i64,      1 }, // paddq
688 
689     { ISD::MUL,  MVT::v32i8,     17 }, // extend/pmullw/trunc sequence.
690     { ISD::MUL,  MVT::v16i8,      7 }, // extend/pmullw/trunc sequence.
691     { ISD::MUL,  MVT::v16i16,     1 }, // pmullw
692     { ISD::MUL,  MVT::v8i32,      2 }, // pmulld (Haswell from agner.org)
693     { ISD::MUL,  MVT::v4i64,      8 }, // 3*pmuludq/3*shift/2*add
694 
695     { ISD::FADD, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
696     { ISD::FADD, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
697     { ISD::FSUB, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
698     { ISD::FSUB, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
699     { ISD::FMUL, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
700     { ISD::FMUL, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
701 
702     { ISD::FDIV, MVT::f32,        7 }, // Haswell from http://www.agner.org/
703     { ISD::FDIV, MVT::v4f32,      7 }, // Haswell from http://www.agner.org/
704     { ISD::FDIV, MVT::v8f32,     14 }, // Haswell from http://www.agner.org/
705     { ISD::FDIV, MVT::f64,       14 }, // Haswell from http://www.agner.org/
706     { ISD::FDIV, MVT::v2f64,     14 }, // Haswell from http://www.agner.org/
707     { ISD::FDIV, MVT::v4f64,     28 }, // Haswell from http://www.agner.org/
708   };
709 
710   // Look for AVX2 lowering tricks for custom cases.
711   if (ST->hasAVX2())
712     if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second))
713       return LT.first * Entry->Cost;
714 
715   static const CostTblEntry AVX1CostTable[] = {
716     // We don't have to scalarize unsupported ops. We can issue two half-sized
717     // operations and we only need to extract the upper YMM half.
718     // Two ops + 1 extract + 1 insert = 4.
719     { ISD::MUL,     MVT::v16i16,     4 },
720     { ISD::MUL,     MVT::v8i32,      4 },
721     { ISD::SUB,     MVT::v32i8,      4 },
722     { ISD::ADD,     MVT::v32i8,      4 },
723     { ISD::SUB,     MVT::v16i16,     4 },
724     { ISD::ADD,     MVT::v16i16,     4 },
725     { ISD::SUB,     MVT::v8i32,      4 },
726     { ISD::ADD,     MVT::v8i32,      4 },
727     { ISD::SUB,     MVT::v4i64,      4 },
728     { ISD::ADD,     MVT::v4i64,      4 },
729 
730     // A v4i64 multiply is custom lowered as two split v2i64 vectors that then
731     // are lowered as a series of long multiplies(3), shifts(3) and adds(2)
732     // Because we believe v4i64 to be a legal type, we must also include the
733     // extract+insert in the cost table. Therefore, the cost here is 18
734     // instead of 8.
735     { ISD::MUL,     MVT::v4i64,     18 },
736 
737     { ISD::MUL,     MVT::v32i8,     26 }, // extend/pmullw/trunc sequence.
738 
739     { ISD::FDIV,    MVT::f32,       14 }, // SNB from http://www.agner.org/
740     { ISD::FDIV,    MVT::v4f32,     14 }, // SNB from http://www.agner.org/
741     { ISD::FDIV,    MVT::v8f32,     28 }, // SNB from http://www.agner.org/
742     { ISD::FDIV,    MVT::f64,       22 }, // SNB from http://www.agner.org/
743     { ISD::FDIV,    MVT::v2f64,     22 }, // SNB from http://www.agner.org/
744     { ISD::FDIV,    MVT::v4f64,     44 }, // SNB from http://www.agner.org/
745   };
746 
747   if (ST->hasAVX())
748     if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, LT.second))
749       return LT.first * Entry->Cost;
750 
751   static const CostTblEntry SSE42CostTable[] = {
752     { ISD::FADD, MVT::f64,     1 }, // Nehalem from http://www.agner.org/
753     { ISD::FADD, MVT::f32,     1 }, // Nehalem from http://www.agner.org/
754     { ISD::FADD, MVT::v2f64,   1 }, // Nehalem from http://www.agner.org/
755     { ISD::FADD, MVT::v4f32,   1 }, // Nehalem from http://www.agner.org/
756 
757     { ISD::FSUB, MVT::f64,     1 }, // Nehalem from http://www.agner.org/
758     { ISD::FSUB, MVT::f32 ,    1 }, // Nehalem from http://www.agner.org/
759     { ISD::FSUB, MVT::v2f64,   1 }, // Nehalem from http://www.agner.org/
760     { ISD::FSUB, MVT::v4f32,   1 }, // Nehalem from http://www.agner.org/
761 
762     { ISD::FMUL, MVT::f64,     1 }, // Nehalem from http://www.agner.org/
763     { ISD::FMUL, MVT::f32,     1 }, // Nehalem from http://www.agner.org/
764     { ISD::FMUL, MVT::v2f64,   1 }, // Nehalem from http://www.agner.org/
765     { ISD::FMUL, MVT::v4f32,   1 }, // Nehalem from http://www.agner.org/
766 
767     { ISD::FDIV,  MVT::f32,   14 }, // Nehalem from http://www.agner.org/
768     { ISD::FDIV,  MVT::v4f32, 14 }, // Nehalem from http://www.agner.org/
769     { ISD::FDIV,  MVT::f64,   22 }, // Nehalem from http://www.agner.org/
770     { ISD::FDIV,  MVT::v2f64, 22 }, // Nehalem from http://www.agner.org/
771   };
772 
773   if (ST->hasSSE42())
774     if (const auto *Entry = CostTableLookup(SSE42CostTable, ISD, LT.second))
775       return LT.first * Entry->Cost;
776 
777   static const CostTblEntry SSE41CostTable[] = {
778     { ISD::SHL,  MVT::v16i8,      11 }, // pblendvb sequence.
779     { ISD::SHL,  MVT::v32i8,  2*11+2 }, // pblendvb sequence + split.
780     { ISD::SHL,  MVT::v8i16,      14 }, // pblendvb sequence.
781     { ISD::SHL,  MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
782     { ISD::SHL,  MVT::v4i32,       4 }, // pslld/paddd/cvttps2dq/pmulld
783     { ISD::SHL,  MVT::v8i32,   2*4+2 }, // pslld/paddd/cvttps2dq/pmulld + split
784 
785     { ISD::SRL,  MVT::v16i8,      12 }, // pblendvb sequence.
786     { ISD::SRL,  MVT::v32i8,  2*12+2 }, // pblendvb sequence + split.
787     { ISD::SRL,  MVT::v8i16,      14 }, // pblendvb sequence.
788     { ISD::SRL,  MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
789     { ISD::SRL,  MVT::v4i32,      11 }, // Shift each lane + blend.
790     { ISD::SRL,  MVT::v8i32,  2*11+2 }, // Shift each lane + blend + split.
791 
792     { ISD::SRA,  MVT::v16i8,      24 }, // pblendvb sequence.
793     { ISD::SRA,  MVT::v32i8,  2*24+2 }, // pblendvb sequence + split.
794     { ISD::SRA,  MVT::v8i16,      14 }, // pblendvb sequence.
795     { ISD::SRA,  MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
796     { ISD::SRA,  MVT::v4i32,      12 }, // Shift each lane + blend.
797     { ISD::SRA,  MVT::v8i32,  2*12+2 }, // Shift each lane + blend + split.
798 
799     { ISD::MUL,  MVT::v4i32,       2 }  // pmulld (Nehalem from agner.org)
800   };
801 
802   if (ST->hasSSE41())
803     if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second))
804       return LT.first * Entry->Cost;
805 
806   static const CostTblEntry SSE2CostTable[] = {
807     // We don't correctly identify costs of casts because they are marked as
808     // custom.
809     { ISD::SHL,  MVT::v16i8,      26 }, // cmpgtb sequence.
810     { ISD::SHL,  MVT::v8i16,      32 }, // cmpgtb sequence.
811     { ISD::SHL,  MVT::v4i32,     2*5 }, // We optimized this using mul.
812     { ISD::SHL,  MVT::v2i64,       4 }, // splat+shuffle sequence.
813     { ISD::SHL,  MVT::v4i64,   2*4+2 }, // splat+shuffle sequence + split.
814 
815     { ISD::SRL,  MVT::v16i8,      26 }, // cmpgtb sequence.
816     { ISD::SRL,  MVT::v8i16,      32 }, // cmpgtb sequence.
817     { ISD::SRL,  MVT::v4i32,      16 }, // Shift each lane + blend.
818     { ISD::SRL,  MVT::v2i64,       4 }, // splat+shuffle sequence.
819     { ISD::SRL,  MVT::v4i64,   2*4+2 }, // splat+shuffle sequence + split.
820 
821     { ISD::SRA,  MVT::v16i8,      54 }, // unpacked cmpgtb sequence.
822     { ISD::SRA,  MVT::v8i16,      32 }, // cmpgtb sequence.
823     { ISD::SRA,  MVT::v4i32,      16 }, // Shift each lane + blend.
824     { ISD::SRA,  MVT::v2i64,      12 }, // srl/xor/sub sequence.
825     { ISD::SRA,  MVT::v4i64,  2*12+2 }, // srl/xor/sub sequence+split.
826 
827     { ISD::MUL,  MVT::v16i8,      12 }, // extend/pmullw/trunc sequence.
828     { ISD::MUL,  MVT::v8i16,       1 }, // pmullw
829     { ISD::MUL,  MVT::v4i32,       6 }, // 3*pmuludq/4*shuffle
830     { ISD::MUL,  MVT::v2i64,       8 }, // 3*pmuludq/3*shift/2*add
831 
832     { ISD::FDIV, MVT::f32,        23 }, // Pentium IV from http://www.agner.org/
833     { ISD::FDIV, MVT::v4f32,      39 }, // Pentium IV from http://www.agner.org/
834     { ISD::FDIV, MVT::f64,        38 }, // Pentium IV from http://www.agner.org/
835     { ISD::FDIV, MVT::v2f64,      69 }, // Pentium IV from http://www.agner.org/
836 
837     { ISD::FADD, MVT::f32,         2 }, // Pentium IV from http://www.agner.org/
838     { ISD::FADD, MVT::f64,         2 }, // Pentium IV from http://www.agner.org/
839 
840     { ISD::FSUB, MVT::f32,         2 }, // Pentium IV from http://www.agner.org/
841     { ISD::FSUB, MVT::f64,         2 }, // Pentium IV from http://www.agner.org/
842   };
843 
844   if (ST->hasSSE2())
845     if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second))
846       return LT.first * Entry->Cost;
847 
848   static const CostTblEntry SSE1CostTable[] = {
849     { ISD::FDIV, MVT::f32,   17 }, // Pentium III from http://www.agner.org/
850     { ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/
851 
852     { ISD::FADD, MVT::f32,    1 }, // Pentium III from http://www.agner.org/
853     { ISD::FADD, MVT::v4f32,  2 }, // Pentium III from http://www.agner.org/
854 
855     { ISD::FSUB, MVT::f32,    1 }, // Pentium III from http://www.agner.org/
856     { ISD::FSUB, MVT::v4f32,  2 }, // Pentium III from http://www.agner.org/
857 
858     { ISD::ADD, MVT::i8,      1 }, // Pentium III from http://www.agner.org/
859     { ISD::ADD, MVT::i16,     1 }, // Pentium III from http://www.agner.org/
860     { ISD::ADD, MVT::i32,     1 }, // Pentium III from http://www.agner.org/
861 
862     { ISD::SUB, MVT::i8,      1 }, // Pentium III from http://www.agner.org/
863     { ISD::SUB, MVT::i16,     1 }, // Pentium III from http://www.agner.org/
864     { ISD::SUB, MVT::i32,     1 }, // Pentium III from http://www.agner.org/
865   };
866 
867   if (ST->hasSSE1())
868     if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second))
869       return LT.first * Entry->Cost;
870 
871   // It is not a good idea to vectorize division. We have to scalarize it and
872   // in the process we will often end up having to spilling regular
873   // registers. The overhead of division is going to dominate most kernels
874   // anyways so try hard to prevent vectorization of division - it is
875   // generally a bad idea. Assume somewhat arbitrarily that we have to be able
876   // to hide "20 cycles" for each lane.
877   if (LT.second.isVector() && (ISD == ISD::SDIV || ISD == ISD::SREM ||
878                                ISD == ISD::UDIV || ISD == ISD::UREM)) {
879     int ScalarCost = getArithmeticInstrCost(
880         Opcode, Ty->getScalarType(), Op1Info, Op2Info,
881         TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
882     return 20 * LT.first * LT.second.getVectorNumElements() * ScalarCost;
883   }
884 
885   // Fallback to the default implementation.
886   return BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
887 }
888 
889 int X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
890                                Type *SubTp) {
891   // 64-bit packed float vectors (v2f32) are widened to type v4f32.
892   // 64-bit packed integer vectors (v2i32) are widened to type v4i32.
893   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
894 
895   // Treat Transpose as 2-op shuffles - there's no difference in lowering.
896   if (Kind == TTI::SK_Transpose)
897     Kind = TTI::SK_PermuteTwoSrc;
898 
899   // For Broadcasts we are splatting the first element from the first input
900   // register, so only need to reference that input and all the output
901   // registers are the same.
902   if (Kind == TTI::SK_Broadcast)
903     LT.first = 1;
904 
905   // Subvector extractions are free if they start at the beginning of a
906   // vector and cheap if the subvectors are aligned.
907   if (Kind == TTI::SK_ExtractSubvector && LT.second.isVector()) {
908     int NumElts = LT.second.getVectorNumElements();
909     if ((Index % NumElts) == 0)
910       return 0;
911     std::pair<int, MVT> SubLT = TLI->getTypeLegalizationCost(DL, SubTp);
912     if (SubLT.second.isVector()) {
913       int NumSubElts = SubLT.second.getVectorNumElements();
914       if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0)
915         return SubLT.first;
916       // Handle some cases for widening legalization. For now we only handle
917       // cases where the original subvector was naturally aligned and evenly
918       // fit in its legalized subvector type.
919       // FIXME: Remove some of the alignment restrictions.
920       // FIXME: We can use permq for 64-bit or larger extracts from 256-bit
921       // vectors.
922       int OrigSubElts = SubTp->getVectorNumElements();
923       if (NumSubElts > OrigSubElts &&
924           (Index % OrigSubElts) == 0 && (NumSubElts % OrigSubElts) == 0 &&
925           LT.second.getVectorElementType() ==
926             SubLT.second.getVectorElementType() &&
927           LT.second.getVectorElementType().getSizeInBits() ==
928             Tp->getVectorElementType()->getPrimitiveSizeInBits()) {
929         assert(NumElts >= NumSubElts && NumElts > OrigSubElts &&
930                "Unexpected number of elements!");
931         Type *VecTy = VectorType::get(Tp->getVectorElementType(),
932                                       LT.second.getVectorNumElements());
933         Type *SubTy = VectorType::get(Tp->getVectorElementType(),
934                                       SubLT.second.getVectorNumElements());
935         int ExtractIndex = alignDown((Index % NumElts), NumSubElts);
936         int ExtractCost = getShuffleCost(TTI::SK_ExtractSubvector, VecTy,
937                                          ExtractIndex, SubTy);
938 
939         // If the original size is 32-bits or more, we can use pshufd. Otherwise
940         // if we have SSSE3 we can use pshufb.
941         if (SubTp->getPrimitiveSizeInBits() >= 32 || ST->hasSSSE3())
942           return ExtractCost + 1; // pshufd or pshufb
943 
944         assert(SubTp->getPrimitiveSizeInBits() == 16 &&
945                "Unexpected vector size");
946 
947         return ExtractCost + 2; // worst case pshufhw + pshufd
948       }
949     }
950   }
951 
952   // We are going to permute multiple sources and the result will be in multiple
953   // destinations. Providing an accurate cost only for splits where the element
954   // type remains the same.
955   if (Kind == TTI::SK_PermuteSingleSrc && LT.first != 1) {
956     MVT LegalVT = LT.second;
957     if (LegalVT.isVector() &&
958         LegalVT.getVectorElementType().getSizeInBits() ==
959             Tp->getVectorElementType()->getPrimitiveSizeInBits() &&
960         LegalVT.getVectorNumElements() < Tp->getVectorNumElements()) {
961 
962       unsigned VecTySize = DL.getTypeStoreSize(Tp);
963       unsigned LegalVTSize = LegalVT.getStoreSize();
964       // Number of source vectors after legalization:
965       unsigned NumOfSrcs = (VecTySize + LegalVTSize - 1) / LegalVTSize;
966       // Number of destination vectors after legalization:
967       unsigned NumOfDests = LT.first;
968 
969       Type *SingleOpTy = VectorType::get(Tp->getVectorElementType(),
970                                          LegalVT.getVectorNumElements());
971 
972       unsigned NumOfShuffles = (NumOfSrcs - 1) * NumOfDests;
973       return NumOfShuffles *
974              getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy, 0, nullptr);
975     }
976 
977     return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
978   }
979 
980   // For 2-input shuffles, we must account for splitting the 2 inputs into many.
981   if (Kind == TTI::SK_PermuteTwoSrc && LT.first != 1) {
982     // We assume that source and destination have the same vector type.
983     int NumOfDests = LT.first;
984     int NumOfShufflesPerDest = LT.first * 2 - 1;
985     LT.first = NumOfDests * NumOfShufflesPerDest;
986   }
987 
988   static const CostTblEntry AVX512VBMIShuffleTbl[] = {
989       {TTI::SK_Reverse, MVT::v64i8, 1}, // vpermb
990       {TTI::SK_Reverse, MVT::v32i8, 1}, // vpermb
991 
992       {TTI::SK_PermuteSingleSrc, MVT::v64i8, 1}, // vpermb
993       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 1}, // vpermb
994 
995       {TTI::SK_PermuteTwoSrc, MVT::v64i8, 1}, // vpermt2b
996       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 1}, // vpermt2b
997       {TTI::SK_PermuteTwoSrc, MVT::v16i8, 1}  // vpermt2b
998   };
999 
1000   if (ST->hasVBMI())
1001     if (const auto *Entry =
1002             CostTableLookup(AVX512VBMIShuffleTbl, Kind, LT.second))
1003       return LT.first * Entry->Cost;
1004 
1005   static const CostTblEntry AVX512BWShuffleTbl[] = {
1006       {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw
1007       {TTI::SK_Broadcast, MVT::v64i8, 1},  // vpbroadcastb
1008 
1009       {TTI::SK_Reverse, MVT::v32i16, 1}, // vpermw
1010       {TTI::SK_Reverse, MVT::v16i16, 1}, // vpermw
1011       {TTI::SK_Reverse, MVT::v64i8, 2},  // pshufb + vshufi64x2
1012 
1013       {TTI::SK_PermuteSingleSrc, MVT::v32i16, 1}, // vpermw
1014       {TTI::SK_PermuteSingleSrc, MVT::v16i16, 1}, // vpermw
1015       {TTI::SK_PermuteSingleSrc, MVT::v8i16, 1},  // vpermw
1016       {TTI::SK_PermuteSingleSrc, MVT::v64i8, 8},  // extend to v32i16
1017       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 3},  // vpermw + zext/trunc
1018 
1019       {TTI::SK_PermuteTwoSrc, MVT::v32i16, 1}, // vpermt2w
1020       {TTI::SK_PermuteTwoSrc, MVT::v16i16, 1}, // vpermt2w
1021       {TTI::SK_PermuteTwoSrc, MVT::v8i16, 1},  // vpermt2w
1022       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 3},  // zext + vpermt2w + trunc
1023       {TTI::SK_PermuteTwoSrc, MVT::v64i8, 19}, // 6 * v32i8 + 1
1024       {TTI::SK_PermuteTwoSrc, MVT::v16i8, 3}   // zext + vpermt2w + trunc
1025   };
1026 
1027   if (ST->hasBWI())
1028     if (const auto *Entry =
1029             CostTableLookup(AVX512BWShuffleTbl, Kind, LT.second))
1030       return LT.first * Entry->Cost;
1031 
1032   static const CostTblEntry AVX512ShuffleTbl[] = {
1033       {TTI::SK_Broadcast, MVT::v8f64, 1},  // vbroadcastpd
1034       {TTI::SK_Broadcast, MVT::v16f32, 1}, // vbroadcastps
1035       {TTI::SK_Broadcast, MVT::v8i64, 1},  // vpbroadcastq
1036       {TTI::SK_Broadcast, MVT::v16i32, 1}, // vpbroadcastd
1037 
1038       {TTI::SK_Reverse, MVT::v8f64, 1},  // vpermpd
1039       {TTI::SK_Reverse, MVT::v16f32, 1}, // vpermps
1040       {TTI::SK_Reverse, MVT::v8i64, 1},  // vpermq
1041       {TTI::SK_Reverse, MVT::v16i32, 1}, // vpermd
1042 
1043       {TTI::SK_PermuteSingleSrc, MVT::v8f64, 1},  // vpermpd
1044       {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1},  // vpermpd
1045       {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1},  // vpermpd
1046       {TTI::SK_PermuteSingleSrc, MVT::v16f32, 1}, // vpermps
1047       {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1},  // vpermps
1048       {TTI::SK_PermuteSingleSrc, MVT::v4f32, 1},  // vpermps
1049       {TTI::SK_PermuteSingleSrc, MVT::v8i64, 1},  // vpermq
1050       {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1},  // vpermq
1051       {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1},  // vpermq
1052       {TTI::SK_PermuteSingleSrc, MVT::v16i32, 1}, // vpermd
1053       {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1},  // vpermd
1054       {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1},  // vpermd
1055       {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1},  // pshufb
1056 
1057       {TTI::SK_PermuteTwoSrc, MVT::v8f64, 1},  // vpermt2pd
1058       {TTI::SK_PermuteTwoSrc, MVT::v16f32, 1}, // vpermt2ps
1059       {TTI::SK_PermuteTwoSrc, MVT::v8i64, 1},  // vpermt2q
1060       {TTI::SK_PermuteTwoSrc, MVT::v16i32, 1}, // vpermt2d
1061       {TTI::SK_PermuteTwoSrc, MVT::v4f64, 1},  // vpermt2pd
1062       {TTI::SK_PermuteTwoSrc, MVT::v8f32, 1},  // vpermt2ps
1063       {TTI::SK_PermuteTwoSrc, MVT::v4i64, 1},  // vpermt2q
1064       {TTI::SK_PermuteTwoSrc, MVT::v8i32, 1},  // vpermt2d
1065       {TTI::SK_PermuteTwoSrc, MVT::v2f64, 1},  // vpermt2pd
1066       {TTI::SK_PermuteTwoSrc, MVT::v4f32, 1},  // vpermt2ps
1067       {TTI::SK_PermuteTwoSrc, MVT::v2i64, 1},  // vpermt2q
1068       {TTI::SK_PermuteTwoSrc, MVT::v4i32, 1}   // vpermt2d
1069   };
1070 
1071   if (ST->hasAVX512())
1072     if (const auto *Entry = CostTableLookup(AVX512ShuffleTbl, Kind, LT.second))
1073       return LT.first * Entry->Cost;
1074 
1075   static const CostTblEntry AVX2ShuffleTbl[] = {
1076       {TTI::SK_Broadcast, MVT::v4f64, 1},  // vbroadcastpd
1077       {TTI::SK_Broadcast, MVT::v8f32, 1},  // vbroadcastps
1078       {TTI::SK_Broadcast, MVT::v4i64, 1},  // vpbroadcastq
1079       {TTI::SK_Broadcast, MVT::v8i32, 1},  // vpbroadcastd
1080       {TTI::SK_Broadcast, MVT::v16i16, 1}, // vpbroadcastw
1081       {TTI::SK_Broadcast, MVT::v32i8, 1},  // vpbroadcastb
1082 
1083       {TTI::SK_Reverse, MVT::v4f64, 1},  // vpermpd
1084       {TTI::SK_Reverse, MVT::v8f32, 1},  // vpermps
1085       {TTI::SK_Reverse, MVT::v4i64, 1},  // vpermq
1086       {TTI::SK_Reverse, MVT::v8i32, 1},  // vpermd
1087       {TTI::SK_Reverse, MVT::v16i16, 2}, // vperm2i128 + pshufb
1088       {TTI::SK_Reverse, MVT::v32i8, 2},  // vperm2i128 + pshufb
1089 
1090       {TTI::SK_Select, MVT::v16i16, 1}, // vpblendvb
1091       {TTI::SK_Select, MVT::v32i8, 1},  // vpblendvb
1092 
1093       {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1},  // vpermpd
1094       {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1},  // vpermps
1095       {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1},  // vpermq
1096       {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1},  // vpermd
1097       {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vperm2i128 + 2*vpshufb
1098                                                   // + vpblendvb
1099       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4},  // vperm2i128 + 2*vpshufb
1100                                                   // + vpblendvb
1101 
1102       {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3},  // 2*vpermpd + vblendpd
1103       {TTI::SK_PermuteTwoSrc, MVT::v8f32, 3},  // 2*vpermps + vblendps
1104       {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3},  // 2*vpermq + vpblendd
1105       {TTI::SK_PermuteTwoSrc, MVT::v8i32, 3},  // 2*vpermd + vpblendd
1106       {TTI::SK_PermuteTwoSrc, MVT::v16i16, 7}, // 2*vperm2i128 + 4*vpshufb
1107                                                // + vpblendvb
1108       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 7},  // 2*vperm2i128 + 4*vpshufb
1109                                                // + vpblendvb
1110   };
1111 
1112   if (ST->hasAVX2())
1113     if (const auto *Entry = CostTableLookup(AVX2ShuffleTbl, Kind, LT.second))
1114       return LT.first * Entry->Cost;
1115 
1116   static const CostTblEntry XOPShuffleTbl[] = {
1117       {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2},  // vperm2f128 + vpermil2pd
1118       {TTI::SK_PermuteSingleSrc, MVT::v8f32, 2},  // vperm2f128 + vpermil2ps
1119       {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2},  // vperm2f128 + vpermil2pd
1120       {TTI::SK_PermuteSingleSrc, MVT::v8i32, 2},  // vperm2f128 + vpermil2ps
1121       {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vextractf128 + 2*vpperm
1122                                                   // + vinsertf128
1123       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4},  // vextractf128 + 2*vpperm
1124                                                   // + vinsertf128
1125 
1126       {TTI::SK_PermuteTwoSrc, MVT::v16i16, 9}, // 2*vextractf128 + 6*vpperm
1127                                                // + vinsertf128
1128       {TTI::SK_PermuteTwoSrc, MVT::v8i16, 1},  // vpperm
1129       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 9},  // 2*vextractf128 + 6*vpperm
1130                                                // + vinsertf128
1131       {TTI::SK_PermuteTwoSrc, MVT::v16i8, 1},  // vpperm
1132   };
1133 
1134   if (ST->hasXOP())
1135     if (const auto *Entry = CostTableLookup(XOPShuffleTbl, Kind, LT.second))
1136       return LT.first * Entry->Cost;
1137 
1138   static const CostTblEntry AVX1ShuffleTbl[] = {
1139       {TTI::SK_Broadcast, MVT::v4f64, 2},  // vperm2f128 + vpermilpd
1140       {TTI::SK_Broadcast, MVT::v8f32, 2},  // vperm2f128 + vpermilps
1141       {TTI::SK_Broadcast, MVT::v4i64, 2},  // vperm2f128 + vpermilpd
1142       {TTI::SK_Broadcast, MVT::v8i32, 2},  // vperm2f128 + vpermilps
1143       {TTI::SK_Broadcast, MVT::v16i16, 3}, // vpshuflw + vpshufd + vinsertf128
1144       {TTI::SK_Broadcast, MVT::v32i8, 2},  // vpshufb + vinsertf128
1145 
1146       {TTI::SK_Reverse, MVT::v4f64, 2},  // vperm2f128 + vpermilpd
1147       {TTI::SK_Reverse, MVT::v8f32, 2},  // vperm2f128 + vpermilps
1148       {TTI::SK_Reverse, MVT::v4i64, 2},  // vperm2f128 + vpermilpd
1149       {TTI::SK_Reverse, MVT::v8i32, 2},  // vperm2f128 + vpermilps
1150       {TTI::SK_Reverse, MVT::v16i16, 4}, // vextractf128 + 2*pshufb
1151                                          // + vinsertf128
1152       {TTI::SK_Reverse, MVT::v32i8, 4},  // vextractf128 + 2*pshufb
1153                                          // + vinsertf128
1154 
1155       {TTI::SK_Select, MVT::v4i64, 1},  // vblendpd
1156       {TTI::SK_Select, MVT::v4f64, 1},  // vblendpd
1157       {TTI::SK_Select, MVT::v8i32, 1},  // vblendps
1158       {TTI::SK_Select, MVT::v8f32, 1},  // vblendps
1159       {TTI::SK_Select, MVT::v16i16, 3}, // vpand + vpandn + vpor
1160       {TTI::SK_Select, MVT::v32i8, 3},  // vpand + vpandn + vpor
1161 
1162       {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2},  // vperm2f128 + vshufpd
1163       {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2},  // vperm2f128 + vshufpd
1164       {TTI::SK_PermuteSingleSrc, MVT::v8f32, 4},  // 2*vperm2f128 + 2*vshufps
1165       {TTI::SK_PermuteSingleSrc, MVT::v8i32, 4},  // 2*vperm2f128 + 2*vshufps
1166       {TTI::SK_PermuteSingleSrc, MVT::v16i16, 8}, // vextractf128 + 4*pshufb
1167                                                   // + 2*por + vinsertf128
1168       {TTI::SK_PermuteSingleSrc, MVT::v32i8, 8},  // vextractf128 + 4*pshufb
1169                                                   // + 2*por + vinsertf128
1170 
1171       {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3},   // 2*vperm2f128 + vshufpd
1172       {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3},   // 2*vperm2f128 + vshufpd
1173       {TTI::SK_PermuteTwoSrc, MVT::v8f32, 4},   // 2*vperm2f128 + 2*vshufps
1174       {TTI::SK_PermuteTwoSrc, MVT::v8i32, 4},   // 2*vperm2f128 + 2*vshufps
1175       {TTI::SK_PermuteTwoSrc, MVT::v16i16, 15}, // 2*vextractf128 + 8*pshufb
1176                                                 // + 4*por + vinsertf128
1177       {TTI::SK_PermuteTwoSrc, MVT::v32i8, 15},  // 2*vextractf128 + 8*pshufb
1178                                                 // + 4*por + vinsertf128
1179   };
1180 
1181   if (ST->hasAVX())
1182     if (const auto *Entry = CostTableLookup(AVX1ShuffleTbl, Kind, LT.second))
1183       return LT.first * Entry->Cost;
1184 
1185   static const CostTblEntry SSE41ShuffleTbl[] = {
1186       {TTI::SK_Select, MVT::v2i64, 1}, // pblendw
1187       {TTI::SK_Select, MVT::v2f64, 1}, // movsd
1188       {TTI::SK_Select, MVT::v4i32, 1}, // pblendw
1189       {TTI::SK_Select, MVT::v4f32, 1}, // blendps
1190       {TTI::SK_Select, MVT::v8i16, 1}, // pblendw
1191       {TTI::SK_Select, MVT::v16i8, 1}  // pblendvb
1192   };
1193 
1194   if (ST->hasSSE41())
1195     if (const auto *Entry = CostTableLookup(SSE41ShuffleTbl, Kind, LT.second))
1196       return LT.first * Entry->Cost;
1197 
1198   static const CostTblEntry SSSE3ShuffleTbl[] = {
1199       {TTI::SK_Broadcast, MVT::v8i16, 1}, // pshufb
1200       {TTI::SK_Broadcast, MVT::v16i8, 1}, // pshufb
1201 
1202       {TTI::SK_Reverse, MVT::v8i16, 1}, // pshufb
1203       {TTI::SK_Reverse, MVT::v16i8, 1}, // pshufb
1204 
1205       {TTI::SK_Select, MVT::v8i16, 3}, // 2*pshufb + por
1206       {TTI::SK_Select, MVT::v16i8, 3}, // 2*pshufb + por
1207 
1208       {TTI::SK_PermuteSingleSrc, MVT::v8i16, 1}, // pshufb
1209       {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb
1210 
1211       {TTI::SK_PermuteTwoSrc, MVT::v8i16, 3}, // 2*pshufb + por
1212       {TTI::SK_PermuteTwoSrc, MVT::v16i8, 3}, // 2*pshufb + por
1213   };
1214 
1215   if (ST->hasSSSE3())
1216     if (const auto *Entry = CostTableLookup(SSSE3ShuffleTbl, Kind, LT.second))
1217       return LT.first * Entry->Cost;
1218 
1219   static const CostTblEntry SSE2ShuffleTbl[] = {
1220       {TTI::SK_Broadcast, MVT::v2f64, 1}, // shufpd
1221       {TTI::SK_Broadcast, MVT::v2i64, 1}, // pshufd
1222       {TTI::SK_Broadcast, MVT::v4i32, 1}, // pshufd
1223       {TTI::SK_Broadcast, MVT::v8i16, 2}, // pshuflw + pshufd
1224       {TTI::SK_Broadcast, MVT::v16i8, 3}, // unpck + pshuflw + pshufd
1225 
1226       {TTI::SK_Reverse, MVT::v2f64, 1}, // shufpd
1227       {TTI::SK_Reverse, MVT::v2i64, 1}, // pshufd
1228       {TTI::SK_Reverse, MVT::v4i32, 1}, // pshufd
1229       {TTI::SK_Reverse, MVT::v8i16, 3}, // pshuflw + pshufhw + pshufd
1230       {TTI::SK_Reverse, MVT::v16i8, 9}, // 2*pshuflw + 2*pshufhw
1231                                         // + 2*pshufd + 2*unpck + packus
1232 
1233       {TTI::SK_Select, MVT::v2i64, 1}, // movsd
1234       {TTI::SK_Select, MVT::v2f64, 1}, // movsd
1235       {TTI::SK_Select, MVT::v4i32, 2}, // 2*shufps
1236       {TTI::SK_Select, MVT::v8i16, 3}, // pand + pandn + por
1237       {TTI::SK_Select, MVT::v16i8, 3}, // pand + pandn + por
1238 
1239       {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // shufpd
1240       {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // pshufd
1241       {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // pshufd
1242       {TTI::SK_PermuteSingleSrc, MVT::v8i16, 5}, // 2*pshuflw + 2*pshufhw
1243                                                   // + pshufd/unpck
1244     { TTI::SK_PermuteSingleSrc, MVT::v16i8, 10 }, // 2*pshuflw + 2*pshufhw
1245                                                   // + 2*pshufd + 2*unpck + 2*packus
1246 
1247     { TTI::SK_PermuteTwoSrc,    MVT::v2f64,  1 }, // shufpd
1248     { TTI::SK_PermuteTwoSrc,    MVT::v2i64,  1 }, // shufpd
1249     { TTI::SK_PermuteTwoSrc,    MVT::v4i32,  2 }, // 2*{unpck,movsd,pshufd}
1250     { TTI::SK_PermuteTwoSrc,    MVT::v8i16,  8 }, // blend+permute
1251     { TTI::SK_PermuteTwoSrc,    MVT::v16i8, 13 }, // blend+permute
1252   };
1253 
1254   if (ST->hasSSE2())
1255     if (const auto *Entry = CostTableLookup(SSE2ShuffleTbl, Kind, LT.second))
1256       return LT.first * Entry->Cost;
1257 
1258   static const CostTblEntry SSE1ShuffleTbl[] = {
1259     { TTI::SK_Broadcast,        MVT::v4f32, 1 }, // shufps
1260     { TTI::SK_Reverse,          MVT::v4f32, 1 }, // shufps
1261     { TTI::SK_Select,           MVT::v4f32, 2 }, // 2*shufps
1262     { TTI::SK_PermuteSingleSrc, MVT::v4f32, 1 }, // shufps
1263     { TTI::SK_PermuteTwoSrc,    MVT::v4f32, 2 }, // 2*shufps
1264   };
1265 
1266   if (ST->hasSSE1())
1267     if (const auto *Entry = CostTableLookup(SSE1ShuffleTbl, Kind, LT.second))
1268       return LT.first * Entry->Cost;
1269 
1270   return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
1271 }
1272 
1273 int X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
1274                                  const Instruction *I) {
1275   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1276   assert(ISD && "Invalid opcode");
1277 
1278   // FIXME: Need a better design of the cost table to handle non-simple types of
1279   // potential massive combinations (elem_num x src_type x dst_type).
1280 
1281   static const TypeConversionCostTblEntry AVX512BWConversionTbl[] {
1282     { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8, 1 },
1283     { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8, 1 },
1284 
1285     // Mask sign extend has an instruction.
1286     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i1,  1 },
1287     { ISD::SIGN_EXTEND, MVT::v16i8,  MVT::v16i1, 1 },
1288     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 },
1289     { ISD::SIGN_EXTEND, MVT::v32i8,  MVT::v32i1, 1 },
1290     { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i1, 1 },
1291     { ISD::SIGN_EXTEND, MVT::v64i8,  MVT::v64i1, 1 },
1292 
1293     // Mask zero extend is a load + broadcast.
1294     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i1,  2 },
1295     { ISD::ZERO_EXTEND, MVT::v16i8,  MVT::v16i1, 2 },
1296     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 2 },
1297     { ISD::ZERO_EXTEND, MVT::v32i8,  MVT::v32i1, 2 },
1298     { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i1, 2 },
1299     { ISD::ZERO_EXTEND, MVT::v64i8,  MVT::v64i1, 2 },
1300   };
1301 
1302   static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = {
1303     { ISD::SINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
1304     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
1305     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
1306     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
1307     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
1308     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
1309 
1310     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
1311     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
1312     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
1313     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
1314     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
1315     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
1316 
1317     { ISD::FP_TO_SINT,  MVT::v2i64,  MVT::v2f32,  1 },
1318     { ISD::FP_TO_SINT,  MVT::v4i64,  MVT::v4f32,  1 },
1319     { ISD::FP_TO_SINT,  MVT::v8i64,  MVT::v8f32,  1 },
1320     { ISD::FP_TO_SINT,  MVT::v2i64,  MVT::v2f64,  1 },
1321     { ISD::FP_TO_SINT,  MVT::v4i64,  MVT::v4f64,  1 },
1322     { ISD::FP_TO_SINT,  MVT::v8i64,  MVT::v8f64,  1 },
1323 
1324     { ISD::FP_TO_UINT,  MVT::v2i64,  MVT::v2f32,  1 },
1325     { ISD::FP_TO_UINT,  MVT::v4i64,  MVT::v4f32,  1 },
1326     { ISD::FP_TO_UINT,  MVT::v8i64,  MVT::v8f32,  1 },
1327     { ISD::FP_TO_UINT,  MVT::v2i64,  MVT::v2f64,  1 },
1328     { ISD::FP_TO_UINT,  MVT::v4i64,  MVT::v4f64,  1 },
1329     { ISD::FP_TO_UINT,  MVT::v8i64,  MVT::v8f64,  1 },
1330   };
1331 
1332   // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and
1333   // 256-bit wide vectors.
1334 
1335   static const TypeConversionCostTblEntry AVX512FConversionTbl[] = {
1336     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v8f32,  1 },
1337     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v16f32, 3 },
1338     { ISD::FP_ROUND,  MVT::v8f32,   MVT::v8f64,  1 },
1339 
1340     { ISD::TRUNCATE,  MVT::v16i8,   MVT::v16i32, 1 },
1341     { ISD::TRUNCATE,  MVT::v16i16,  MVT::v16i32, 1 },
1342     { ISD::TRUNCATE,  MVT::v8i16,   MVT::v8i64,  1 },
1343     { ISD::TRUNCATE,  MVT::v8i32,   MVT::v8i64,  1 },
1344 
1345     // v16i1 -> v16i32 - load + broadcast
1346     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
1347     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
1348     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
1349     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
1350     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1351     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1352     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i8,   1 },
1353     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i8,   1 },
1354     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
1355     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
1356     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
1357     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
1358 
1359     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
1360     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
1361     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i8,   2 },
1362     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i8,  2 },
1363     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
1364     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i16, 2 },
1365     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
1366     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
1367 
1368     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
1369     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
1370     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i8,   2 },
1371     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i8,   2 },
1372     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i8,   2 },
1373     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i8,   2 },
1374     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i8,  2 },
1375     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i16,  5 },
1376     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i16,  2 },
1377     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  2 },
1378     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
1379     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i16, 2 },
1380     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  2 },
1381     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i32,  1 },
1382     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  1 },
1383     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  1 },
1384     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  1 },
1385     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
1386     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
1387     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  5 },
1388     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64, 26 },
1389     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  5 },
1390     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  5 },
1391     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  5 },
1392 
1393     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i64,    1 },
1394     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i64,    1 },
1395     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f32,    1 },
1396     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f64,    1 },
1397 
1398     { ISD::FP_TO_UINT,  MVT::v2i32,  MVT::v2f32,  1 },
1399     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  1 },
1400     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f64,  1 },
1401     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f32,  1 },
1402     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f64,  1 },
1403     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v8f64,  2 },
1404     { ISD::FP_TO_UINT,  MVT::v8i8,   MVT::v8f64,  2 },
1405     { ISD::FP_TO_UINT,  MVT::v16i32, MVT::v16f32, 1 },
1406     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v16f32, 2 },
1407     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v16f32, 2 },
1408   };
1409 
1410   static const TypeConversionCostTblEntry AVX2ConversionTbl[] = {
1411     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
1412     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
1413     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
1414     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
1415     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,   1 },
1416     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,   1 },
1417     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   1 },
1418     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   1 },
1419     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
1420     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
1421     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16,  1 },
1422     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16,  1 },
1423     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
1424     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
1425     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
1426     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
1427 
1428     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i64,  2 },
1429     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i64,  2 },
1430     { ISD::TRUNCATE,    MVT::v4i32,  MVT::v4i64,  2 },
1431     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  2 },
1432     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  2 },
1433     { ISD::TRUNCATE,    MVT::v8i32,  MVT::v8i64,  4 },
1434 
1435     { ISD::FP_EXTEND,   MVT::v8f64,  MVT::v8f32,  3 },
1436     { ISD::FP_ROUND,    MVT::v8f32,  MVT::v8f64,  3 },
1437 
1438     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  8 },
1439   };
1440 
1441   static const TypeConversionCostTblEntry AVXConversionTbl[] = {
1442     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,  6 },
1443     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,  4 },
1444     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,  7 },
1445     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,  4 },
1446     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,  4 },
1447     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,  4 },
1448     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,  4 },
1449     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,  4 },
1450     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
1451     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
1452     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16, 4 },
1453     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16, 3 },
1454     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16, 4 },
1455     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16, 4 },
1456     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32, 4 },
1457     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32, 4 },
1458 
1459     { ISD::TRUNCATE,    MVT::v16i8, MVT::v16i16, 4 },
1460     { ISD::TRUNCATE,    MVT::v8i8,  MVT::v8i32,  4 },
1461     { ISD::TRUNCATE,    MVT::v8i16, MVT::v8i32,  5 },
1462     { ISD::TRUNCATE,    MVT::v4i8,  MVT::v4i64,  4 },
1463     { ISD::TRUNCATE,    MVT::v4i16, MVT::v4i64,  4 },
1464     { ISD::TRUNCATE,    MVT::v4i32, MVT::v4i64,  4 },
1465     { ISD::TRUNCATE,    MVT::v8i8,  MVT::v8i64, 11 },
1466     { ISD::TRUNCATE,    MVT::v8i16, MVT::v8i64,  9 },
1467     { ISD::TRUNCATE,    MVT::v8i32, MVT::v8i64,  9 },
1468     { ISD::TRUNCATE,    MVT::v16i8, MVT::v16i64, 11 },
1469 
1470     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i1,  3 },
1471     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i1,  3 },
1472     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i1,  8 },
1473     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i8,  3 },
1474     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i8,  3 },
1475     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i8,  8 },
1476     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i16, 3 },
1477     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i16, 3 },
1478     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
1479     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },
1480     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i32, 1 },
1481     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i32, 1 },
1482 
1483     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i1,  7 },
1484     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i1,  7 },
1485     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i1,  6 },
1486     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i8,  2 },
1487     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i8,  2 },
1488     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i8,  5 },
1489     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
1490     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i16, 2 },
1491     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
1492     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i32, 6 },
1493     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i32, 6 },
1494     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i32, 6 },
1495     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i32, 9 },
1496     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i64, 5 },
1497     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i64, 6 },
1498     // The generic code to compute the scalar overhead is currently broken.
1499     // Workaround this limitation by estimating the scalarization overhead
1500     // here. We have roughly 10 instructions per scalar element.
1501     // Multiply that by the vector width.
1502     // FIXME: remove that when PR19268 is fixed.
1503     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
1504     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
1505 
1506     { ISD::FP_TO_SINT,  MVT::v4i8,  MVT::v4f32, 1 },
1507     { ISD::FP_TO_SINT,  MVT::v8i8,  MVT::v8f32, 7 },
1508     // This node is expanded into scalarized operations but BasicTTI is overly
1509     // optimistic estimating its cost.  It computes 3 per element (one
1510     // vector-extract, one scalar conversion and one vector-insert).  The
1511     // problem is that the inserts form a read-modify-write chain so latency
1512     // should be factored in too.  Inflating the cost per element by 1.
1513     { ISD::FP_TO_UINT,  MVT::v8i32, MVT::v8f32, 8*4 },
1514     { ISD::FP_TO_UINT,  MVT::v4i32, MVT::v4f64, 4*4 },
1515 
1516     { ISD::FP_EXTEND,   MVT::v4f64,  MVT::v4f32,  1 },
1517     { ISD::FP_ROUND,    MVT::v4f32,  MVT::v4f64,  1 },
1518   };
1519 
1520   static const TypeConversionCostTblEntry SSE41ConversionTbl[] = {
1521     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8,    2 },
1522     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8,    2 },
1523     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16,   2 },
1524     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16,   2 },
1525     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32,   2 },
1526     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32,   2 },
1527 
1528     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i8,   1 },
1529     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i8,   2 },
1530     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i8,   1 },
1531     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i8,   1 },
1532     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
1533     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
1534     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   2 },
1535     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   2 },
1536     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
1537     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
1538     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  4 },
1539     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  4 },
1540     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
1541     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
1542     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
1543     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
1544     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 4 },
1545     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 4 },
1546 
1547     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  2 },
1548     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  1 },
1549     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i32,  1 },
1550     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i32,  1 },
1551     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  3 },
1552     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  3 },
1553     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 6 },
1554     { ISD::TRUNCATE,    MVT::v2i8,   MVT::v2i64,  1 }, // PSHUFB
1555 
1556     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i64,    4 },
1557     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i64,    4 },
1558   };
1559 
1560   static const TypeConversionCostTblEntry SSE2ConversionTbl[] = {
1561     // These are somewhat magic numbers justified by looking at the output of
1562     // Intel's IACA, running some kernels and making sure when we take
1563     // legalization into account the throughput will be overestimated.
1564     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
1565     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
1566     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
1567     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
1568     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 },
1569     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 2*10 },
1570     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 2*10 },
1571     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
1572     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
1573 
1574     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
1575     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
1576     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
1577     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
1578     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
1579     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 8 },
1580     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 6 },
1581     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
1582 
1583     { ISD::FP_TO_SINT,  MVT::v4i16,  MVT::v4f32,  2 },
1584     { ISD::FP_TO_SINT,  MVT::v2i16,  MVT::v2f64,  2 },
1585 
1586     { ISD::FP_TO_SINT,  MVT::v2i32,  MVT::v2f64,  3 },
1587 
1588     { ISD::UINT_TO_FP,  MVT::f32,    MVT::i64,    6 },
1589     { ISD::UINT_TO_FP,  MVT::f64,    MVT::i64,    6 },
1590 
1591     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f32,    4 },
1592     { ISD::FP_TO_UINT,  MVT::i64,    MVT::f64,    4 },
1593 
1594     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i8,   1 },
1595     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i8,   6 },
1596     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i8,   2 },
1597     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i8,   3 },
1598     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,   4 },
1599     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,   8 },
1600     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
1601     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i8,   2 },
1602     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   6 },
1603     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   6 },
1604     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  3 },
1605     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  4 },
1606     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  9 },
1607     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  12 },
1608     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
1609     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i16,  2 },
1610     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
1611     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16,  10 },
1612     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  3 },
1613     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  4 },
1614     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 6 },
1615     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 8 },
1616     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  3 },
1617     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  5 },
1618 
1619     { ISD::TRUNCATE,    MVT::v2i8,   MVT::v2i16,  2 }, // PAND+PACKUSWB
1620     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  4 },
1621     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  2 },
1622     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 3 },
1623     { ISD::TRUNCATE,    MVT::v2i8,   MVT::v2i32,  3 }, // PAND+3*PACKUSWB
1624     { ISD::TRUNCATE,    MVT::v2i16,  MVT::v2i32,  1 },
1625     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i32,  3 },
1626     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i32,  3 },
1627     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  4 },
1628     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i32, 7 },
1629     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  5 },
1630     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 10 },
1631     { ISD::TRUNCATE,    MVT::v2i8,   MVT::v2i64,  4 }, // PAND+3*PACKUSWB
1632     { ISD::TRUNCATE,    MVT::v2i16,  MVT::v2i64,  2 }, // PSHUFD+PSHUFLW
1633     { ISD::TRUNCATE,    MVT::v2i32,  MVT::v2i64,  1 }, // PSHUFD
1634   };
1635 
1636   std::pair<int, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src);
1637   std::pair<int, MVT> LTDest = TLI->getTypeLegalizationCost(DL, Dst);
1638 
1639   if (ST->hasSSE2() && !ST->hasAVX()) {
1640     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
1641                                                    LTDest.second, LTSrc.second))
1642       return LTSrc.first * Entry->Cost;
1643   }
1644 
1645   EVT SrcTy = TLI->getValueType(DL, Src);
1646   EVT DstTy = TLI->getValueType(DL, Dst);
1647 
1648   // The function getSimpleVT only handles simple value types.
1649   if (!SrcTy.isSimple() || !DstTy.isSimple())
1650     return BaseT::getCastInstrCost(Opcode, Dst, Src);
1651 
1652   MVT SimpleSrcTy = SrcTy.getSimpleVT();
1653   MVT SimpleDstTy = DstTy.getSimpleVT();
1654 
1655   // Make sure that neither type is going to be split before using the
1656   // AVX512 tables. This handles -mprefer-vector-width=256
1657   // with -min-legal-vector-width<=256
1658   if (TLI->getTypeAction(SimpleSrcTy) != TargetLowering::TypeSplitVector &&
1659       TLI->getTypeAction(SimpleDstTy) != TargetLowering::TypeSplitVector) {
1660     if (ST->hasBWI())
1661       if (const auto *Entry = ConvertCostTableLookup(AVX512BWConversionTbl, ISD,
1662                                                      SimpleDstTy, SimpleSrcTy))
1663         return Entry->Cost;
1664 
1665     if (ST->hasDQI())
1666       if (const auto *Entry = ConvertCostTableLookup(AVX512DQConversionTbl, ISD,
1667                                                      SimpleDstTy, SimpleSrcTy))
1668         return Entry->Cost;
1669 
1670     if (ST->hasAVX512())
1671       if (const auto *Entry = ConvertCostTableLookup(AVX512FConversionTbl, ISD,
1672                                                      SimpleDstTy, SimpleSrcTy))
1673         return Entry->Cost;
1674   }
1675 
1676   if (ST->hasAVX2()) {
1677     if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
1678                                                    SimpleDstTy, SimpleSrcTy))
1679       return Entry->Cost;
1680   }
1681 
1682   if (ST->hasAVX()) {
1683     if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
1684                                                    SimpleDstTy, SimpleSrcTy))
1685       return Entry->Cost;
1686   }
1687 
1688   if (ST->hasSSE41()) {
1689     if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
1690                                                    SimpleDstTy, SimpleSrcTy))
1691       return Entry->Cost;
1692   }
1693 
1694   if (ST->hasSSE2()) {
1695     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
1696                                                    SimpleDstTy, SimpleSrcTy))
1697       return Entry->Cost;
1698   }
1699 
1700   return BaseT::getCastInstrCost(Opcode, Dst, Src, I);
1701 }
1702 
1703 int X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
1704                                    const Instruction *I) {
1705   // Legalize the type.
1706   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
1707 
1708   MVT MTy = LT.second;
1709 
1710   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1711   assert(ISD && "Invalid opcode");
1712 
1713   unsigned ExtraCost = 0;
1714   if (I && (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)) {
1715     // Some vector comparison predicates cost extra instructions.
1716     if (MTy.isVector() &&
1717         !((ST->hasXOP() && (!ST->hasAVX2() || MTy.is128BitVector())) ||
1718           (ST->hasAVX512() && 32 <= MTy.getScalarSizeInBits()) ||
1719           ST->hasBWI())) {
1720       switch (cast<CmpInst>(I)->getPredicate()) {
1721       case CmpInst::Predicate::ICMP_NE:
1722         // xor(cmpeq(x,y),-1)
1723         ExtraCost = 1;
1724         break;
1725       case CmpInst::Predicate::ICMP_SGE:
1726       case CmpInst::Predicate::ICMP_SLE:
1727         // xor(cmpgt(x,y),-1)
1728         ExtraCost = 1;
1729         break;
1730       case CmpInst::Predicate::ICMP_ULT:
1731       case CmpInst::Predicate::ICMP_UGT:
1732         // cmpgt(xor(x,signbit),xor(y,signbit))
1733         // xor(cmpeq(pmaxu(x,y),x),-1)
1734         ExtraCost = 2;
1735         break;
1736       case CmpInst::Predicate::ICMP_ULE:
1737       case CmpInst::Predicate::ICMP_UGE:
1738         if ((ST->hasSSE41() && MTy.getScalarSizeInBits() == 32) ||
1739             (ST->hasSSE2() && MTy.getScalarSizeInBits() < 32)) {
1740           // cmpeq(psubus(x,y),0)
1741           // cmpeq(pminu(x,y),x)
1742           ExtraCost = 1;
1743         } else {
1744           // xor(cmpgt(xor(x,signbit),xor(y,signbit)),-1)
1745           ExtraCost = 3;
1746         }
1747         break;
1748       default:
1749         break;
1750       }
1751     }
1752   }
1753 
1754   static const CostTblEntry SLMCostTbl[] = {
1755     // slm pcmpeq/pcmpgt throughput is 2
1756     { ISD::SETCC,   MVT::v2i64,   2 },
1757   };
1758 
1759   static const CostTblEntry AVX512BWCostTbl[] = {
1760     { ISD::SETCC,   MVT::v32i16,  1 },
1761     { ISD::SETCC,   MVT::v64i8,   1 },
1762 
1763     { ISD::SELECT,  MVT::v32i16,  1 },
1764     { ISD::SELECT,  MVT::v64i8,   1 },
1765   };
1766 
1767   static const CostTblEntry AVX512CostTbl[] = {
1768     { ISD::SETCC,   MVT::v8i64,   1 },
1769     { ISD::SETCC,   MVT::v16i32,  1 },
1770     { ISD::SETCC,   MVT::v8f64,   1 },
1771     { ISD::SETCC,   MVT::v16f32,  1 },
1772 
1773     { ISD::SELECT,  MVT::v8i64,   1 },
1774     { ISD::SELECT,  MVT::v16i32,  1 },
1775     { ISD::SELECT,  MVT::v8f64,   1 },
1776     { ISD::SELECT,  MVT::v16f32,  1 },
1777   };
1778 
1779   static const CostTblEntry AVX2CostTbl[] = {
1780     { ISD::SETCC,   MVT::v4i64,   1 },
1781     { ISD::SETCC,   MVT::v8i32,   1 },
1782     { ISD::SETCC,   MVT::v16i16,  1 },
1783     { ISD::SETCC,   MVT::v32i8,   1 },
1784 
1785     { ISD::SELECT,  MVT::v4i64,   1 }, // pblendvb
1786     { ISD::SELECT,  MVT::v8i32,   1 }, // pblendvb
1787     { ISD::SELECT,  MVT::v16i16,  1 }, // pblendvb
1788     { ISD::SELECT,  MVT::v32i8,   1 }, // pblendvb
1789   };
1790 
1791   static const CostTblEntry AVX1CostTbl[] = {
1792     { ISD::SETCC,   MVT::v4f64,   1 },
1793     { ISD::SETCC,   MVT::v8f32,   1 },
1794     // AVX1 does not support 8-wide integer compare.
1795     { ISD::SETCC,   MVT::v4i64,   4 },
1796     { ISD::SETCC,   MVT::v8i32,   4 },
1797     { ISD::SETCC,   MVT::v16i16,  4 },
1798     { ISD::SETCC,   MVT::v32i8,   4 },
1799 
1800     { ISD::SELECT,  MVT::v4f64,   1 }, // vblendvpd
1801     { ISD::SELECT,  MVT::v8f32,   1 }, // vblendvps
1802     { ISD::SELECT,  MVT::v4i64,   1 }, // vblendvpd
1803     { ISD::SELECT,  MVT::v8i32,   1 }, // vblendvps
1804     { ISD::SELECT,  MVT::v16i16,  3 }, // vandps + vandnps + vorps
1805     { ISD::SELECT,  MVT::v32i8,   3 }, // vandps + vandnps + vorps
1806   };
1807 
1808   static const CostTblEntry SSE42CostTbl[] = {
1809     { ISD::SETCC,   MVT::v2f64,   1 },
1810     { ISD::SETCC,   MVT::v4f32,   1 },
1811     { ISD::SETCC,   MVT::v2i64,   1 },
1812   };
1813 
1814   static const CostTblEntry SSE41CostTbl[] = {
1815     { ISD::SELECT,  MVT::v2f64,   1 }, // blendvpd
1816     { ISD::SELECT,  MVT::v4f32,   1 }, // blendvps
1817     { ISD::SELECT,  MVT::v2i64,   1 }, // pblendvb
1818     { ISD::SELECT,  MVT::v4i32,   1 }, // pblendvb
1819     { ISD::SELECT,  MVT::v8i16,   1 }, // pblendvb
1820     { ISD::SELECT,  MVT::v16i8,   1 }, // pblendvb
1821   };
1822 
1823   static const CostTblEntry SSE2CostTbl[] = {
1824     { ISD::SETCC,   MVT::v2f64,   2 },
1825     { ISD::SETCC,   MVT::f64,     1 },
1826     { ISD::SETCC,   MVT::v2i64,   8 },
1827     { ISD::SETCC,   MVT::v4i32,   1 },
1828     { ISD::SETCC,   MVT::v8i16,   1 },
1829     { ISD::SETCC,   MVT::v16i8,   1 },
1830 
1831     { ISD::SELECT,  MVT::v2f64,   3 }, // andpd + andnpd + orpd
1832     { ISD::SELECT,  MVT::v2i64,   3 }, // pand + pandn + por
1833     { ISD::SELECT,  MVT::v4i32,   3 }, // pand + pandn + por
1834     { ISD::SELECT,  MVT::v8i16,   3 }, // pand + pandn + por
1835     { ISD::SELECT,  MVT::v16i8,   3 }, // pand + pandn + por
1836   };
1837 
1838   static const CostTblEntry SSE1CostTbl[] = {
1839     { ISD::SETCC,   MVT::v4f32,   2 },
1840     { ISD::SETCC,   MVT::f32,     1 },
1841 
1842     { ISD::SELECT,  MVT::v4f32,   3 }, // andps + andnps + orps
1843   };
1844 
1845   if (ST->isSLM())
1846     if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy))
1847       return LT.first * (ExtraCost + Entry->Cost);
1848 
1849   if (ST->hasBWI())
1850     if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
1851       return LT.first * (ExtraCost + Entry->Cost);
1852 
1853   if (ST->hasAVX512())
1854     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
1855       return LT.first * (ExtraCost + Entry->Cost);
1856 
1857   if (ST->hasAVX2())
1858     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
1859       return LT.first * (ExtraCost + Entry->Cost);
1860 
1861   if (ST->hasAVX())
1862     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
1863       return LT.first * (ExtraCost + Entry->Cost);
1864 
1865   if (ST->hasSSE42())
1866     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
1867       return LT.first * (ExtraCost + Entry->Cost);
1868 
1869   if (ST->hasSSE41())
1870     if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
1871       return LT.first * (ExtraCost + Entry->Cost);
1872 
1873   if (ST->hasSSE2())
1874     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
1875       return LT.first * (ExtraCost + Entry->Cost);
1876 
1877   if (ST->hasSSE1())
1878     if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
1879       return LT.first * (ExtraCost + Entry->Cost);
1880 
1881   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, I);
1882 }
1883 
1884 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; }
1885 
1886 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1887                                       ArrayRef<Type *> Tys, FastMathFlags FMF,
1888                                       unsigned ScalarizationCostPassed) {
1889   // Costs should match the codegen from:
1890   // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll
1891   // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll
1892   // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll
1893   // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll
1894   // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll
1895   static const CostTblEntry AVX512CDCostTbl[] = {
1896     { ISD::CTLZ,       MVT::v8i64,   1 },
1897     { ISD::CTLZ,       MVT::v16i32,  1 },
1898     { ISD::CTLZ,       MVT::v32i16,  8 },
1899     { ISD::CTLZ,       MVT::v64i8,  20 },
1900     { ISD::CTLZ,       MVT::v4i64,   1 },
1901     { ISD::CTLZ,       MVT::v8i32,   1 },
1902     { ISD::CTLZ,       MVT::v16i16,  4 },
1903     { ISD::CTLZ,       MVT::v32i8,  10 },
1904     { ISD::CTLZ,       MVT::v2i64,   1 },
1905     { ISD::CTLZ,       MVT::v4i32,   1 },
1906     { ISD::CTLZ,       MVT::v8i16,   4 },
1907     { ISD::CTLZ,       MVT::v16i8,   4 },
1908   };
1909   static const CostTblEntry AVX512BWCostTbl[] = {
1910     { ISD::BITREVERSE, MVT::v8i64,   5 },
1911     { ISD::BITREVERSE, MVT::v16i32,  5 },
1912     { ISD::BITREVERSE, MVT::v32i16,  5 },
1913     { ISD::BITREVERSE, MVT::v64i8,   5 },
1914     { ISD::CTLZ,       MVT::v8i64,  23 },
1915     { ISD::CTLZ,       MVT::v16i32, 22 },
1916     { ISD::CTLZ,       MVT::v32i16, 18 },
1917     { ISD::CTLZ,       MVT::v64i8,  17 },
1918     { ISD::CTPOP,      MVT::v8i64,   7 },
1919     { ISD::CTPOP,      MVT::v16i32, 11 },
1920     { ISD::CTPOP,      MVT::v32i16,  9 },
1921     { ISD::CTPOP,      MVT::v64i8,   6 },
1922     { ISD::CTTZ,       MVT::v8i64,  10 },
1923     { ISD::CTTZ,       MVT::v16i32, 14 },
1924     { ISD::CTTZ,       MVT::v32i16, 12 },
1925     { ISD::CTTZ,       MVT::v64i8,   9 },
1926     { ISD::SADDSAT,    MVT::v32i16,  1 },
1927     { ISD::SADDSAT,    MVT::v64i8,   1 },
1928     { ISD::SSUBSAT,    MVT::v32i16,  1 },
1929     { ISD::SSUBSAT,    MVT::v64i8,   1 },
1930     { ISD::UADDSAT,    MVT::v32i16,  1 },
1931     { ISD::UADDSAT,    MVT::v64i8,   1 },
1932     { ISD::USUBSAT,    MVT::v32i16,  1 },
1933     { ISD::USUBSAT,    MVT::v64i8,   1 },
1934   };
1935   static const CostTblEntry AVX512CostTbl[] = {
1936     { ISD::BITREVERSE, MVT::v8i64,  36 },
1937     { ISD::BITREVERSE, MVT::v16i32, 24 },
1938     { ISD::CTLZ,       MVT::v8i64,  29 },
1939     { ISD::CTLZ,       MVT::v16i32, 35 },
1940     { ISD::CTPOP,      MVT::v8i64,  16 },
1941     { ISD::CTPOP,      MVT::v16i32, 24 },
1942     { ISD::CTTZ,       MVT::v8i64,  20 },
1943     { ISD::CTTZ,       MVT::v16i32, 28 },
1944     { ISD::USUBSAT,    MVT::v16i32,  2 }, // pmaxud + psubd
1945     { ISD::USUBSAT,    MVT::v2i64,   2 }, // pmaxuq + psubq
1946     { ISD::USUBSAT,    MVT::v4i64,   2 }, // pmaxuq + psubq
1947     { ISD::USUBSAT,    MVT::v8i64,   2 }, // pmaxuq + psubq
1948     { ISD::UADDSAT,    MVT::v16i32,  3 }, // not + pminud + paddd
1949     { ISD::UADDSAT,    MVT::v2i64,   3 }, // not + pminuq + paddq
1950     { ISD::UADDSAT,    MVT::v4i64,   3 }, // not + pminuq + paddq
1951     { ISD::UADDSAT,    MVT::v8i64,   3 }, // not + pminuq + paddq
1952   };
1953   static const CostTblEntry XOPCostTbl[] = {
1954     { ISD::BITREVERSE, MVT::v4i64,   4 },
1955     { ISD::BITREVERSE, MVT::v8i32,   4 },
1956     { ISD::BITREVERSE, MVT::v16i16,  4 },
1957     { ISD::BITREVERSE, MVT::v32i8,   4 },
1958     { ISD::BITREVERSE, MVT::v2i64,   1 },
1959     { ISD::BITREVERSE, MVT::v4i32,   1 },
1960     { ISD::BITREVERSE, MVT::v8i16,   1 },
1961     { ISD::BITREVERSE, MVT::v16i8,   1 },
1962     { ISD::BITREVERSE, MVT::i64,     3 },
1963     { ISD::BITREVERSE, MVT::i32,     3 },
1964     { ISD::BITREVERSE, MVT::i16,     3 },
1965     { ISD::BITREVERSE, MVT::i8,      3 }
1966   };
1967   static const CostTblEntry AVX2CostTbl[] = {
1968     { ISD::BITREVERSE, MVT::v4i64,   5 },
1969     { ISD::BITREVERSE, MVT::v8i32,   5 },
1970     { ISD::BITREVERSE, MVT::v16i16,  5 },
1971     { ISD::BITREVERSE, MVT::v32i8,   5 },
1972     { ISD::BSWAP,      MVT::v4i64,   1 },
1973     { ISD::BSWAP,      MVT::v8i32,   1 },
1974     { ISD::BSWAP,      MVT::v16i16,  1 },
1975     { ISD::CTLZ,       MVT::v4i64,  23 },
1976     { ISD::CTLZ,       MVT::v8i32,  18 },
1977     { ISD::CTLZ,       MVT::v16i16, 14 },
1978     { ISD::CTLZ,       MVT::v32i8,   9 },
1979     { ISD::CTPOP,      MVT::v4i64,   7 },
1980     { ISD::CTPOP,      MVT::v8i32,  11 },
1981     { ISD::CTPOP,      MVT::v16i16,  9 },
1982     { ISD::CTPOP,      MVT::v32i8,   6 },
1983     { ISD::CTTZ,       MVT::v4i64,  10 },
1984     { ISD::CTTZ,       MVT::v8i32,  14 },
1985     { ISD::CTTZ,       MVT::v16i16, 12 },
1986     { ISD::CTTZ,       MVT::v32i8,   9 },
1987     { ISD::SADDSAT,    MVT::v16i16,  1 },
1988     { ISD::SADDSAT,    MVT::v32i8,   1 },
1989     { ISD::SSUBSAT,    MVT::v16i16,  1 },
1990     { ISD::SSUBSAT,    MVT::v32i8,   1 },
1991     { ISD::UADDSAT,    MVT::v16i16,  1 },
1992     { ISD::UADDSAT,    MVT::v32i8,   1 },
1993     { ISD::UADDSAT,    MVT::v8i32,   3 }, // not + pminud + paddd
1994     { ISD::USUBSAT,    MVT::v16i16,  1 },
1995     { ISD::USUBSAT,    MVT::v32i8,   1 },
1996     { ISD::USUBSAT,    MVT::v8i32,   2 }, // pmaxud + psubd
1997     { ISD::FSQRT,      MVT::f32,     7 }, // Haswell from http://www.agner.org/
1998     { ISD::FSQRT,      MVT::v4f32,   7 }, // Haswell from http://www.agner.org/
1999     { ISD::FSQRT,      MVT::v8f32,  14 }, // Haswell from http://www.agner.org/
2000     { ISD::FSQRT,      MVT::f64,    14 }, // Haswell from http://www.agner.org/
2001     { ISD::FSQRT,      MVT::v2f64,  14 }, // Haswell from http://www.agner.org/
2002     { ISD::FSQRT,      MVT::v4f64,  28 }, // Haswell from http://www.agner.org/
2003   };
2004   static const CostTblEntry AVX1CostTbl[] = {
2005     { ISD::BITREVERSE, MVT::v4i64,  12 }, // 2 x 128-bit Op + extract/insert
2006     { ISD::BITREVERSE, MVT::v8i32,  12 }, // 2 x 128-bit Op + extract/insert
2007     { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert
2008     { ISD::BITREVERSE, MVT::v32i8,  12 }, // 2 x 128-bit Op + extract/insert
2009     { ISD::BSWAP,      MVT::v4i64,   4 },
2010     { ISD::BSWAP,      MVT::v8i32,   4 },
2011     { ISD::BSWAP,      MVT::v16i16,  4 },
2012     { ISD::CTLZ,       MVT::v4i64,  48 }, // 2 x 128-bit Op + extract/insert
2013     { ISD::CTLZ,       MVT::v8i32,  38 }, // 2 x 128-bit Op + extract/insert
2014     { ISD::CTLZ,       MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert
2015     { ISD::CTLZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
2016     { ISD::CTPOP,      MVT::v4i64,  16 }, // 2 x 128-bit Op + extract/insert
2017     { ISD::CTPOP,      MVT::v8i32,  24 }, // 2 x 128-bit Op + extract/insert
2018     { ISD::CTPOP,      MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert
2019     { ISD::CTPOP,      MVT::v32i8,  14 }, // 2 x 128-bit Op + extract/insert
2020     { ISD::CTTZ,       MVT::v4i64,  22 }, // 2 x 128-bit Op + extract/insert
2021     { ISD::CTTZ,       MVT::v8i32,  30 }, // 2 x 128-bit Op + extract/insert
2022     { ISD::CTTZ,       MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert
2023     { ISD::CTTZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
2024     { ISD::SADDSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
2025     { ISD::SADDSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
2026     { ISD::SSUBSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
2027     { ISD::SSUBSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
2028     { ISD::UADDSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
2029     { ISD::UADDSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
2030     { ISD::UADDSAT,    MVT::v8i32,   8 }, // 2 x 128-bit Op + extract/insert
2031     { ISD::USUBSAT,    MVT::v16i16,  4 }, // 2 x 128-bit Op + extract/insert
2032     { ISD::USUBSAT,    MVT::v32i8,   4 }, // 2 x 128-bit Op + extract/insert
2033     { ISD::USUBSAT,    MVT::v8i32,   6 }, // 2 x 128-bit Op + extract/insert
2034     { ISD::FSQRT,      MVT::f32,    14 }, // SNB from http://www.agner.org/
2035     { ISD::FSQRT,      MVT::v4f32,  14 }, // SNB from http://www.agner.org/
2036     { ISD::FSQRT,      MVT::v8f32,  28 }, // SNB from http://www.agner.org/
2037     { ISD::FSQRT,      MVT::f64,    21 }, // SNB from http://www.agner.org/
2038     { ISD::FSQRT,      MVT::v2f64,  21 }, // SNB from http://www.agner.org/
2039     { ISD::FSQRT,      MVT::v4f64,  43 }, // SNB from http://www.agner.org/
2040   };
2041   static const CostTblEntry GLMCostTbl[] = {
2042     { ISD::FSQRT, MVT::f32,   19 }, // sqrtss
2043     { ISD::FSQRT, MVT::v4f32, 37 }, // sqrtps
2044     { ISD::FSQRT, MVT::f64,   34 }, // sqrtsd
2045     { ISD::FSQRT, MVT::v2f64, 67 }, // sqrtpd
2046   };
2047   static const CostTblEntry SLMCostTbl[] = {
2048     { ISD::FSQRT, MVT::f32,   20 }, // sqrtss
2049     { ISD::FSQRT, MVT::v4f32, 40 }, // sqrtps
2050     { ISD::FSQRT, MVT::f64,   35 }, // sqrtsd
2051     { ISD::FSQRT, MVT::v2f64, 70 }, // sqrtpd
2052   };
2053   static const CostTblEntry SSE42CostTbl[] = {
2054     { ISD::USUBSAT,    MVT::v4i32,   2 }, // pmaxud + psubd
2055     { ISD::UADDSAT,    MVT::v4i32,   3 }, // not + pminud + paddd
2056     { ISD::FSQRT,      MVT::f32,    18 }, // Nehalem from http://www.agner.org/
2057     { ISD::FSQRT,      MVT::v4f32,  18 }, // Nehalem from http://www.agner.org/
2058   };
2059   static const CostTblEntry SSSE3CostTbl[] = {
2060     { ISD::BITREVERSE, MVT::v2i64,   5 },
2061     { ISD::BITREVERSE, MVT::v4i32,   5 },
2062     { ISD::BITREVERSE, MVT::v8i16,   5 },
2063     { ISD::BITREVERSE, MVT::v16i8,   5 },
2064     { ISD::BSWAP,      MVT::v2i64,   1 },
2065     { ISD::BSWAP,      MVT::v4i32,   1 },
2066     { ISD::BSWAP,      MVT::v8i16,   1 },
2067     { ISD::CTLZ,       MVT::v2i64,  23 },
2068     { ISD::CTLZ,       MVT::v4i32,  18 },
2069     { ISD::CTLZ,       MVT::v8i16,  14 },
2070     { ISD::CTLZ,       MVT::v16i8,   9 },
2071     { ISD::CTPOP,      MVT::v2i64,   7 },
2072     { ISD::CTPOP,      MVT::v4i32,  11 },
2073     { ISD::CTPOP,      MVT::v8i16,   9 },
2074     { ISD::CTPOP,      MVT::v16i8,   6 },
2075     { ISD::CTTZ,       MVT::v2i64,  10 },
2076     { ISD::CTTZ,       MVT::v4i32,  14 },
2077     { ISD::CTTZ,       MVT::v8i16,  12 },
2078     { ISD::CTTZ,       MVT::v16i8,   9 }
2079   };
2080   static const CostTblEntry SSE2CostTbl[] = {
2081     { ISD::BITREVERSE, MVT::v2i64,  29 },
2082     { ISD::BITREVERSE, MVT::v4i32,  27 },
2083     { ISD::BITREVERSE, MVT::v8i16,  27 },
2084     { ISD::BITREVERSE, MVT::v16i8,  20 },
2085     { ISD::BSWAP,      MVT::v2i64,   7 },
2086     { ISD::BSWAP,      MVT::v4i32,   7 },
2087     { ISD::BSWAP,      MVT::v8i16,   7 },
2088     { ISD::CTLZ,       MVT::v2i64,  25 },
2089     { ISD::CTLZ,       MVT::v4i32,  26 },
2090     { ISD::CTLZ,       MVT::v8i16,  20 },
2091     { ISD::CTLZ,       MVT::v16i8,  17 },
2092     { ISD::CTPOP,      MVT::v2i64,  12 },
2093     { ISD::CTPOP,      MVT::v4i32,  15 },
2094     { ISD::CTPOP,      MVT::v8i16,  13 },
2095     { ISD::CTPOP,      MVT::v16i8,  10 },
2096     { ISD::CTTZ,       MVT::v2i64,  14 },
2097     { ISD::CTTZ,       MVT::v4i32,  18 },
2098     { ISD::CTTZ,       MVT::v8i16,  16 },
2099     { ISD::CTTZ,       MVT::v16i8,  13 },
2100     { ISD::SADDSAT,    MVT::v8i16,   1 },
2101     { ISD::SADDSAT,    MVT::v16i8,   1 },
2102     { ISD::SSUBSAT,    MVT::v8i16,   1 },
2103     { ISD::SSUBSAT,    MVT::v16i8,   1 },
2104     { ISD::UADDSAT,    MVT::v8i16,   1 },
2105     { ISD::UADDSAT,    MVT::v16i8,   1 },
2106     { ISD::USUBSAT,    MVT::v8i16,   1 },
2107     { ISD::USUBSAT,    MVT::v16i8,   1 },
2108     { ISD::FSQRT,      MVT::f64,    32 }, // Nehalem from http://www.agner.org/
2109     { ISD::FSQRT,      MVT::v2f64,  32 }, // Nehalem from http://www.agner.org/
2110   };
2111   static const CostTblEntry SSE1CostTbl[] = {
2112     { ISD::FSQRT,      MVT::f32,    28 }, // Pentium III from http://www.agner.org/
2113     { ISD::FSQRT,      MVT::v4f32,  56 }, // Pentium III from http://www.agner.org/
2114   };
2115   static const CostTblEntry LZCNT64CostTbl[] = { // 64-bit targets
2116     { ISD::CTLZ,       MVT::i64,     1 },
2117   };
2118   static const CostTblEntry LZCNT32CostTbl[] = { // 32 or 64-bit targets
2119     { ISD::CTLZ,       MVT::i32,     1 },
2120     { ISD::CTLZ,       MVT::i16,     1 },
2121     { ISD::CTLZ,       MVT::i8,      1 },
2122   };
2123   static const CostTblEntry POPCNT64CostTbl[] = { // 64-bit targets
2124     { ISD::CTPOP,      MVT::i64,     1 },
2125   };
2126   static const CostTblEntry POPCNT32CostTbl[] = { // 32 or 64-bit targets
2127     { ISD::CTPOP,      MVT::i32,     1 },
2128     { ISD::CTPOP,      MVT::i16,     1 },
2129     { ISD::CTPOP,      MVT::i8,      1 },
2130   };
2131   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
2132     { ISD::BITREVERSE, MVT::i64,    14 },
2133     { ISD::CTLZ,       MVT::i64,     4 }, // BSR+XOR or BSR+XOR+CMOV
2134     { ISD::CTPOP,      MVT::i64,    10 },
2135     { ISD::SADDO,      MVT::i64,     1 },
2136     { ISD::UADDO,      MVT::i64,     1 },
2137   };
2138   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
2139     { ISD::BITREVERSE, MVT::i32,    14 },
2140     { ISD::BITREVERSE, MVT::i16,    14 },
2141     { ISD::BITREVERSE, MVT::i8,     11 },
2142     { ISD::CTLZ,       MVT::i32,     4 }, // BSR+XOR or BSR+XOR+CMOV
2143     { ISD::CTLZ,       MVT::i16,     4 }, // BSR+XOR or BSR+XOR+CMOV
2144     { ISD::CTLZ,       MVT::i8,      4 }, // BSR+XOR or BSR+XOR+CMOV
2145     { ISD::CTPOP,      MVT::i32,     8 },
2146     { ISD::CTPOP,      MVT::i16,     9 },
2147     { ISD::CTPOP,      MVT::i8,      7 },
2148     { ISD::SADDO,      MVT::i32,     1 },
2149     { ISD::SADDO,      MVT::i16,     1 },
2150     { ISD::SADDO,      MVT::i8,      1 },
2151     { ISD::UADDO,      MVT::i32,     1 },
2152     { ISD::UADDO,      MVT::i16,     1 },
2153     { ISD::UADDO,      MVT::i8,      1 },
2154   };
2155 
2156   Type *OpTy = RetTy;
2157   unsigned ISD = ISD::DELETED_NODE;
2158   switch (IID) {
2159   default:
2160     break;
2161   case Intrinsic::bitreverse:
2162     ISD = ISD::BITREVERSE;
2163     break;
2164   case Intrinsic::bswap:
2165     ISD = ISD::BSWAP;
2166     break;
2167   case Intrinsic::ctlz:
2168     ISD = ISD::CTLZ;
2169     break;
2170   case Intrinsic::ctpop:
2171     ISD = ISD::CTPOP;
2172     break;
2173   case Intrinsic::cttz:
2174     ISD = ISD::CTTZ;
2175     break;
2176   case Intrinsic::sadd_sat:
2177     ISD = ISD::SADDSAT;
2178     break;
2179   case Intrinsic::ssub_sat:
2180     ISD = ISD::SSUBSAT;
2181     break;
2182   case Intrinsic::uadd_sat:
2183     ISD = ISD::UADDSAT;
2184     break;
2185   case Intrinsic::usub_sat:
2186     ISD = ISD::USUBSAT;
2187     break;
2188   case Intrinsic::sqrt:
2189     ISD = ISD::FSQRT;
2190     break;
2191   case Intrinsic::sadd_with_overflow:
2192   case Intrinsic::ssub_with_overflow:
2193     // SSUBO has same costs so don't duplicate.
2194     ISD = ISD::SADDO;
2195     OpTy = RetTy->getContainedType(0);
2196     break;
2197   case Intrinsic::uadd_with_overflow:
2198   case Intrinsic::usub_with_overflow:
2199     // USUBO has same costs so don't duplicate.
2200     ISD = ISD::UADDO;
2201     OpTy = RetTy->getContainedType(0);
2202     break;
2203   }
2204 
2205   if (ISD != ISD::DELETED_NODE) {
2206     // Legalize the type.
2207     std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, OpTy);
2208     MVT MTy = LT.second;
2209 
2210     // Attempt to lookup cost.
2211     if (ST->useGLMDivSqrtCosts())
2212       if (const auto *Entry = CostTableLookup(GLMCostTbl, ISD, MTy))
2213         return LT.first * Entry->Cost;
2214 
2215     if (ST->isSLM())
2216       if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy))
2217         return LT.first * Entry->Cost;
2218 
2219     if (ST->hasCDI())
2220       if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy))
2221         return LT.first * Entry->Cost;
2222 
2223     if (ST->hasBWI())
2224       if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
2225         return LT.first * Entry->Cost;
2226 
2227     if (ST->hasAVX512())
2228       if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
2229         return LT.first * Entry->Cost;
2230 
2231     if (ST->hasXOP())
2232       if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
2233         return LT.first * Entry->Cost;
2234 
2235     if (ST->hasAVX2())
2236       if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
2237         return LT.first * Entry->Cost;
2238 
2239     if (ST->hasAVX())
2240       if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
2241         return LT.first * Entry->Cost;
2242 
2243     if (ST->hasSSE42())
2244       if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
2245         return LT.first * Entry->Cost;
2246 
2247     if (ST->hasSSSE3())
2248       if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy))
2249         return LT.first * Entry->Cost;
2250 
2251     if (ST->hasSSE2())
2252       if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
2253         return LT.first * Entry->Cost;
2254 
2255     if (ST->hasSSE1())
2256       if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
2257         return LT.first * Entry->Cost;
2258 
2259     if (ST->hasLZCNT()) {
2260       if (ST->is64Bit())
2261         if (const auto *Entry = CostTableLookup(LZCNT64CostTbl, ISD, MTy))
2262           return LT.first * Entry->Cost;
2263 
2264       if (const auto *Entry = CostTableLookup(LZCNT32CostTbl, ISD, MTy))
2265         return LT.first * Entry->Cost;
2266     }
2267 
2268     if (ST->hasPOPCNT()) {
2269       if (ST->is64Bit())
2270         if (const auto *Entry = CostTableLookup(POPCNT64CostTbl, ISD, MTy))
2271           return LT.first * Entry->Cost;
2272 
2273       if (const auto *Entry = CostTableLookup(POPCNT32CostTbl, ISD, MTy))
2274         return LT.first * Entry->Cost;
2275     }
2276 
2277     // TODO - add BMI (TZCNT) scalar handling
2278 
2279     if (ST->is64Bit())
2280       if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
2281         return LT.first * Entry->Cost;
2282 
2283     if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
2284       return LT.first * Entry->Cost;
2285   }
2286 
2287   return BaseT::getIntrinsicInstrCost(IID, RetTy, Tys, FMF, ScalarizationCostPassed);
2288 }
2289 
2290 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
2291                                       ArrayRef<Value *> Args, FastMathFlags FMF,
2292                                       unsigned VF) {
2293   static const CostTblEntry AVX512CostTbl[] = {
2294     { ISD::ROTL,       MVT::v8i64,   1 },
2295     { ISD::ROTL,       MVT::v4i64,   1 },
2296     { ISD::ROTL,       MVT::v2i64,   1 },
2297     { ISD::ROTL,       MVT::v16i32,  1 },
2298     { ISD::ROTL,       MVT::v8i32,   1 },
2299     { ISD::ROTL,       MVT::v4i32,   1 },
2300     { ISD::ROTR,       MVT::v8i64,   1 },
2301     { ISD::ROTR,       MVT::v4i64,   1 },
2302     { ISD::ROTR,       MVT::v2i64,   1 },
2303     { ISD::ROTR,       MVT::v16i32,  1 },
2304     { ISD::ROTR,       MVT::v8i32,   1 },
2305     { ISD::ROTR,       MVT::v4i32,   1 }
2306   };
2307   // XOP: ROTL = VPROT(X,Y), ROTR = VPROT(X,SUB(0,Y))
2308   static const CostTblEntry XOPCostTbl[] = {
2309     { ISD::ROTL,       MVT::v4i64,   4 },
2310     { ISD::ROTL,       MVT::v8i32,   4 },
2311     { ISD::ROTL,       MVT::v16i16,  4 },
2312     { ISD::ROTL,       MVT::v32i8,   4 },
2313     { ISD::ROTL,       MVT::v2i64,   1 },
2314     { ISD::ROTL,       MVT::v4i32,   1 },
2315     { ISD::ROTL,       MVT::v8i16,   1 },
2316     { ISD::ROTL,       MVT::v16i8,   1 },
2317     { ISD::ROTR,       MVT::v4i64,   6 },
2318     { ISD::ROTR,       MVT::v8i32,   6 },
2319     { ISD::ROTR,       MVT::v16i16,  6 },
2320     { ISD::ROTR,       MVT::v32i8,   6 },
2321     { ISD::ROTR,       MVT::v2i64,   2 },
2322     { ISD::ROTR,       MVT::v4i32,   2 },
2323     { ISD::ROTR,       MVT::v8i16,   2 },
2324     { ISD::ROTR,       MVT::v16i8,   2 }
2325   };
2326   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
2327     { ISD::ROTL,       MVT::i64,     1 },
2328     { ISD::ROTR,       MVT::i64,     1 },
2329     { ISD::FSHL,       MVT::i64,     4 }
2330   };
2331   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
2332     { ISD::ROTL,       MVT::i32,     1 },
2333     { ISD::ROTL,       MVT::i16,     1 },
2334     { ISD::ROTL,       MVT::i8,      1 },
2335     { ISD::ROTR,       MVT::i32,     1 },
2336     { ISD::ROTR,       MVT::i16,     1 },
2337     { ISD::ROTR,       MVT::i8,      1 },
2338     { ISD::FSHL,       MVT::i32,     4 },
2339     { ISD::FSHL,       MVT::i16,     4 },
2340     { ISD::FSHL,       MVT::i8,      4 }
2341   };
2342 
2343   unsigned ISD = ISD::DELETED_NODE;
2344   switch (IID) {
2345   default:
2346     break;
2347   case Intrinsic::fshl:
2348     ISD = ISD::FSHL;
2349     if (Args[0] == Args[1])
2350       ISD = ISD::ROTL;
2351     break;
2352   case Intrinsic::fshr:
2353     // FSHR has same costs so don't duplicate.
2354     ISD = ISD::FSHL;
2355     if (Args[0] == Args[1])
2356       ISD = ISD::ROTR;
2357     break;
2358   }
2359 
2360   if (ISD != ISD::DELETED_NODE) {
2361     // Legalize the type.
2362     std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
2363     MVT MTy = LT.second;
2364 
2365     // Attempt to lookup cost.
2366     if (ST->hasAVX512())
2367       if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
2368         return LT.first * Entry->Cost;
2369 
2370     if (ST->hasXOP())
2371       if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
2372         return LT.first * Entry->Cost;
2373 
2374     if (ST->is64Bit())
2375       if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
2376         return LT.first * Entry->Cost;
2377 
2378     if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
2379       return LT.first * Entry->Cost;
2380   }
2381 
2382   return BaseT::getIntrinsicInstrCost(IID, RetTy, Args, FMF, VF);
2383 }
2384 
2385 int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
2386   static const CostTblEntry SLMCostTbl[] = {
2387      { ISD::EXTRACT_VECTOR_ELT,       MVT::i8,      4 },
2388      { ISD::EXTRACT_VECTOR_ELT,       MVT::i16,     4 },
2389      { ISD::EXTRACT_VECTOR_ELT,       MVT::i32,     4 },
2390      { ISD::EXTRACT_VECTOR_ELT,       MVT::i64,     7 }
2391    };
2392 
2393   assert(Val->isVectorTy() && "This must be a vector type");
2394 
2395   Type *ScalarType = Val->getScalarType();
2396 
2397   if (Index != -1U) {
2398     // Legalize the type.
2399     std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Val);
2400 
2401     // This type is legalized to a scalar type.
2402     if (!LT.second.isVector())
2403       return 0;
2404 
2405     // The type may be split. Normalize the index to the new type.
2406     unsigned Width = LT.second.getVectorNumElements();
2407     Index = Index % Width;
2408 
2409     if (Index == 0) {
2410       // Floating point scalars are already located in index #0.
2411       if (ScalarType->isFloatingPointTy())
2412         return 0;
2413 
2414       // Assume movd/movq XMM <-> GPR is relatively cheap on all targets.
2415       if (ScalarType->isIntegerTy())
2416         return 1;
2417     }
2418 
2419     int ISD = TLI->InstructionOpcodeToISD(Opcode);
2420     assert(ISD && "Unexpected vector opcode");
2421     MVT MScalarTy = LT.second.getScalarType();
2422     if (ST->isSLM())
2423       if (auto *Entry = CostTableLookup(SLMCostTbl, ISD, MScalarTy))
2424         return LT.first * Entry->Cost;
2425   }
2426 
2427   // Add to the base cost if we know that the extracted element of a vector is
2428   // destined to be moved to and used in the integer register file.
2429   int RegisterFileMoveCost = 0;
2430   if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy())
2431     RegisterFileMoveCost = 1;
2432 
2433   return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost;
2434 }
2435 
2436 int X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
2437                                 MaybeAlign Alignment, unsigned AddressSpace,
2438                                 const Instruction *I) {
2439   // Handle non-power-of-two vectors such as <3 x float>
2440   if (VectorType *VTy = dyn_cast<VectorType>(Src)) {
2441     unsigned NumElem = VTy->getVectorNumElements();
2442 
2443     // Handle a few common cases:
2444     // <3 x float>
2445     if (NumElem == 3 && VTy->getScalarSizeInBits() == 32)
2446       // Cost = 64 bit store + extract + 32 bit store.
2447       return 3;
2448 
2449     // <3 x double>
2450     if (NumElem == 3 && VTy->getScalarSizeInBits() == 64)
2451       // Cost = 128 bit store + unpack + 64 bit store.
2452       return 3;
2453 
2454     // Assume that all other non-power-of-two numbers are scalarized.
2455     if (!isPowerOf2_32(NumElem)) {
2456       int Cost = BaseT::getMemoryOpCost(Opcode, VTy->getScalarType(), Alignment,
2457                                         AddressSpace);
2458       int SplitCost = getScalarizationOverhead(Src, Opcode == Instruction::Load,
2459                                                Opcode == Instruction::Store);
2460       return NumElem * Cost + SplitCost;
2461     }
2462   }
2463 
2464   // Legalize the type.
2465   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
2466   assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
2467          "Invalid Opcode");
2468 
2469   // Each load/store unit costs 1.
2470   int Cost = LT.first * 1;
2471 
2472   // This isn't exactly right. We're using slow unaligned 32-byte accesses as a
2473   // proxy for a double-pumped AVX memory interface such as on Sandybridge.
2474   if (LT.second.getStoreSize() == 32 && ST->isUnalignedMem32Slow())
2475     Cost *= 2;
2476 
2477   return Cost;
2478 }
2479 
2480 int X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy,
2481                                       unsigned Alignment,
2482                                       unsigned AddressSpace) {
2483   bool IsLoad = (Instruction::Load == Opcode);
2484   bool IsStore = (Instruction::Store == Opcode);
2485 
2486   VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy);
2487   if (!SrcVTy)
2488     // To calculate scalar take the regular cost, without mask
2489     return getMemoryOpCost(Opcode, SrcTy, MaybeAlign(Alignment), AddressSpace);
2490 
2491   unsigned NumElem = SrcVTy->getVectorNumElements();
2492   VectorType *MaskTy =
2493       VectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem);
2494   if ((IsLoad && !isLegalMaskedLoad(SrcVTy, MaybeAlign(Alignment))) ||
2495       (IsStore && !isLegalMaskedStore(SrcVTy, MaybeAlign(Alignment))) ||
2496       !isPowerOf2_32(NumElem)) {
2497     // Scalarization
2498     int MaskSplitCost = getScalarizationOverhead(MaskTy, false, true);
2499     int ScalarCompareCost = getCmpSelInstrCost(
2500         Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr);
2501     int BranchCost = getCFInstrCost(Instruction::Br);
2502     int MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost);
2503 
2504     int ValueSplitCost = getScalarizationOverhead(SrcVTy, IsLoad, IsStore);
2505     int MemopCost =
2506         NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
2507                                          MaybeAlign(Alignment), AddressSpace);
2508     return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost;
2509   }
2510 
2511   // Legalize the type.
2512   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy);
2513   auto VT = TLI->getValueType(DL, SrcVTy);
2514   int Cost = 0;
2515   if (VT.isSimple() && LT.second != VT.getSimpleVT() &&
2516       LT.second.getVectorNumElements() == NumElem)
2517     // Promotion requires expand/truncate for data and a shuffle for mask.
2518     Cost += getShuffleCost(TTI::SK_PermuteTwoSrc, SrcVTy, 0, nullptr) +
2519             getShuffleCost(TTI::SK_PermuteTwoSrc, MaskTy, 0, nullptr);
2520 
2521   else if (LT.second.getVectorNumElements() > NumElem) {
2522     VectorType *NewMaskTy = VectorType::get(MaskTy->getVectorElementType(),
2523                                             LT.second.getVectorNumElements());
2524     // Expanding requires fill mask with zeroes
2525     Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, 0, MaskTy);
2526   }
2527 
2528   // Pre-AVX512 - each maskmov load costs 2 + store costs ~8.
2529   if (!ST->hasAVX512())
2530     return Cost + LT.first * (IsLoad ? 2 : 8);
2531 
2532   // AVX-512 masked load/store is cheapper
2533   return Cost + LT.first;
2534 }
2535 
2536 int X86TTIImpl::getAddressComputationCost(Type *Ty, ScalarEvolution *SE,
2537                                           const SCEV *Ptr) {
2538   // Address computations in vectorized code with non-consecutive addresses will
2539   // likely result in more instructions compared to scalar code where the
2540   // computation can more often be merged into the index mode. The resulting
2541   // extra micro-ops can significantly decrease throughput.
2542   const unsigned NumVectorInstToHideOverhead = 10;
2543 
2544   // Cost modeling of Strided Access Computation is hidden by the indexing
2545   // modes of X86 regardless of the stride value. We dont believe that there
2546   // is a difference between constant strided access in gerenal and constant
2547   // strided value which is less than or equal to 64.
2548   // Even in the case of (loop invariant) stride whose value is not known at
2549   // compile time, the address computation will not incur more than one extra
2550   // ADD instruction.
2551   if (Ty->isVectorTy() && SE) {
2552     if (!BaseT::isStridedAccess(Ptr))
2553       return NumVectorInstToHideOverhead;
2554     if (!BaseT::getConstantStrideStep(SE, Ptr))
2555       return 1;
2556   }
2557 
2558   return BaseT::getAddressComputationCost(Ty, SE, Ptr);
2559 }
2560 
2561 int X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, Type *ValTy,
2562                                            bool IsPairwise) {
2563   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
2564   // and make it as the cost.
2565 
2566   static const CostTblEntry SLMCostTblPairWise[] = {
2567     { ISD::FADD,  MVT::v2f64,   3 },
2568     { ISD::ADD,   MVT::v2i64,   5 },
2569   };
2570 
2571   static const CostTblEntry SSE2CostTblPairWise[] = {
2572     { ISD::FADD,  MVT::v2f64,   2 },
2573     { ISD::FADD,  MVT::v4f32,   4 },
2574     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
2575     { ISD::ADD,   MVT::v2i32,   2 }, // FIXME: chosen to be less than v4i32.
2576     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.5".
2577     { ISD::ADD,   MVT::v2i16,   3 }, // FIXME: chosen to be less than v4i16
2578     { ISD::ADD,   MVT::v4i16,   4 }, // FIXME: chosen to be less than v8i16
2579     { ISD::ADD,   MVT::v8i16,   5 },
2580     { ISD::ADD,   MVT::v2i8,    2 },
2581     { ISD::ADD,   MVT::v4i8,    2 },
2582     { ISD::ADD,   MVT::v8i8,    2 },
2583     { ISD::ADD,   MVT::v16i8,   3 },
2584   };
2585 
2586   static const CostTblEntry AVX1CostTblPairWise[] = {
2587     { ISD::FADD,  MVT::v4f64,   5 },
2588     { ISD::FADD,  MVT::v8f32,   7 },
2589     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
2590     { ISD::ADD,   MVT::v4i64,   5 },      // The data reported by the IACA tool is "4.8".
2591     { ISD::ADD,   MVT::v8i32,   5 },
2592     { ISD::ADD,   MVT::v16i16,  6 },
2593     { ISD::ADD,   MVT::v32i8,   4 },
2594   };
2595 
2596   static const CostTblEntry SLMCostTblNoPairWise[] = {
2597     { ISD::FADD,  MVT::v2f64,   3 },
2598     { ISD::ADD,   MVT::v2i64,   5 },
2599   };
2600 
2601   static const CostTblEntry SSE2CostTblNoPairWise[] = {
2602     { ISD::FADD,  MVT::v2f64,   2 },
2603     { ISD::FADD,  MVT::v4f32,   4 },
2604     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
2605     { ISD::ADD,   MVT::v2i32,   2 }, // FIXME: chosen to be less than v4i32
2606     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.3".
2607     { ISD::ADD,   MVT::v2i16,   2 },      // The data reported by the IACA tool is "4.3".
2608     { ISD::ADD,   MVT::v4i16,   3 },      // The data reported by the IACA tool is "4.3".
2609     { ISD::ADD,   MVT::v8i16,   4 },      // The data reported by the IACA tool is "4.3".
2610     { ISD::ADD,   MVT::v2i8,    2 },
2611     { ISD::ADD,   MVT::v4i8,    2 },
2612     { ISD::ADD,   MVT::v8i8,    2 },
2613     { ISD::ADD,   MVT::v16i8,   3 },
2614   };
2615 
2616   static const CostTblEntry AVX1CostTblNoPairWise[] = {
2617     { ISD::FADD,  MVT::v4f64,   3 },
2618     { ISD::FADD,  MVT::v4f32,   3 },
2619     { ISD::FADD,  MVT::v8f32,   4 },
2620     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
2621     { ISD::ADD,   MVT::v4i64,   3 },
2622     { ISD::ADD,   MVT::v8i32,   5 },
2623     { ISD::ADD,   MVT::v16i16,  5 },
2624     { ISD::ADD,   MVT::v32i8,   4 },
2625   };
2626 
2627   int ISD = TLI->InstructionOpcodeToISD(Opcode);
2628   assert(ISD && "Invalid opcode");
2629 
2630   // Before legalizing the type, give a chance to look up illegal narrow types
2631   // in the table.
2632   // FIXME: Is there a better way to do this?
2633   EVT VT = TLI->getValueType(DL, ValTy);
2634   if (VT.isSimple()) {
2635     MVT MTy = VT.getSimpleVT();
2636     if (IsPairwise) {
2637       if (ST->isSLM())
2638         if (const auto *Entry = CostTableLookup(SLMCostTblPairWise, ISD, MTy))
2639           return Entry->Cost;
2640 
2641       if (ST->hasAVX())
2642         if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy))
2643           return Entry->Cost;
2644 
2645       if (ST->hasSSE2())
2646         if (const auto *Entry = CostTableLookup(SSE2CostTblPairWise, ISD, MTy))
2647           return Entry->Cost;
2648     } else {
2649       if (ST->isSLM())
2650         if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy))
2651           return Entry->Cost;
2652 
2653       if (ST->hasAVX())
2654         if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
2655           return Entry->Cost;
2656 
2657       if (ST->hasSSE2())
2658         if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
2659           return Entry->Cost;
2660     }
2661   }
2662 
2663   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
2664 
2665   MVT MTy = LT.second;
2666 
2667   if (IsPairwise) {
2668     if (ST->isSLM())
2669       if (const auto *Entry = CostTableLookup(SLMCostTblPairWise, ISD, MTy))
2670         return LT.first * Entry->Cost;
2671 
2672     if (ST->hasAVX())
2673       if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy))
2674         return LT.first * Entry->Cost;
2675 
2676     if (ST->hasSSE2())
2677       if (const auto *Entry = CostTableLookup(SSE2CostTblPairWise, ISD, MTy))
2678         return LT.first * Entry->Cost;
2679   } else {
2680     if (ST->isSLM())
2681       if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy))
2682         return LT.first * Entry->Cost;
2683 
2684     if (ST->hasAVX())
2685       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
2686         return LT.first * Entry->Cost;
2687 
2688     if (ST->hasSSE2())
2689       if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
2690         return LT.first * Entry->Cost;
2691   }
2692 
2693   // FIXME: These assume a naive kshift+binop lowering, which is probably
2694   // conservative in most cases.
2695   // FIXME: This doesn't cost large types like v128i1 correctly.
2696   static const CostTblEntry AVX512BoolReduction[] = {
2697     { ISD::AND,  MVT::v2i1,   3 },
2698     { ISD::AND,  MVT::v4i1,   5 },
2699     { ISD::AND,  MVT::v8i1,   7 },
2700     { ISD::AND,  MVT::v16i1,  9 },
2701     { ISD::AND,  MVT::v32i1, 11 },
2702     { ISD::AND,  MVT::v64i1, 13 },
2703     { ISD::OR,   MVT::v2i1,   3 },
2704     { ISD::OR,   MVT::v4i1,   5 },
2705     { ISD::OR,   MVT::v8i1,   7 },
2706     { ISD::OR,   MVT::v16i1,  9 },
2707     { ISD::OR,   MVT::v32i1, 11 },
2708     { ISD::OR,   MVT::v64i1, 13 },
2709   };
2710 
2711   static const CostTblEntry AVX2BoolReduction[] = {
2712     { ISD::AND,  MVT::v16i16,  2 }, // vpmovmskb + cmp
2713     { ISD::AND,  MVT::v32i8,   2 }, // vpmovmskb + cmp
2714     { ISD::OR,   MVT::v16i16,  2 }, // vpmovmskb + cmp
2715     { ISD::OR,   MVT::v32i8,   2 }, // vpmovmskb + cmp
2716   };
2717 
2718   static const CostTblEntry AVX1BoolReduction[] = {
2719     { ISD::AND,  MVT::v4i64,   2 }, // vmovmskpd + cmp
2720     { ISD::AND,  MVT::v8i32,   2 }, // vmovmskps + cmp
2721     { ISD::AND,  MVT::v16i16,  4 }, // vextractf128 + vpand + vpmovmskb + cmp
2722     { ISD::AND,  MVT::v32i8,   4 }, // vextractf128 + vpand + vpmovmskb + cmp
2723     { ISD::OR,   MVT::v4i64,   2 }, // vmovmskpd + cmp
2724     { ISD::OR,   MVT::v8i32,   2 }, // vmovmskps + cmp
2725     { ISD::OR,   MVT::v16i16,  4 }, // vextractf128 + vpor + vpmovmskb + cmp
2726     { ISD::OR,   MVT::v32i8,   4 }, // vextractf128 + vpor + vpmovmskb + cmp
2727   };
2728 
2729   static const CostTblEntry SSE2BoolReduction[] = {
2730     { ISD::AND,  MVT::v2i64,   2 }, // movmskpd + cmp
2731     { ISD::AND,  MVT::v4i32,   2 }, // movmskps + cmp
2732     { ISD::AND,  MVT::v8i16,   2 }, // pmovmskb + cmp
2733     { ISD::AND,  MVT::v16i8,   2 }, // pmovmskb + cmp
2734     { ISD::OR,   MVT::v2i64,   2 }, // movmskpd + cmp
2735     { ISD::OR,   MVT::v4i32,   2 }, // movmskps + cmp
2736     { ISD::OR,   MVT::v8i16,   2 }, // pmovmskb + cmp
2737     { ISD::OR,   MVT::v16i8,   2 }, // pmovmskb + cmp
2738   };
2739 
2740   // Handle bool allof/anyof patterns.
2741   if (!IsPairwise && ValTy->getVectorElementType()->isIntegerTy(1)) {
2742     if (ST->hasAVX512())
2743       if (const auto *Entry = CostTableLookup(AVX512BoolReduction, ISD, MTy))
2744         return LT.first * Entry->Cost;
2745     if (ST->hasAVX2())
2746       if (const auto *Entry = CostTableLookup(AVX2BoolReduction, ISD, MTy))
2747         return LT.first * Entry->Cost;
2748     if (ST->hasAVX())
2749       if (const auto *Entry = CostTableLookup(AVX1BoolReduction, ISD, MTy))
2750         return LT.first * Entry->Cost;
2751     if (ST->hasSSE2())
2752       if (const auto *Entry = CostTableLookup(SSE2BoolReduction, ISD, MTy))
2753         return LT.first * Entry->Cost;
2754   }
2755 
2756   return BaseT::getArithmeticReductionCost(Opcode, ValTy, IsPairwise);
2757 }
2758 
2759 int X86TTIImpl::getMinMaxReductionCost(Type *ValTy, Type *CondTy,
2760                                        bool IsPairwise, bool IsUnsigned) {
2761   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
2762 
2763   MVT MTy = LT.second;
2764 
2765   int ISD;
2766   if (ValTy->isIntOrIntVectorTy()) {
2767     ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN;
2768   } else {
2769     assert(ValTy->isFPOrFPVectorTy() &&
2770            "Expected float point or integer vector type.");
2771     ISD = ISD::FMINNUM;
2772   }
2773 
2774   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
2775   // and make it as the cost.
2776 
2777   static const CostTblEntry SSE1CostTblPairWise[] = {
2778       {ISD::FMINNUM, MVT::v4f32, 4},
2779   };
2780 
2781   static const CostTblEntry SSE2CostTblPairWise[] = {
2782       {ISD::FMINNUM, MVT::v2f64, 3},
2783       {ISD::SMIN, MVT::v2i64, 6},
2784       {ISD::UMIN, MVT::v2i64, 8},
2785       {ISD::SMIN, MVT::v4i32, 6},
2786       {ISD::UMIN, MVT::v4i32, 8},
2787       {ISD::SMIN, MVT::v8i16, 4},
2788       {ISD::UMIN, MVT::v8i16, 6},
2789       {ISD::SMIN, MVT::v16i8, 8},
2790       {ISD::UMIN, MVT::v16i8, 6},
2791   };
2792 
2793   static const CostTblEntry SSE41CostTblPairWise[] = {
2794       {ISD::FMINNUM, MVT::v4f32, 2},
2795       {ISD::SMIN, MVT::v2i64, 9},
2796       {ISD::UMIN, MVT::v2i64,10},
2797       {ISD::SMIN, MVT::v4i32, 1}, // The data reported by the IACA is "1.5"
2798       {ISD::UMIN, MVT::v4i32, 2}, // The data reported by the IACA is "1.8"
2799       {ISD::SMIN, MVT::v8i16, 2},
2800       {ISD::UMIN, MVT::v8i16, 2},
2801       {ISD::SMIN, MVT::v16i8, 3},
2802       {ISD::UMIN, MVT::v16i8, 3},
2803   };
2804 
2805   static const CostTblEntry SSE42CostTblPairWise[] = {
2806       {ISD::SMIN, MVT::v2i64, 7}, // The data reported by the IACA is "6.8"
2807       {ISD::UMIN, MVT::v2i64, 8}, // The data reported by the IACA is "8.6"
2808   };
2809 
2810   static const CostTblEntry AVX1CostTblPairWise[] = {
2811       {ISD::FMINNUM, MVT::v4f32, 1},
2812       {ISD::FMINNUM, MVT::v4f64, 1},
2813       {ISD::FMINNUM, MVT::v8f32, 2},
2814       {ISD::SMIN, MVT::v2i64, 3},
2815       {ISD::UMIN, MVT::v2i64, 3},
2816       {ISD::SMIN, MVT::v4i32, 1},
2817       {ISD::UMIN, MVT::v4i32, 1},
2818       {ISD::SMIN, MVT::v8i16, 1},
2819       {ISD::UMIN, MVT::v8i16, 1},
2820       {ISD::SMIN, MVT::v16i8, 2},
2821       {ISD::UMIN, MVT::v16i8, 2},
2822       {ISD::SMIN, MVT::v4i64, 7},
2823       {ISD::UMIN, MVT::v4i64, 7},
2824       {ISD::SMIN, MVT::v8i32, 3},
2825       {ISD::UMIN, MVT::v8i32, 3},
2826       {ISD::SMIN, MVT::v16i16, 3},
2827       {ISD::UMIN, MVT::v16i16, 3},
2828       {ISD::SMIN, MVT::v32i8, 3},
2829       {ISD::UMIN, MVT::v32i8, 3},
2830   };
2831 
2832   static const CostTblEntry AVX2CostTblPairWise[] = {
2833       {ISD::SMIN, MVT::v4i64, 2},
2834       {ISD::UMIN, MVT::v4i64, 2},
2835       {ISD::SMIN, MVT::v8i32, 1},
2836       {ISD::UMIN, MVT::v8i32, 1},
2837       {ISD::SMIN, MVT::v16i16, 1},
2838       {ISD::UMIN, MVT::v16i16, 1},
2839       {ISD::SMIN, MVT::v32i8, 2},
2840       {ISD::UMIN, MVT::v32i8, 2},
2841   };
2842 
2843   static const CostTblEntry AVX512CostTblPairWise[] = {
2844       {ISD::FMINNUM, MVT::v8f64, 1},
2845       {ISD::FMINNUM, MVT::v16f32, 2},
2846       {ISD::SMIN, MVT::v8i64, 2},
2847       {ISD::UMIN, MVT::v8i64, 2},
2848       {ISD::SMIN, MVT::v16i32, 1},
2849       {ISD::UMIN, MVT::v16i32, 1},
2850   };
2851 
2852   static const CostTblEntry SSE1CostTblNoPairWise[] = {
2853       {ISD::FMINNUM, MVT::v4f32, 4},
2854   };
2855 
2856   static const CostTblEntry SSE2CostTblNoPairWise[] = {
2857       {ISD::FMINNUM, MVT::v2f64, 3},
2858       {ISD::SMIN, MVT::v2i64, 6},
2859       {ISD::UMIN, MVT::v2i64, 8},
2860       {ISD::SMIN, MVT::v4i32, 6},
2861       {ISD::UMIN, MVT::v4i32, 8},
2862       {ISD::SMIN, MVT::v8i16, 4},
2863       {ISD::UMIN, MVT::v8i16, 6},
2864       {ISD::SMIN, MVT::v16i8, 8},
2865       {ISD::UMIN, MVT::v16i8, 6},
2866   };
2867 
2868   static const CostTblEntry SSE41CostTblNoPairWise[] = {
2869       {ISD::FMINNUM, MVT::v4f32, 3},
2870       {ISD::SMIN, MVT::v2i64, 9},
2871       {ISD::UMIN, MVT::v2i64,11},
2872       {ISD::SMIN, MVT::v4i32, 1}, // The data reported by the IACA is "1.5"
2873       {ISD::UMIN, MVT::v4i32, 2}, // The data reported by the IACA is "1.8"
2874       {ISD::SMIN, MVT::v8i16, 1}, // The data reported by the IACA is "1.5"
2875       {ISD::UMIN, MVT::v8i16, 2}, // The data reported by the IACA is "1.8"
2876       {ISD::SMIN, MVT::v16i8, 3},
2877       {ISD::UMIN, MVT::v16i8, 3},
2878   };
2879 
2880   static const CostTblEntry SSE42CostTblNoPairWise[] = {
2881       {ISD::SMIN, MVT::v2i64, 7}, // The data reported by the IACA is "6.8"
2882       {ISD::UMIN, MVT::v2i64, 9}, // The data reported by the IACA is "8.6"
2883   };
2884 
2885   static const CostTblEntry AVX1CostTblNoPairWise[] = {
2886       {ISD::FMINNUM, MVT::v4f32, 1},
2887       {ISD::FMINNUM, MVT::v4f64, 1},
2888       {ISD::FMINNUM, MVT::v8f32, 1},
2889       {ISD::SMIN, MVT::v2i64, 3},
2890       {ISD::UMIN, MVT::v2i64, 3},
2891       {ISD::SMIN, MVT::v4i32, 1},
2892       {ISD::UMIN, MVT::v4i32, 1},
2893       {ISD::SMIN, MVT::v8i16, 1},
2894       {ISD::UMIN, MVT::v8i16, 1},
2895       {ISD::SMIN, MVT::v16i8, 2},
2896       {ISD::UMIN, MVT::v16i8, 2},
2897       {ISD::SMIN, MVT::v4i64, 7},
2898       {ISD::UMIN, MVT::v4i64, 7},
2899       {ISD::SMIN, MVT::v8i32, 2},
2900       {ISD::UMIN, MVT::v8i32, 2},
2901       {ISD::SMIN, MVT::v16i16, 2},
2902       {ISD::UMIN, MVT::v16i16, 2},
2903       {ISD::SMIN, MVT::v32i8, 2},
2904       {ISD::UMIN, MVT::v32i8, 2},
2905   };
2906 
2907   static const CostTblEntry AVX2CostTblNoPairWise[] = {
2908       {ISD::SMIN, MVT::v4i64, 1},
2909       {ISD::UMIN, MVT::v4i64, 1},
2910       {ISD::SMIN, MVT::v8i32, 1},
2911       {ISD::UMIN, MVT::v8i32, 1},
2912       {ISD::SMIN, MVT::v16i16, 1},
2913       {ISD::UMIN, MVT::v16i16, 1},
2914       {ISD::SMIN, MVT::v32i8, 1},
2915       {ISD::UMIN, MVT::v32i8, 1},
2916   };
2917 
2918   static const CostTblEntry AVX512CostTblNoPairWise[] = {
2919       {ISD::FMINNUM, MVT::v8f64, 1},
2920       {ISD::FMINNUM, MVT::v16f32, 2},
2921       {ISD::SMIN, MVT::v8i64, 1},
2922       {ISD::UMIN, MVT::v8i64, 1},
2923       {ISD::SMIN, MVT::v16i32, 1},
2924       {ISD::UMIN, MVT::v16i32, 1},
2925   };
2926 
2927   if (IsPairwise) {
2928     if (ST->hasAVX512())
2929       if (const auto *Entry = CostTableLookup(AVX512CostTblPairWise, ISD, MTy))
2930         return LT.first * Entry->Cost;
2931 
2932     if (ST->hasAVX2())
2933       if (const auto *Entry = CostTableLookup(AVX2CostTblPairWise, ISD, MTy))
2934         return LT.first * Entry->Cost;
2935 
2936     if (ST->hasAVX())
2937       if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy))
2938         return LT.first * Entry->Cost;
2939 
2940     if (ST->hasSSE42())
2941       if (const auto *Entry = CostTableLookup(SSE42CostTblPairWise, ISD, MTy))
2942         return LT.first * Entry->Cost;
2943 
2944     if (ST->hasSSE41())
2945       if (const auto *Entry = CostTableLookup(SSE41CostTblPairWise, ISD, MTy))
2946         return LT.first * Entry->Cost;
2947 
2948     if (ST->hasSSE2())
2949       if (const auto *Entry = CostTableLookup(SSE2CostTblPairWise, ISD, MTy))
2950         return LT.first * Entry->Cost;
2951 
2952     if (ST->hasSSE1())
2953       if (const auto *Entry = CostTableLookup(SSE1CostTblPairWise, ISD, MTy))
2954         return LT.first * Entry->Cost;
2955   } else {
2956     if (ST->hasAVX512())
2957       if (const auto *Entry =
2958               CostTableLookup(AVX512CostTblNoPairWise, ISD, MTy))
2959         return LT.first * Entry->Cost;
2960 
2961     if (ST->hasAVX2())
2962       if (const auto *Entry = CostTableLookup(AVX2CostTblNoPairWise, ISD, MTy))
2963         return LT.first * Entry->Cost;
2964 
2965     if (ST->hasAVX())
2966       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
2967         return LT.first * Entry->Cost;
2968 
2969     if (ST->hasSSE42())
2970       if (const auto *Entry = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy))
2971         return LT.first * Entry->Cost;
2972 
2973     if (ST->hasSSE41())
2974       if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy))
2975         return LT.first * Entry->Cost;
2976 
2977     if (ST->hasSSE2())
2978       if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
2979         return LT.first * Entry->Cost;
2980 
2981     if (ST->hasSSE1())
2982       if (const auto *Entry = CostTableLookup(SSE1CostTblNoPairWise, ISD, MTy))
2983         return LT.first * Entry->Cost;
2984   }
2985 
2986   return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsPairwise, IsUnsigned);
2987 }
2988 
2989 /// Calculate the cost of materializing a 64-bit value. This helper
2990 /// method might only calculate a fraction of a larger immediate. Therefore it
2991 /// is valid to return a cost of ZERO.
2992 int X86TTIImpl::getIntImmCost(int64_t Val) {
2993   if (Val == 0)
2994     return TTI::TCC_Free;
2995 
2996   if (isInt<32>(Val))
2997     return TTI::TCC_Basic;
2998 
2999   return 2 * TTI::TCC_Basic;
3000 }
3001 
3002 int X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) {
3003   assert(Ty->isIntegerTy());
3004 
3005   unsigned BitSize = Ty->getPrimitiveSizeInBits();
3006   if (BitSize == 0)
3007     return ~0U;
3008 
3009   // Never hoist constants larger than 128bit, because this might lead to
3010   // incorrect code generation or assertions in codegen.
3011   // Fixme: Create a cost model for types larger than i128 once the codegen
3012   // issues have been fixed.
3013   if (BitSize > 128)
3014     return TTI::TCC_Free;
3015 
3016   if (Imm == 0)
3017     return TTI::TCC_Free;
3018 
3019   // Sign-extend all constants to a multiple of 64-bit.
3020   APInt ImmVal = Imm;
3021   if (BitSize % 64 != 0)
3022     ImmVal = Imm.sext(alignTo(BitSize, 64));
3023 
3024   // Split the constant into 64-bit chunks and calculate the cost for each
3025   // chunk.
3026   int Cost = 0;
3027   for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
3028     APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
3029     int64_t Val = Tmp.getSExtValue();
3030     Cost += getIntImmCost(Val);
3031   }
3032   // We need at least one instruction to materialize the constant.
3033   return std::max(1, Cost);
3034 }
3035 
3036 int X86TTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx, const APInt &Imm,
3037                               Type *Ty) {
3038   assert(Ty->isIntegerTy());
3039 
3040   unsigned BitSize = Ty->getPrimitiveSizeInBits();
3041   // There is no cost model for constants with a bit size of 0. Return TCC_Free
3042   // here, so that constant hoisting will ignore this constant.
3043   if (BitSize == 0)
3044     return TTI::TCC_Free;
3045 
3046   unsigned ImmIdx = ~0U;
3047   switch (Opcode) {
3048   default:
3049     return TTI::TCC_Free;
3050   case Instruction::GetElementPtr:
3051     // Always hoist the base address of a GetElementPtr. This prevents the
3052     // creation of new constants for every base constant that gets constant
3053     // folded with the offset.
3054     if (Idx == 0)
3055       return 2 * TTI::TCC_Basic;
3056     return TTI::TCC_Free;
3057   case Instruction::Store:
3058     ImmIdx = 0;
3059     break;
3060   case Instruction::ICmp:
3061     // This is an imperfect hack to prevent constant hoisting of
3062     // compares that might be trying to check if a 64-bit value fits in
3063     // 32-bits. The backend can optimize these cases using a right shift by 32.
3064     // Ideally we would check the compare predicate here. There also other
3065     // similar immediates the backend can use shifts for.
3066     if (Idx == 1 && Imm.getBitWidth() == 64) {
3067       uint64_t ImmVal = Imm.getZExtValue();
3068       if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff)
3069         return TTI::TCC_Free;
3070     }
3071     ImmIdx = 1;
3072     break;
3073   case Instruction::And:
3074     // We support 64-bit ANDs with immediates with 32-bits of leading zeroes
3075     // by using a 32-bit operation with implicit zero extension. Detect such
3076     // immediates here as the normal path expects bit 31 to be sign extended.
3077     if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue()))
3078       return TTI::TCC_Free;
3079     ImmIdx = 1;
3080     break;
3081   case Instruction::Add:
3082   case Instruction::Sub:
3083     // For add/sub, we can use the opposite instruction for INT32_MIN.
3084     if (Idx == 1 && Imm.getBitWidth() == 64 && Imm.getZExtValue() == 0x80000000)
3085       return TTI::TCC_Free;
3086     ImmIdx = 1;
3087     break;
3088   case Instruction::UDiv:
3089   case Instruction::SDiv:
3090   case Instruction::URem:
3091   case Instruction::SRem:
3092     // Division by constant is typically expanded later into a different
3093     // instruction sequence. This completely changes the constants.
3094     // Report them as "free" to stop ConstantHoist from marking them as opaque.
3095     return TTI::TCC_Free;
3096   case Instruction::Mul:
3097   case Instruction::Or:
3098   case Instruction::Xor:
3099     ImmIdx = 1;
3100     break;
3101   // Always return TCC_Free for the shift value of a shift instruction.
3102   case Instruction::Shl:
3103   case Instruction::LShr:
3104   case Instruction::AShr:
3105     if (Idx == 1)
3106       return TTI::TCC_Free;
3107     break;
3108   case Instruction::Trunc:
3109   case Instruction::ZExt:
3110   case Instruction::SExt:
3111   case Instruction::IntToPtr:
3112   case Instruction::PtrToInt:
3113   case Instruction::BitCast:
3114   case Instruction::PHI:
3115   case Instruction::Call:
3116   case Instruction::Select:
3117   case Instruction::Ret:
3118   case Instruction::Load:
3119     break;
3120   }
3121 
3122   if (Idx == ImmIdx) {
3123     int NumConstants = divideCeil(BitSize, 64);
3124     int Cost = X86TTIImpl::getIntImmCost(Imm, Ty);
3125     return (Cost <= NumConstants * TTI::TCC_Basic)
3126                ? static_cast<int>(TTI::TCC_Free)
3127                : Cost;
3128   }
3129 
3130   return X86TTIImpl::getIntImmCost(Imm, Ty);
3131 }
3132 
3133 int X86TTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
3134                                     const APInt &Imm, Type *Ty) {
3135   assert(Ty->isIntegerTy());
3136 
3137   unsigned BitSize = Ty->getPrimitiveSizeInBits();
3138   // There is no cost model for constants with a bit size of 0. Return TCC_Free
3139   // here, so that constant hoisting will ignore this constant.
3140   if (BitSize == 0)
3141     return TTI::TCC_Free;
3142 
3143   switch (IID) {
3144   default:
3145     return TTI::TCC_Free;
3146   case Intrinsic::sadd_with_overflow:
3147   case Intrinsic::uadd_with_overflow:
3148   case Intrinsic::ssub_with_overflow:
3149   case Intrinsic::usub_with_overflow:
3150   case Intrinsic::smul_with_overflow:
3151   case Intrinsic::umul_with_overflow:
3152     if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
3153       return TTI::TCC_Free;
3154     break;
3155   case Intrinsic::experimental_stackmap:
3156     if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
3157       return TTI::TCC_Free;
3158     break;
3159   case Intrinsic::experimental_patchpoint_void:
3160   case Intrinsic::experimental_patchpoint_i64:
3161     if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
3162       return TTI::TCC_Free;
3163     break;
3164   }
3165   return X86TTIImpl::getIntImmCost(Imm, Ty);
3166 }
3167 
3168 unsigned X86TTIImpl::getUserCost(const User *U,
3169                                  ArrayRef<const Value *> Operands) {
3170   if (isa<StoreInst>(U)) {
3171     Value *Ptr = U->getOperand(1);
3172     // Store instruction with index and scale costs 2 Uops.
3173     // Check the preceding GEP to identify non-const indices.
3174     if (auto GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
3175       if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); }))
3176         return TTI::TCC_Basic * 2;
3177     }
3178     return TTI::TCC_Basic;
3179   }
3180   return BaseT::getUserCost(U, Operands);
3181 }
3182 
3183 // Return an average cost of Gather / Scatter instruction, maybe improved later
3184 int X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy, Value *Ptr,
3185                                 unsigned Alignment, unsigned AddressSpace) {
3186 
3187   assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost");
3188   unsigned VF = SrcVTy->getVectorNumElements();
3189 
3190   // Try to reduce index size from 64 bit (default for GEP)
3191   // to 32. It is essential for VF 16. If the index can't be reduced to 32, the
3192   // operation will use 16 x 64 indices which do not fit in a zmm and needs
3193   // to split. Also check that the base pointer is the same for all lanes,
3194   // and that there's at most one variable index.
3195   auto getIndexSizeInBits = [](Value *Ptr, const DataLayout& DL) {
3196     unsigned IndexSize = DL.getPointerSizeInBits();
3197     GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3198     if (IndexSize < 64 || !GEP)
3199       return IndexSize;
3200 
3201     unsigned NumOfVarIndices = 0;
3202     Value *Ptrs = GEP->getPointerOperand();
3203     if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs))
3204       return IndexSize;
3205     for (unsigned i = 1; i < GEP->getNumOperands(); ++i) {
3206       if (isa<Constant>(GEP->getOperand(i)))
3207         continue;
3208       Type *IndxTy = GEP->getOperand(i)->getType();
3209       if (IndxTy->isVectorTy())
3210         IndxTy = IndxTy->getVectorElementType();
3211       if ((IndxTy->getPrimitiveSizeInBits() == 64 &&
3212           !isa<SExtInst>(GEP->getOperand(i))) ||
3213          ++NumOfVarIndices > 1)
3214         return IndexSize; // 64
3215     }
3216     return (unsigned)32;
3217   };
3218 
3219 
3220   // Trying to reduce IndexSize to 32 bits for vector 16.
3221   // By default the IndexSize is equal to pointer size.
3222   unsigned IndexSize = (ST->hasAVX512() && VF >= 16)
3223                            ? getIndexSizeInBits(Ptr, DL)
3224                            : DL.getPointerSizeInBits();
3225 
3226   Type *IndexVTy = VectorType::get(IntegerType::get(SrcVTy->getContext(),
3227                                                     IndexSize), VF);
3228   std::pair<int, MVT> IdxsLT = TLI->getTypeLegalizationCost(DL, IndexVTy);
3229   std::pair<int, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, SrcVTy);
3230   int SplitFactor = std::max(IdxsLT.first, SrcLT.first);
3231   if (SplitFactor > 1) {
3232     // Handle splitting of vector of pointers
3233     Type *SplitSrcTy = VectorType::get(SrcVTy->getScalarType(), VF / SplitFactor);
3234     return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment,
3235                                          AddressSpace);
3236   }
3237 
3238   // The gather / scatter cost is given by Intel architects. It is a rough
3239   // number since we are looking at one instruction in a time.
3240   const int GSOverhead = (Opcode == Instruction::Load)
3241                              ? ST->getGatherOverhead()
3242                              : ST->getScatterOverhead();
3243   return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
3244                                            MaybeAlign(Alignment), AddressSpace);
3245 }
3246 
3247 /// Return the cost of full scalarization of gather / scatter operation.
3248 ///
3249 /// Opcode - Load or Store instruction.
3250 /// SrcVTy - The type of the data vector that should be gathered or scattered.
3251 /// VariableMask - The mask is non-constant at compile time.
3252 /// Alignment - Alignment for one element.
3253 /// AddressSpace - pointer[s] address space.
3254 ///
3255 int X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy,
3256                                 bool VariableMask, unsigned Alignment,
3257                                 unsigned AddressSpace) {
3258   unsigned VF = SrcVTy->getVectorNumElements();
3259 
3260   int MaskUnpackCost = 0;
3261   if (VariableMask) {
3262     VectorType *MaskTy =
3263       VectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF);
3264     MaskUnpackCost = getScalarizationOverhead(MaskTy, false, true);
3265     int ScalarCompareCost =
3266       getCmpSelInstrCost(Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()),
3267                          nullptr);
3268     int BranchCost = getCFInstrCost(Instruction::Br);
3269     MaskUnpackCost += VF * (BranchCost + ScalarCompareCost);
3270   }
3271 
3272   // The cost of the scalar loads/stores.
3273   int MemoryOpCost = VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
3274                                           MaybeAlign(Alignment), AddressSpace);
3275 
3276   int InsertExtractCost = 0;
3277   if (Opcode == Instruction::Load)
3278     for (unsigned i = 0; i < VF; ++i)
3279       // Add the cost of inserting each scalar load into the vector
3280       InsertExtractCost +=
3281         getVectorInstrCost(Instruction::InsertElement, SrcVTy, i);
3282   else
3283     for (unsigned i = 0; i < VF; ++i)
3284       // Add the cost of extracting each element out of the data vector
3285       InsertExtractCost +=
3286         getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i);
3287 
3288   return MemoryOpCost + MaskUnpackCost + InsertExtractCost;
3289 }
3290 
3291 /// Calculate the cost of Gather / Scatter operation
3292 int X86TTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *SrcVTy,
3293                                        Value *Ptr, bool VariableMask,
3294                                        unsigned Alignment) {
3295   assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter");
3296   unsigned VF = SrcVTy->getVectorNumElements();
3297   PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3298   if (!PtrTy && Ptr->getType()->isVectorTy())
3299     PtrTy = dyn_cast<PointerType>(Ptr->getType()->getVectorElementType());
3300   assert(PtrTy && "Unexpected type for Ptr argument");
3301   unsigned AddressSpace = PtrTy->getAddressSpace();
3302 
3303   bool Scalarize = false;
3304   if ((Opcode == Instruction::Load &&
3305        !isLegalMaskedGather(SrcVTy, MaybeAlign(Alignment))) ||
3306       (Opcode == Instruction::Store &&
3307        !isLegalMaskedScatter(SrcVTy, MaybeAlign(Alignment))))
3308     Scalarize = true;
3309   // Gather / Scatter for vector 2 is not profitable on KNL / SKX
3310   // Vector-4 of gather/scatter instruction does not exist on KNL.
3311   // We can extend it to 8 elements, but zeroing upper bits of
3312   // the mask vector will add more instructions. Right now we give the scalar
3313   // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter instruction
3314   // is better in the VariableMask case.
3315   if (ST->hasAVX512() && (VF == 2 || (VF == 4 && !ST->hasVLX())))
3316     Scalarize = true;
3317 
3318   if (Scalarize)
3319     return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment,
3320                            AddressSpace);
3321 
3322   return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace);
3323 }
3324 
3325 bool X86TTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1,
3326                                TargetTransformInfo::LSRCost &C2) {
3327     // X86 specific here are "instruction number 1st priority".
3328     return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost,
3329                     C1.NumIVMuls, C1.NumBaseAdds,
3330                     C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
3331            std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost,
3332                     C2.NumIVMuls, C2.NumBaseAdds,
3333                     C2.ScaleCost, C2.ImmCost, C2.SetupCost);
3334 }
3335 
3336 bool X86TTIImpl::canMacroFuseCmp() {
3337   return ST->hasMacroFusion() || ST->hasBranchFusion();
3338 }
3339 
3340 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy, MaybeAlign Alignment) {
3341   if (!ST->hasAVX())
3342     return false;
3343 
3344   // The backend can't handle a single element vector.
3345   if (isa<VectorType>(DataTy) && DataTy->getVectorNumElements() == 1)
3346     return false;
3347   Type *ScalarTy = DataTy->getScalarType();
3348 
3349   if (ScalarTy->isPointerTy())
3350     return true;
3351 
3352   if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
3353     return true;
3354 
3355   if (!ScalarTy->isIntegerTy())
3356     return false;
3357 
3358   unsigned IntWidth = ScalarTy->getIntegerBitWidth();
3359   return IntWidth == 32 || IntWidth == 64 ||
3360          ((IntWidth == 8 || IntWidth == 16) && ST->hasBWI());
3361 }
3362 
3363 bool X86TTIImpl::isLegalMaskedStore(Type *DataType, MaybeAlign Alignment) {
3364   return isLegalMaskedLoad(DataType, Alignment);
3365 }
3366 
3367 bool X86TTIImpl::isLegalNTLoad(Type *DataType, Align Alignment) {
3368   unsigned DataSize = DL.getTypeStoreSize(DataType);
3369   // The only supported nontemporal loads are for aligned vectors of 16 or 32
3370   // bytes.  Note that 32-byte nontemporal vector loads are supported by AVX2
3371   // (the equivalent stores only require AVX).
3372   if (Alignment >= DataSize && (DataSize == 16 || DataSize == 32))
3373     return DataSize == 16 ?  ST->hasSSE1() : ST->hasAVX2();
3374 
3375   return false;
3376 }
3377 
3378 bool X86TTIImpl::isLegalNTStore(Type *DataType, Align Alignment) {
3379   unsigned DataSize = DL.getTypeStoreSize(DataType);
3380 
3381   // SSE4A supports nontemporal stores of float and double at arbitrary
3382   // alignment.
3383   if (ST->hasSSE4A() && (DataType->isFloatTy() || DataType->isDoubleTy()))
3384     return true;
3385 
3386   // Besides the SSE4A subtarget exception above, only aligned stores are
3387   // available nontemporaly on any other subtarget.  And only stores with a size
3388   // of 4..32 bytes (powers of 2, only) are permitted.
3389   if (Alignment < DataSize || DataSize < 4 || DataSize > 32 ||
3390       !isPowerOf2_32(DataSize))
3391     return false;
3392 
3393   // 32-byte vector nontemporal stores are supported by AVX (the equivalent
3394   // loads require AVX2).
3395   if (DataSize == 32)
3396     return ST->hasAVX();
3397   else if (DataSize == 16)
3398     return ST->hasSSE1();
3399   return true;
3400 }
3401 
3402 bool X86TTIImpl::isLegalMaskedExpandLoad(Type *DataTy) {
3403   if (!isa<VectorType>(DataTy))
3404     return false;
3405 
3406   if (!ST->hasAVX512())
3407     return false;
3408 
3409   // The backend can't handle a single element vector.
3410   if (DataTy->getVectorNumElements() == 1)
3411     return false;
3412 
3413   Type *ScalarTy = DataTy->getVectorElementType();
3414 
3415   if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
3416     return true;
3417 
3418   if (!ScalarTy->isIntegerTy())
3419     return false;
3420 
3421   unsigned IntWidth = ScalarTy->getIntegerBitWidth();
3422   return IntWidth == 32 || IntWidth == 64 ||
3423          ((IntWidth == 8 || IntWidth == 16) && ST->hasVBMI2());
3424 }
3425 
3426 bool X86TTIImpl::isLegalMaskedCompressStore(Type *DataTy) {
3427   return isLegalMaskedExpandLoad(DataTy);
3428 }
3429 
3430 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy, MaybeAlign Alignment) {
3431   // Some CPUs have better gather performance than others.
3432   // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only
3433   // enable gather with a -march.
3434   if (!(ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2())))
3435     return false;
3436 
3437   // This function is called now in two cases: from the Loop Vectorizer
3438   // and from the Scalarizer.
3439   // When the Loop Vectorizer asks about legality of the feature,
3440   // the vectorization factor is not calculated yet. The Loop Vectorizer
3441   // sends a scalar type and the decision is based on the width of the
3442   // scalar element.
3443   // Later on, the cost model will estimate usage this intrinsic based on
3444   // the vector type.
3445   // The Scalarizer asks again about legality. It sends a vector type.
3446   // In this case we can reject non-power-of-2 vectors.
3447   // We also reject single element vectors as the type legalizer can't
3448   // scalarize it.
3449   if (isa<VectorType>(DataTy)) {
3450     unsigned NumElts = DataTy->getVectorNumElements();
3451     if (NumElts == 1 || !isPowerOf2_32(NumElts))
3452       return false;
3453   }
3454   Type *ScalarTy = DataTy->getScalarType();
3455   if (ScalarTy->isPointerTy())
3456     return true;
3457 
3458   if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
3459     return true;
3460 
3461   if (!ScalarTy->isIntegerTy())
3462     return false;
3463 
3464   unsigned IntWidth = ScalarTy->getIntegerBitWidth();
3465   return IntWidth == 32 || IntWidth == 64;
3466 }
3467 
3468 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType, MaybeAlign Alignment) {
3469   // AVX2 doesn't support scatter
3470   if (!ST->hasAVX512())
3471     return false;
3472   return isLegalMaskedGather(DataType, Alignment);
3473 }
3474 
3475 bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) {
3476   EVT VT = TLI->getValueType(DL, DataType);
3477   return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT);
3478 }
3479 
3480 bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
3481   return false;
3482 }
3483 
3484 bool X86TTIImpl::areInlineCompatible(const Function *Caller,
3485                                      const Function *Callee) const {
3486   const TargetMachine &TM = getTLI()->getTargetMachine();
3487 
3488   // Work this as a subsetting of subtarget features.
3489   const FeatureBitset &CallerBits =
3490       TM.getSubtargetImpl(*Caller)->getFeatureBits();
3491   const FeatureBitset &CalleeBits =
3492       TM.getSubtargetImpl(*Callee)->getFeatureBits();
3493 
3494   FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList;
3495   FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList;
3496   return (RealCallerBits & RealCalleeBits) == RealCalleeBits;
3497 }
3498 
3499 bool X86TTIImpl::areFunctionArgsABICompatible(
3500     const Function *Caller, const Function *Callee,
3501     SmallPtrSetImpl<Argument *> &Args) const {
3502   if (!BaseT::areFunctionArgsABICompatible(Caller, Callee, Args))
3503     return false;
3504 
3505   // If we get here, we know the target features match. If one function
3506   // considers 512-bit vectors legal and the other does not, consider them
3507   // incompatible.
3508   // FIXME Look at the arguments and only consider 512 bit or larger vectors?
3509   const TargetMachine &TM = getTLI()->getTargetMachine();
3510 
3511   return TM.getSubtarget<X86Subtarget>(*Caller).useAVX512Regs() ==
3512          TM.getSubtarget<X86Subtarget>(*Callee).useAVX512Regs();
3513 }
3514 
3515 X86TTIImpl::TTI::MemCmpExpansionOptions
3516 X86TTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
3517   TTI::MemCmpExpansionOptions Options;
3518   Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize);
3519   Options.NumLoadsPerBlock = 2;
3520   if (IsZeroCmp) {
3521     // Only enable vector loads for equality comparison. Right now the vector
3522     // version is not as fast for three way compare (see #33329).
3523     const unsigned PreferredWidth = ST->getPreferVectorWidth();
3524     if (PreferredWidth >= 512 && ST->hasAVX512()) Options.LoadSizes.push_back(64);
3525     if (PreferredWidth >= 256 && ST->hasAVX()) Options.LoadSizes.push_back(32);
3526     if (PreferredWidth >= 128 && ST->hasSSE2()) Options.LoadSizes.push_back(16);
3527     // All GPR and vector loads can be unaligned.
3528     Options.AllowOverlappingLoads = true;
3529   }
3530   if (ST->is64Bit()) {
3531     Options.LoadSizes.push_back(8);
3532   }
3533   Options.LoadSizes.push_back(4);
3534   Options.LoadSizes.push_back(2);
3535   Options.LoadSizes.push_back(1);
3536   return Options;
3537 }
3538 
3539 bool X86TTIImpl::enableInterleavedAccessVectorization() {
3540   // TODO: We expect this to be beneficial regardless of arch,
3541   // but there are currently some unexplained performance artifacts on Atom.
3542   // As a temporary solution, disable on Atom.
3543   return !(ST->isAtom());
3544 }
3545 
3546 // Get estimation for interleaved load/store operations for AVX2.
3547 // \p Factor is the interleaved-access factor (stride) - number of
3548 // (interleaved) elements in the group.
3549 // \p Indices contains the indices for a strided load: when the
3550 // interleaved load has gaps they indicate which elements are used.
3551 // If Indices is empty (or if the number of indices is equal to the size
3552 // of the interleaved-access as given in \p Factor) the access has no gaps.
3553 //
3554 // As opposed to AVX-512, AVX2 does not have generic shuffles that allow
3555 // computing the cost using a generic formula as a function of generic
3556 // shuffles. We therefore use a lookup table instead, filled according to
3557 // the instruction sequences that codegen currently generates.
3558 int X86TTIImpl::getInterleavedMemoryOpCostAVX2(unsigned Opcode, Type *VecTy,
3559                                                unsigned Factor,
3560                                                ArrayRef<unsigned> Indices,
3561                                                unsigned Alignment,
3562                                                unsigned AddressSpace,
3563                                                bool UseMaskForCond,
3564                                                bool UseMaskForGaps) {
3565 
3566   if (UseMaskForCond || UseMaskForGaps)
3567     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
3568                                              Alignment, AddressSpace,
3569                                              UseMaskForCond, UseMaskForGaps);
3570 
3571   // We currently Support only fully-interleaved groups, with no gaps.
3572   // TODO: Support also strided loads (interleaved-groups with gaps).
3573   if (Indices.size() && Indices.size() != Factor)
3574     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
3575                                              Alignment, AddressSpace);
3576 
3577   // VecTy for interleave memop is <VF*Factor x Elt>.
3578   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
3579   // VecTy = <12 x i32>.
3580   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
3581 
3582   // This function can be called with VecTy=<6xi128>, Factor=3, in which case
3583   // the VF=2, while v2i128 is an unsupported MVT vector type
3584   // (see MachineValueType.h::getVectorVT()).
3585   if (!LegalVT.isVector())
3586     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
3587                                              Alignment, AddressSpace);
3588 
3589   unsigned VF = VecTy->getVectorNumElements() / Factor;
3590   Type *ScalarTy = VecTy->getVectorElementType();
3591 
3592   // Calculate the number of memory operations (NumOfMemOps), required
3593   // for load/store the VecTy.
3594   unsigned VecTySize = DL.getTypeStoreSize(VecTy);
3595   unsigned LegalVTSize = LegalVT.getStoreSize();
3596   unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize;
3597 
3598   // Get the cost of one memory operation.
3599   Type *SingleMemOpTy = VectorType::get(VecTy->getVectorElementType(),
3600                                         LegalVT.getVectorNumElements());
3601   unsigned MemOpCost = getMemoryOpCost(Opcode, SingleMemOpTy,
3602                                        MaybeAlign(Alignment), AddressSpace);
3603 
3604   VectorType *VT = VectorType::get(ScalarTy, VF);
3605   EVT ETy = TLI->getValueType(DL, VT);
3606   if (!ETy.isSimple())
3607     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
3608                                              Alignment, AddressSpace);
3609 
3610   // TODO: Complete for other data-types and strides.
3611   // Each combination of Stride, ElementTy and VF results in a different
3612   // sequence; The cost tables are therefore accessed with:
3613   // Factor (stride) and VectorType=VFxElemType.
3614   // The Cost accounts only for the shuffle sequence;
3615   // The cost of the loads/stores is accounted for separately.
3616   //
3617   static const CostTblEntry AVX2InterleavedLoadTbl[] = {
3618     { 2, MVT::v4i64, 6 }, //(load 8i64 and) deinterleave into 2 x 4i64
3619     { 2, MVT::v4f64, 6 }, //(load 8f64 and) deinterleave into 2 x 4f64
3620 
3621     { 3, MVT::v2i8,  10 }, //(load 6i8 and)  deinterleave into 3 x 2i8
3622     { 3, MVT::v4i8,  4 },  //(load 12i8 and) deinterleave into 3 x 4i8
3623     { 3, MVT::v8i8,  9 },  //(load 24i8 and) deinterleave into 3 x 8i8
3624     { 3, MVT::v16i8, 11},  //(load 48i8 and) deinterleave into 3 x 16i8
3625     { 3, MVT::v32i8, 13},  //(load 96i8 and) deinterleave into 3 x 32i8
3626     { 3, MVT::v8f32, 17 }, //(load 24f32 and)deinterleave into 3 x 8f32
3627 
3628     { 4, MVT::v2i8,  12 }, //(load 8i8 and)   deinterleave into 4 x 2i8
3629     { 4, MVT::v4i8,  4 },  //(load 16i8 and)  deinterleave into 4 x 4i8
3630     { 4, MVT::v8i8,  20 }, //(load 32i8 and)  deinterleave into 4 x 8i8
3631     { 4, MVT::v16i8, 39 }, //(load 64i8 and)  deinterleave into 4 x 16i8
3632     { 4, MVT::v32i8, 80 }, //(load 128i8 and) deinterleave into 4 x 32i8
3633 
3634     { 8, MVT::v8f32, 40 }  //(load 64f32 and)deinterleave into 8 x 8f32
3635   };
3636 
3637   static const CostTblEntry AVX2InterleavedStoreTbl[] = {
3638     { 2, MVT::v4i64, 6 }, //interleave into 2 x 4i64 into 8i64 (and store)
3639     { 2, MVT::v4f64, 6 }, //interleave into 2 x 4f64 into 8f64 (and store)
3640 
3641     { 3, MVT::v2i8,  7 },  //interleave 3 x 2i8  into 6i8 (and store)
3642     { 3, MVT::v4i8,  8 },  //interleave 3 x 4i8  into 12i8 (and store)
3643     { 3, MVT::v8i8,  11 }, //interleave 3 x 8i8  into 24i8 (and store)
3644     { 3, MVT::v16i8, 11 }, //interleave 3 x 16i8 into 48i8 (and store)
3645     { 3, MVT::v32i8, 13 }, //interleave 3 x 32i8 into 96i8 (and store)
3646 
3647     { 4, MVT::v2i8,  12 }, //interleave 4 x 2i8  into 8i8 (and store)
3648     { 4, MVT::v4i8,  9 },  //interleave 4 x 4i8  into 16i8 (and store)
3649     { 4, MVT::v8i8,  10 }, //interleave 4 x 8i8  into 32i8 (and store)
3650     { 4, MVT::v16i8, 10 }, //interleave 4 x 16i8 into 64i8 (and store)
3651     { 4, MVT::v32i8, 12 }  //interleave 4 x 32i8 into 128i8 (and store)
3652   };
3653 
3654   if (Opcode == Instruction::Load) {
3655     if (const auto *Entry =
3656             CostTableLookup(AVX2InterleavedLoadTbl, Factor, ETy.getSimpleVT()))
3657       return NumOfMemOps * MemOpCost + Entry->Cost;
3658   } else {
3659     assert(Opcode == Instruction::Store &&
3660            "Expected Store Instruction at this  point");
3661     if (const auto *Entry =
3662             CostTableLookup(AVX2InterleavedStoreTbl, Factor, ETy.getSimpleVT()))
3663       return NumOfMemOps * MemOpCost + Entry->Cost;
3664   }
3665 
3666   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
3667                                            Alignment, AddressSpace);
3668 }
3669 
3670 // Get estimation for interleaved load/store operations and strided load.
3671 // \p Indices contains indices for strided load.
3672 // \p Factor - the factor of interleaving.
3673 // AVX-512 provides 3-src shuffles that significantly reduces the cost.
3674 int X86TTIImpl::getInterleavedMemoryOpCostAVX512(unsigned Opcode, Type *VecTy,
3675                                                  unsigned Factor,
3676                                                  ArrayRef<unsigned> Indices,
3677                                                  unsigned Alignment,
3678                                                  unsigned AddressSpace,
3679                                                  bool UseMaskForCond,
3680                                                  bool UseMaskForGaps) {
3681 
3682   if (UseMaskForCond || UseMaskForGaps)
3683     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
3684                                              Alignment, AddressSpace,
3685                                              UseMaskForCond, UseMaskForGaps);
3686 
3687   // VecTy for interleave memop is <VF*Factor x Elt>.
3688   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
3689   // VecTy = <12 x i32>.
3690 
3691   // Calculate the number of memory operations (NumOfMemOps), required
3692   // for load/store the VecTy.
3693   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
3694   unsigned VecTySize = DL.getTypeStoreSize(VecTy);
3695   unsigned LegalVTSize = LegalVT.getStoreSize();
3696   unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize;
3697 
3698   // Get the cost of one memory operation.
3699   Type *SingleMemOpTy = VectorType::get(VecTy->getVectorElementType(),
3700                                         LegalVT.getVectorNumElements());
3701   unsigned MemOpCost = getMemoryOpCost(Opcode, SingleMemOpTy,
3702                                        MaybeAlign(Alignment), AddressSpace);
3703 
3704   unsigned VF = VecTy->getVectorNumElements() / Factor;
3705   MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF);
3706 
3707   if (Opcode == Instruction::Load) {
3708     // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl)
3709     // contain the cost of the optimized shuffle sequence that the
3710     // X86InterleavedAccess pass will generate.
3711     // The cost of loads and stores are computed separately from the table.
3712 
3713     // X86InterleavedAccess support only the following interleaved-access group.
3714     static const CostTblEntry AVX512InterleavedLoadTbl[] = {
3715         {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8
3716         {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8
3717         {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8
3718     };
3719 
3720     if (const auto *Entry =
3721             CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT))
3722       return NumOfMemOps * MemOpCost + Entry->Cost;
3723     //If an entry does not exist, fallback to the default implementation.
3724 
3725     // Kind of shuffle depends on number of loaded values.
3726     // If we load the entire data in one register, we can use a 1-src shuffle.
3727     // Otherwise, we'll merge 2 sources in each operation.
3728     TTI::ShuffleKind ShuffleKind =
3729         (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc;
3730 
3731     unsigned ShuffleCost =
3732         getShuffleCost(ShuffleKind, SingleMemOpTy, 0, nullptr);
3733 
3734     unsigned NumOfLoadsInInterleaveGrp =
3735         Indices.size() ? Indices.size() : Factor;
3736     Type *ResultTy = VectorType::get(VecTy->getVectorElementType(),
3737                                      VecTy->getVectorNumElements() / Factor);
3738     unsigned NumOfResults =
3739         getTLI()->getTypeLegalizationCost(DL, ResultTy).first *
3740         NumOfLoadsInInterleaveGrp;
3741 
3742     // About a half of the loads may be folded in shuffles when we have only
3743     // one result. If we have more than one result, we do not fold loads at all.
3744     unsigned NumOfUnfoldedLoads =
3745         NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2;
3746 
3747     // Get a number of shuffle operations per result.
3748     unsigned NumOfShufflesPerResult =
3749         std::max((unsigned)1, (unsigned)(NumOfMemOps - 1));
3750 
3751     // The SK_MergeTwoSrc shuffle clobbers one of src operands.
3752     // When we have more than one destination, we need additional instructions
3753     // to keep sources.
3754     unsigned NumOfMoves = 0;
3755     if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc)
3756       NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2;
3757 
3758     int Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost +
3759                NumOfUnfoldedLoads * MemOpCost + NumOfMoves;
3760 
3761     return Cost;
3762   }
3763 
3764   // Store.
3765   assert(Opcode == Instruction::Store &&
3766          "Expected Store Instruction at this  point");
3767   // X86InterleavedAccess support only the following interleaved-access group.
3768   static const CostTblEntry AVX512InterleavedStoreTbl[] = {
3769       {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store)
3770       {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store)
3771       {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store)
3772 
3773       {4, MVT::v8i8, 10},  // interleave 4 x 8i8  into 32i8  (and store)
3774       {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8  (and store)
3775       {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store)
3776       {4, MVT::v64i8, 24}  // interleave 4 x 32i8 into 256i8 (and store)
3777   };
3778 
3779   if (const auto *Entry =
3780           CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT))
3781     return NumOfMemOps * MemOpCost + Entry->Cost;
3782   //If an entry does not exist, fallback to the default implementation.
3783 
3784   // There is no strided stores meanwhile. And store can't be folded in
3785   // shuffle.
3786   unsigned NumOfSources = Factor; // The number of values to be merged.
3787   unsigned ShuffleCost =
3788       getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, 0, nullptr);
3789   unsigned NumOfShufflesPerStore = NumOfSources - 1;
3790 
3791   // The SK_MergeTwoSrc shuffle clobbers one of src operands.
3792   // We need additional instructions to keep sources.
3793   unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2;
3794   int Cost = NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) +
3795              NumOfMoves;
3796   return Cost;
3797 }
3798 
3799 int X86TTIImpl::getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
3800                                            unsigned Factor,
3801                                            ArrayRef<unsigned> Indices,
3802                                            unsigned Alignment,
3803                                            unsigned AddressSpace,
3804                                            bool UseMaskForCond,
3805                                            bool UseMaskForGaps) {
3806   auto isSupportedOnAVX512 = [](Type *VecTy, bool HasBW) {
3807     Type *EltTy = VecTy->getVectorElementType();
3808     if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) ||
3809         EltTy->isIntegerTy(32) || EltTy->isPointerTy())
3810       return true;
3811     if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8))
3812       return HasBW;
3813     return false;
3814   };
3815   if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI()))
3816     return getInterleavedMemoryOpCostAVX512(Opcode, VecTy, Factor, Indices,
3817                                             Alignment, AddressSpace,
3818                                             UseMaskForCond, UseMaskForGaps);
3819   if (ST->hasAVX2())
3820     return getInterleavedMemoryOpCostAVX2(Opcode, VecTy, Factor, Indices,
3821                                           Alignment, AddressSpace,
3822                                           UseMaskForCond, UseMaskForGaps);
3823 
3824   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
3825                                            Alignment, AddressSpace,
3826                                            UseMaskForCond, UseMaskForGaps);
3827 }
3828