1 //===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 /// \file
10 /// This file implements a TargetTransformInfo analysis pass specific to the
11 /// X86 target machine. It uses the target's detailed information to provide
12 /// more precise answers to certain TTI queries, while letting the target
13 /// independent and default TTI implementations handle the rest.
14 ///
15 //===----------------------------------------------------------------------===//
16 /// About Cost Model numbers used below it's necessary to say the following:
17 /// the numbers correspond to some "generic" X86 CPU instead of usage of
18 /// concrete CPU model. Usually the numbers correspond to CPU where the feature
19 /// apeared at the first time. For example, if we do Subtarget.hasSSE42() in
20 /// the lookups below the cost is based on Nehalem as that was the first CPU
21 /// to support that feature level and thus has most likely the worst case cost.
22 /// Some examples of other technologies/CPUs:
23 ///   SSE 3   - Pentium4 / Athlon64
24 ///   SSE 4.1 - Penryn
25 ///   SSE 4.2 - Nehalem
26 ///   AVX     - Sandy Bridge
27 ///   AVX2    - Haswell
28 ///   AVX-512 - Xeon Phi / Skylake
29 /// And some examples of instruction target dependent costs (latency)
30 ///                   divss     sqrtss          rsqrtss
31 ///   AMD K7            11-16     19              3
32 ///   Piledriver        9-24      13-15           5
33 ///   Jaguar            14        16              2
34 ///   Pentium II,III    18        30              2
35 ///   Nehalem           7-14      7-18            3
36 ///   Haswell           10-13     11              5
37 /// TODO: Develop and implement  the target dependent cost model and
38 /// specialize cost numbers for different Cost Model Targets such as throughput,
39 /// code size, latency and uop count.
40 //===----------------------------------------------------------------------===//
41 
42 #include "X86TargetTransformInfo.h"
43 #include "llvm/Analysis/TargetTransformInfo.h"
44 #include "llvm/CodeGen/BasicTTIImpl.h"
45 #include "llvm/CodeGen/CostTable.h"
46 #include "llvm/CodeGen/TargetLowering.h"
47 #include "llvm/IR/IntrinsicInst.h"
48 #include "llvm/Support/Debug.h"
49 
50 using namespace llvm;
51 
52 #define DEBUG_TYPE "x86tti"
53 
54 //===----------------------------------------------------------------------===//
55 //
56 // X86 cost model.
57 //
58 //===----------------------------------------------------------------------===//
59 
60 TargetTransformInfo::PopcntSupportKind
61 X86TTIImpl::getPopcntSupport(unsigned TyWidth) {
62   assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
63   // TODO: Currently the __builtin_popcount() implementation using SSE3
64   //   instructions is inefficient. Once the problem is fixed, we should
65   //   call ST->hasSSE3() instead of ST->hasPOPCNT().
66   return ST->hasPOPCNT() ? TTI::PSK_FastHardware : TTI::PSK_Software;
67 }
68 
69 llvm::Optional<unsigned> X86TTIImpl::getCacheSize(
70   TargetTransformInfo::CacheLevel Level) const {
71   switch (Level) {
72   case TargetTransformInfo::CacheLevel::L1D:
73     //   - Penryn
74     //   - Nehalem
75     //   - Westmere
76     //   - Sandy Bridge
77     //   - Ivy Bridge
78     //   - Haswell
79     //   - Broadwell
80     //   - Skylake
81     //   - Kabylake
82     return 32 * 1024;  //  32 KByte
83   case TargetTransformInfo::CacheLevel::L2D:
84     //   - Penryn
85     //   - Nehalem
86     //   - Westmere
87     //   - Sandy Bridge
88     //   - Ivy Bridge
89     //   - Haswell
90     //   - Broadwell
91     //   - Skylake
92     //   - Kabylake
93     return 256 * 1024; // 256 KByte
94   }
95 
96   llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
97 }
98 
99 llvm::Optional<unsigned> X86TTIImpl::getCacheAssociativity(
100   TargetTransformInfo::CacheLevel Level) const {
101   //   - Penryn
102   //   - Nehalem
103   //   - Westmere
104   //   - Sandy Bridge
105   //   - Ivy Bridge
106   //   - Haswell
107   //   - Broadwell
108   //   - Skylake
109   //   - Kabylake
110   switch (Level) {
111   case TargetTransformInfo::CacheLevel::L1D:
112     LLVM_FALLTHROUGH;
113   case TargetTransformInfo::CacheLevel::L2D:
114     return 8;
115   }
116 
117   llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
118 }
119 
120 unsigned X86TTIImpl::getNumberOfRegisters(bool Vector) {
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(
173     unsigned Opcode, Type *Ty,
174     TTI::OperandValueKind Op1Info, TTI::OperandValueKind Op2Info,
175     TTI::OperandValueProperties Opd1PropInfo,
176     TTI::OperandValueProperties Opd2PropInfo,
177     ArrayRef<const Value *> Args) {
178   // Legalize the type.
179   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
180 
181   int ISD = TLI->InstructionOpcodeToISD(Opcode);
182   assert(ISD && "Invalid opcode");
183 
184   static const CostTblEntry GLMCostTable[] = {
185     { ISD::FDIV,  MVT::f32,   18 }, // divss
186     { ISD::FDIV,  MVT::v4f32, 35 }, // divps
187     { ISD::FDIV,  MVT::f64,   33 }, // divsd
188     { ISD::FDIV,  MVT::v2f64, 65 }, // divpd
189   };
190 
191   if (ST->isGLM())
192     if (const auto *Entry = CostTableLookup(GLMCostTable, ISD,
193                                             LT.second))
194       return LT.first * Entry->Cost;
195 
196   static const CostTblEntry SLMCostTable[] = {
197     { ISD::MUL,   MVT::v4i32, 11 }, // pmulld
198     { ISD::MUL,   MVT::v8i16, 2  }, // pmullw
199     { ISD::MUL,   MVT::v16i8, 14 }, // extend/pmullw/trunc sequence.
200     { ISD::FMUL,  MVT::f64,   2  }, // mulsd
201     { ISD::FMUL,  MVT::v2f64, 4  }, // mulpd
202     { ISD::FMUL,  MVT::v4f32, 2  }, // mulps
203     { ISD::FDIV,  MVT::f32,   17 }, // divss
204     { ISD::FDIV,  MVT::v4f32, 39 }, // divps
205     { ISD::FDIV,  MVT::f64,   32 }, // divsd
206     { ISD::FDIV,  MVT::v2f64, 69 }, // divpd
207     { ISD::FADD,  MVT::v2f64, 2  }, // addpd
208     { ISD::FSUB,  MVT::v2f64, 2  }, // subpd
209     // v2i64/v4i64 mul is custom lowered as a series of long:
210     // multiplies(3), shifts(3) and adds(2)
211     // slm muldq version throughput is 2 and addq throughput 4
212     // thus: 3X2 (muldq throughput) + 3X1 (shift throughput) +
213     //       3X4 (addq throughput) = 17
214     { ISD::MUL,   MVT::v2i64, 17 },
215     // slm addq\subq throughput is 4
216     { ISD::ADD,   MVT::v2i64, 4  },
217     { ISD::SUB,   MVT::v2i64, 4  },
218   };
219 
220   if (ST->isSLM()) {
221     if (Args.size() == 2 && ISD == ISD::MUL && LT.second == MVT::v4i32) {
222       // Check if the operands can be shrinked into a smaller datatype.
223       bool Op1Signed = false;
224       unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed);
225       bool Op2Signed = false;
226       unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed);
227 
228       bool signedMode = Op1Signed | Op2Signed;
229       unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize);
230 
231       if (OpMinSize <= 7)
232         return LT.first * 3; // pmullw/sext
233       if (!signedMode && OpMinSize <= 8)
234         return LT.first * 3; // pmullw/zext
235       if (OpMinSize <= 15)
236         return LT.first * 5; // pmullw/pmulhw/pshuf
237       if (!signedMode && OpMinSize <= 16)
238         return LT.first * 5; // pmullw/pmulhw/pshuf
239     }
240 
241     if (const auto *Entry = CostTableLookup(SLMCostTable, ISD,
242                                             LT.second)) {
243       return LT.first * Entry->Cost;
244     }
245   }
246 
247   if ((ISD == ISD::SDIV || ISD == ISD::SREM || ISD == ISD::UDIV ||
248        ISD == ISD::UREM) &&
249       (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
250        Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
251       Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
252     if (ISD == ISD::SDIV || ISD == ISD::SREM) {
253       // On X86, vector signed division by constants power-of-two are
254       // normally expanded to the sequence SRA + SRL + ADD + SRA.
255       // The OperandValue properties may not be the same as that of the previous
256       // operation; conservatively assume OP_None.
257       int Cost =
258           2 * getArithmeticInstrCost(Instruction::AShr, Ty, Op1Info, Op2Info,
259                                      TargetTransformInfo::OP_None,
260                                      TargetTransformInfo::OP_None);
261       Cost += getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info,
262                                      TargetTransformInfo::OP_None,
263                                      TargetTransformInfo::OP_None);
264       Cost += getArithmeticInstrCost(Instruction::Add, Ty, Op1Info, Op2Info,
265                                      TargetTransformInfo::OP_None,
266                                      TargetTransformInfo::OP_None);
267 
268       if (ISD == ISD::SREM) {
269         // For SREM: (X % C) is the equivalent of (X - (X/C)*C)
270         Cost += getArithmeticInstrCost(Instruction::Mul, Ty, Op1Info, Op2Info);
271         Cost += getArithmeticInstrCost(Instruction::Sub, Ty, Op1Info, Op2Info);
272       }
273 
274       return Cost;
275     }
276 
277     // Vector unsigned division/remainder will be simplified to shifts/masks.
278     if (ISD == ISD::UDIV)
279       return getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info,
280                                     TargetTransformInfo::OP_None,
281                                     TargetTransformInfo::OP_None);
282 
283     if (ISD == ISD::UREM)
284       return getArithmeticInstrCost(Instruction::And, Ty, Op1Info, Op2Info,
285                                     TargetTransformInfo::OP_None,
286                                     TargetTransformInfo::OP_None);
287   }
288 
289   static const CostTblEntry AVX512BWUniformConstCostTable[] = {
290     { ISD::SHL,  MVT::v64i8,   2 }, // psllw + pand.
291     { ISD::SRL,  MVT::v64i8,   2 }, // psrlw + pand.
292     { ISD::SRA,  MVT::v64i8,   4 }, // psrlw, pand, pxor, psubb.
293 
294     { ISD::SDIV, MVT::v32i16,  6 }, // vpmulhw sequence
295     { ISD::SREM, MVT::v32i16,  8 }, // vpmulhw+mul+sub sequence
296     { ISD::UDIV, MVT::v32i16,  6 }, // vpmulhuw sequence
297     { ISD::UREM, MVT::v32i16,  8 }, // vpmulhuw+mul+sub sequence
298   };
299 
300   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
301       ST->hasBWI()) {
302     if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD,
303                                             LT.second))
304       return LT.first * Entry->Cost;
305   }
306 
307   static const CostTblEntry AVX512UniformConstCostTable[] = {
308     { ISD::SRA,  MVT::v2i64,   1 },
309     { ISD::SRA,  MVT::v4i64,   1 },
310     { ISD::SRA,  MVT::v8i64,   1 },
311 
312     { ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence
313     { ISD::SREM, MVT::v16i32, 17 }, // vpmuldq+mul+sub sequence
314     { ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence
315     { ISD::UREM, MVT::v16i32, 17 }, // vpmuludq+mul+sub sequence
316   };
317 
318   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
319       ST->hasAVX512()) {
320     if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD,
321                                             LT.second))
322       return LT.first * Entry->Cost;
323   }
324 
325   static const CostTblEntry AVX2UniformConstCostTable[] = {
326     { ISD::SHL,  MVT::v32i8,   2 }, // psllw + pand.
327     { ISD::SRL,  MVT::v32i8,   2 }, // psrlw + pand.
328     { ISD::SRA,  MVT::v32i8,   4 }, // psrlw, pand, pxor, psubb.
329 
330     { ISD::SRA,  MVT::v4i64,   4 }, // 2 x psrad + shuffle.
331 
332     { ISD::SDIV, MVT::v16i16,  6 }, // vpmulhw sequence
333     { ISD::SREM, MVT::v16i16,  8 }, // vpmulhw+mul+sub sequence
334     { ISD::UDIV, MVT::v16i16,  6 }, // vpmulhuw sequence
335     { ISD::UREM, MVT::v16i16,  8 }, // vpmulhuw+mul+sub sequence
336     { ISD::SDIV, MVT::v8i32,  15 }, // vpmuldq sequence
337     { ISD::SREM, MVT::v8i32,  19 }, // vpmuldq+mul+sub sequence
338     { ISD::UDIV, MVT::v8i32,  15 }, // vpmuludq sequence
339     { ISD::UREM, MVT::v8i32,  19 }, // vpmuludq+mul+sub sequence
340   };
341 
342   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
343       ST->hasAVX2()) {
344     if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD,
345                                             LT.second))
346       return LT.first * Entry->Cost;
347   }
348 
349   static const CostTblEntry SSE2UniformConstCostTable[] = {
350     { ISD::SHL,  MVT::v16i8,     2 }, // psllw + pand.
351     { ISD::SRL,  MVT::v16i8,     2 }, // psrlw + pand.
352     { ISD::SRA,  MVT::v16i8,     4 }, // psrlw, pand, pxor, psubb.
353 
354     { ISD::SHL,  MVT::v32i8,   4+2 }, // 2*(psllw + pand) + split.
355     { ISD::SRL,  MVT::v32i8,   4+2 }, // 2*(psrlw + pand) + split.
356     { ISD::SRA,  MVT::v32i8,   8+2 }, // 2*(psrlw, pand, pxor, psubb) + split.
357 
358     { ISD::SDIV, MVT::v16i16, 12+2 }, // 2*pmulhw sequence + split.
359     { ISD::SREM, MVT::v16i16, 16+2 }, // 2*pmulhw+mul+sub sequence + split.
360     { ISD::SDIV, MVT::v8i16,     6 }, // pmulhw sequence
361     { ISD::SREM, MVT::v8i16,     8 }, // pmulhw+mul+sub sequence
362     { ISD::UDIV, MVT::v16i16, 12+2 }, // 2*pmulhuw sequence + split.
363     { ISD::UREM, MVT::v16i16, 16+2 }, // 2*pmulhuw+mul+sub sequence + split.
364     { ISD::UDIV, MVT::v8i16,     6 }, // pmulhuw sequence
365     { ISD::UREM, MVT::v8i16,     8 }, // pmulhuw+mul+sub sequence
366     { ISD::SDIV, MVT::v8i32,  38+2 }, // 2*pmuludq sequence + split.
367     { ISD::SREM, MVT::v8i32,  48+2 }, // 2*pmuludq+mul+sub sequence + split.
368     { ISD::SDIV, MVT::v4i32,    19 }, // pmuludq sequence
369     { ISD::SREM, MVT::v4i32,    24 }, // pmuludq+mul+sub sequence
370     { ISD::UDIV, MVT::v8i32,  30+2 }, // 2*pmuludq sequence + split.
371     { ISD::UREM, MVT::v8i32,  40+2 }, // 2*pmuludq+mul+sub sequence + split.
372     { ISD::UDIV, MVT::v4i32,    15 }, // pmuludq sequence
373     { ISD::UREM, MVT::v4i32,    20 }, // pmuludq+mul+sub sequence
374   };
375 
376   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
377       ST->hasSSE2()) {
378     // pmuldq sequence.
379     if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX())
380       return LT.first * 32;
381     if (ISD == ISD::SREM && LT.second == MVT::v8i32 && ST->hasAVX())
382       return LT.first * 38;
383     if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
384       return LT.first * 15;
385     if (ISD == ISD::SREM && LT.second == MVT::v4i32 && ST->hasSSE41())
386       return LT.first * 20;
387 
388     // XOP has faster vXi8 shifts.
389     if ((ISD != ISD::SHL && ISD != ISD::SRL && ISD != ISD::SRA) ||
390         !ST->hasXOP())
391       if (const auto *Entry =
392               CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second))
393         return LT.first * Entry->Cost;
394   }
395 
396   static const CostTblEntry AVX2UniformCostTable[] = {
397     // Uniform splats are cheaper for the following instructions.
398     { ISD::SHL,  MVT::v16i16, 1 }, // psllw.
399     { ISD::SRL,  MVT::v16i16, 1 }, // psrlw.
400     { ISD::SRA,  MVT::v16i16, 1 }, // psraw.
401   };
402 
403   if (ST->hasAVX2() &&
404       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
405        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
406     if (const auto *Entry =
407             CostTableLookup(AVX2UniformCostTable, ISD, LT.second))
408       return LT.first * Entry->Cost;
409   }
410 
411   static const CostTblEntry SSE2UniformCostTable[] = {
412     // Uniform splats are cheaper for the following instructions.
413     { ISD::SHL,  MVT::v8i16,  1 }, // psllw.
414     { ISD::SHL,  MVT::v4i32,  1 }, // pslld
415     { ISD::SHL,  MVT::v2i64,  1 }, // psllq.
416 
417     { ISD::SRL,  MVT::v8i16,  1 }, // psrlw.
418     { ISD::SRL,  MVT::v4i32,  1 }, // psrld.
419     { ISD::SRL,  MVT::v2i64,  1 }, // psrlq.
420 
421     { ISD::SRA,  MVT::v8i16,  1 }, // psraw.
422     { ISD::SRA,  MVT::v4i32,  1 }, // psrad.
423   };
424 
425   if (ST->hasSSE2() &&
426       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
427        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
428     if (const auto *Entry =
429             CostTableLookup(SSE2UniformCostTable, ISD, LT.second))
430       return LT.first * Entry->Cost;
431   }
432 
433   static const CostTblEntry AVX512DQCostTable[] = {
434     { ISD::MUL,  MVT::v2i64, 1 },
435     { ISD::MUL,  MVT::v4i64, 1 },
436     { ISD::MUL,  MVT::v8i64, 1 }
437   };
438 
439   // Look for AVX512DQ lowering tricks for custom cases.
440   if (ST->hasDQI())
441     if (const auto *Entry = CostTableLookup(AVX512DQCostTable, ISD, LT.second))
442       return LT.first * Entry->Cost;
443 
444   static const CostTblEntry AVX512BWCostTable[] = {
445     { ISD::SHL,   MVT::v8i16,      1 }, // vpsllvw
446     { ISD::SRL,   MVT::v8i16,      1 }, // vpsrlvw
447     { ISD::SRA,   MVT::v8i16,      1 }, // vpsravw
448 
449     { ISD::SHL,   MVT::v16i16,     1 }, // vpsllvw
450     { ISD::SRL,   MVT::v16i16,     1 }, // vpsrlvw
451     { ISD::SRA,   MVT::v16i16,     1 }, // vpsravw
452 
453     { ISD::SHL,   MVT::v32i16,     1 }, // vpsllvw
454     { ISD::SRL,   MVT::v32i16,     1 }, // vpsrlvw
455     { ISD::SRA,   MVT::v32i16,     1 }, // vpsravw
456 
457     { ISD::SHL,   MVT::v64i8,     11 }, // vpblendvb sequence.
458     { ISD::SRL,   MVT::v64i8,     11 }, // vpblendvb sequence.
459     { ISD::SRA,   MVT::v64i8,     24 }, // vpblendvb sequence.
460 
461     { ISD::MUL,   MVT::v64i8,     11 }, // extend/pmullw/trunc sequence.
462     { ISD::MUL,   MVT::v32i8,      4 }, // extend/pmullw/trunc sequence.
463     { ISD::MUL,   MVT::v16i8,      4 }, // extend/pmullw/trunc sequence.
464   };
465 
466   // Look for AVX512BW lowering tricks for custom cases.
467   if (ST->hasBWI())
468     if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD, LT.second))
469       return LT.first * Entry->Cost;
470 
471   static const CostTblEntry AVX512CostTable[] = {
472     { ISD::SHL,     MVT::v16i32,     1 },
473     { ISD::SRL,     MVT::v16i32,     1 },
474     { ISD::SRA,     MVT::v16i32,     1 },
475 
476     { ISD::SHL,     MVT::v8i64,      1 },
477     { ISD::SRL,     MVT::v8i64,      1 },
478 
479     { ISD::SRA,     MVT::v2i64,      1 },
480     { ISD::SRA,     MVT::v4i64,      1 },
481     { ISD::SRA,     MVT::v8i64,      1 },
482 
483     { ISD::MUL,     MVT::v32i8,     13 }, // extend/pmullw/trunc sequence.
484     { ISD::MUL,     MVT::v16i8,      5 }, // extend/pmullw/trunc sequence.
485     { ISD::MUL,     MVT::v16i32,     1 }, // pmulld (Skylake from agner.org)
486     { ISD::MUL,     MVT::v8i32,      1 }, // pmulld (Skylake from agner.org)
487     { ISD::MUL,     MVT::v4i32,      1 }, // pmulld (Skylake from agner.org)
488     { ISD::MUL,     MVT::v8i64,      8 }, // 3*pmuludq/3*shift/2*add
489 
490     { ISD::FADD,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
491     { ISD::FSUB,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
492     { ISD::FMUL,    MVT::v8f64,      1 }, // Skylake from http://www.agner.org/
493 
494     { ISD::FADD,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
495     { ISD::FSUB,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
496     { ISD::FMUL,    MVT::v16f32,     1 }, // Skylake from http://www.agner.org/
497   };
498 
499   if (ST->hasAVX512())
500     if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second))
501       return LT.first * Entry->Cost;
502 
503   static const CostTblEntry AVX2ShiftCostTable[] = {
504     // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
505     // customize them to detect the cases where shift amount is a scalar one.
506     { ISD::SHL,     MVT::v4i32,    1 },
507     { ISD::SRL,     MVT::v4i32,    1 },
508     { ISD::SRA,     MVT::v4i32,    1 },
509     { ISD::SHL,     MVT::v8i32,    1 },
510     { ISD::SRL,     MVT::v8i32,    1 },
511     { ISD::SRA,     MVT::v8i32,    1 },
512     { ISD::SHL,     MVT::v2i64,    1 },
513     { ISD::SRL,     MVT::v2i64,    1 },
514     { ISD::SHL,     MVT::v4i64,    1 },
515     { ISD::SRL,     MVT::v4i64,    1 },
516   };
517 
518   // Look for AVX2 lowering tricks.
519   if (ST->hasAVX2()) {
520     if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
521         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
522          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
523       // On AVX2, a packed v16i16 shift left by a constant build_vector
524       // is lowered into a vector multiply (vpmullw).
525       return getArithmeticInstrCost(Instruction::Mul, Ty, Op1Info, Op2Info,
526                                     TargetTransformInfo::OP_None,
527                                     TargetTransformInfo::OP_None);
528 
529     if (const auto *Entry = CostTableLookup(AVX2ShiftCostTable, ISD, LT.second))
530       return LT.first * Entry->Cost;
531   }
532 
533   static const CostTblEntry XOPShiftCostTable[] = {
534     // 128bit shifts take 1cy, but right shifts require negation beforehand.
535     { ISD::SHL,     MVT::v16i8,    1 },
536     { ISD::SRL,     MVT::v16i8,    2 },
537     { ISD::SRA,     MVT::v16i8,    2 },
538     { ISD::SHL,     MVT::v8i16,    1 },
539     { ISD::SRL,     MVT::v8i16,    2 },
540     { ISD::SRA,     MVT::v8i16,    2 },
541     { ISD::SHL,     MVT::v4i32,    1 },
542     { ISD::SRL,     MVT::v4i32,    2 },
543     { ISD::SRA,     MVT::v4i32,    2 },
544     { ISD::SHL,     MVT::v2i64,    1 },
545     { ISD::SRL,     MVT::v2i64,    2 },
546     { ISD::SRA,     MVT::v2i64,    2 },
547     // 256bit shifts require splitting if AVX2 didn't catch them above.
548     { ISD::SHL,     MVT::v32i8,  2+2 },
549     { ISD::SRL,     MVT::v32i8,  4+2 },
550     { ISD::SRA,     MVT::v32i8,  4+2 },
551     { ISD::SHL,     MVT::v16i16, 2+2 },
552     { ISD::SRL,     MVT::v16i16, 4+2 },
553     { ISD::SRA,     MVT::v16i16, 4+2 },
554     { ISD::SHL,     MVT::v8i32,  2+2 },
555     { ISD::SRL,     MVT::v8i32,  4+2 },
556     { ISD::SRA,     MVT::v8i32,  4+2 },
557     { ISD::SHL,     MVT::v4i64,  2+2 },
558     { ISD::SRL,     MVT::v4i64,  4+2 },
559     { ISD::SRA,     MVT::v4i64,  4+2 },
560   };
561 
562   // Look for XOP lowering tricks.
563   if (ST->hasXOP())
564     if (const auto *Entry = CostTableLookup(XOPShiftCostTable, ISD, LT.second))
565       return LT.first * Entry->Cost;
566 
567   static const CostTblEntry SSE2UniformShiftCostTable[] = {
568     // Uniform splats are cheaper for the following instructions.
569     { ISD::SHL,  MVT::v16i16, 2+2 }, // 2*psllw + split.
570     { ISD::SHL,  MVT::v8i32,  2+2 }, // 2*pslld + split.
571     { ISD::SHL,  MVT::v4i64,  2+2 }, // 2*psllq + split.
572 
573     { ISD::SRL,  MVT::v16i16, 2+2 }, // 2*psrlw + split.
574     { ISD::SRL,  MVT::v8i32,  2+2 }, // 2*psrld + split.
575     { ISD::SRL,  MVT::v4i64,  2+2 }, // 2*psrlq + split.
576 
577     { ISD::SRA,  MVT::v16i16, 2+2 }, // 2*psraw + split.
578     { ISD::SRA,  MVT::v8i32,  2+2 }, // 2*psrad + split.
579     { ISD::SRA,  MVT::v2i64,    4 }, // 2*psrad + shuffle.
580     { ISD::SRA,  MVT::v4i64,  8+2 }, // 2*(2*psrad + shuffle) + split.
581   };
582 
583   if (ST->hasSSE2() &&
584       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
585        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
586 
587     // Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table.
588     if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2())
589       return LT.first * 4; // 2*psrad + shuffle.
590 
591     if (const auto *Entry =
592             CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second))
593       return LT.first * Entry->Cost;
594   }
595 
596   if (ISD == ISD::SHL &&
597       Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
598     MVT VT = LT.second;
599     // Vector shift left by non uniform constant can be lowered
600     // into vector multiply.
601     if (((VT == MVT::v8i16 || VT == MVT::v4i32) && ST->hasSSE2()) ||
602         ((VT == MVT::v16i16 || VT == MVT::v8i32) && ST->hasAVX()))
603       ISD = ISD::MUL;
604   }
605 
606   static const CostTblEntry AVX2CostTable[] = {
607     { ISD::SHL,  MVT::v32i8,     11 }, // vpblendvb sequence.
608     { ISD::SHL,  MVT::v16i16,    10 }, // extend/vpsrlvd/pack sequence.
609 
610     { ISD::SRL,  MVT::v32i8,     11 }, // vpblendvb sequence.
611     { ISD::SRL,  MVT::v16i16,    10 }, // extend/vpsrlvd/pack sequence.
612 
613     { ISD::SRA,  MVT::v32i8,     24 }, // vpblendvb sequence.
614     { ISD::SRA,  MVT::v16i16,    10 }, // extend/vpsravd/pack sequence.
615     { ISD::SRA,  MVT::v2i64,      4 }, // srl/xor/sub sequence.
616     { ISD::SRA,  MVT::v4i64,      4 }, // srl/xor/sub sequence.
617 
618     { ISD::SUB,  MVT::v32i8,      1 }, // psubb
619     { ISD::ADD,  MVT::v32i8,      1 }, // paddb
620     { ISD::SUB,  MVT::v16i16,     1 }, // psubw
621     { ISD::ADD,  MVT::v16i16,     1 }, // paddw
622     { ISD::SUB,  MVT::v8i32,      1 }, // psubd
623     { ISD::ADD,  MVT::v8i32,      1 }, // paddd
624     { ISD::SUB,  MVT::v4i64,      1 }, // psubq
625     { ISD::ADD,  MVT::v4i64,      1 }, // paddq
626 
627     { ISD::MUL,  MVT::v32i8,     17 }, // extend/pmullw/trunc sequence.
628     { ISD::MUL,  MVT::v16i8,      7 }, // extend/pmullw/trunc sequence.
629     { ISD::MUL,  MVT::v16i16,     1 }, // pmullw
630     { ISD::MUL,  MVT::v8i32,      2 }, // pmulld (Haswell from agner.org)
631     { ISD::MUL,  MVT::v4i64,      8 }, // 3*pmuludq/3*shift/2*add
632 
633     { ISD::FADD, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
634     { ISD::FADD, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
635     { ISD::FSUB, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
636     { ISD::FSUB, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
637     { ISD::FMUL, MVT::v4f64,      1 }, // Haswell from http://www.agner.org/
638     { ISD::FMUL, MVT::v8f32,      1 }, // Haswell from http://www.agner.org/
639 
640     { ISD::FDIV, MVT::f32,        7 }, // Haswell from http://www.agner.org/
641     { ISD::FDIV, MVT::v4f32,      7 }, // Haswell from http://www.agner.org/
642     { ISD::FDIV, MVT::v8f32,     14 }, // Haswell from http://www.agner.org/
643     { ISD::FDIV, MVT::f64,       14 }, // Haswell from http://www.agner.org/
644     { ISD::FDIV, MVT::v2f64,     14 }, // Haswell from http://www.agner.org/
645     { ISD::FDIV, MVT::v4f64,     28 }, // Haswell from http://www.agner.org/
646   };
647 
648   // Look for AVX2 lowering tricks for custom cases.
649   if (ST->hasAVX2())
650     if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second))
651       return LT.first * Entry->Cost;
652 
653   static const CostTblEntry AVX1CostTable[] = {
654     // We don't have to scalarize unsupported ops. We can issue two half-sized
655     // operations and we only need to extract the upper YMM half.
656     // Two ops + 1 extract + 1 insert = 4.
657     { ISD::MUL,     MVT::v16i16,     4 },
658     { ISD::MUL,     MVT::v8i32,      4 },
659     { ISD::SUB,     MVT::v32i8,      4 },
660     { ISD::ADD,     MVT::v32i8,      4 },
661     { ISD::SUB,     MVT::v16i16,     4 },
662     { ISD::ADD,     MVT::v16i16,     4 },
663     { ISD::SUB,     MVT::v8i32,      4 },
664     { ISD::ADD,     MVT::v8i32,      4 },
665     { ISD::SUB,     MVT::v4i64,      4 },
666     { ISD::ADD,     MVT::v4i64,      4 },
667 
668     // A v4i64 multiply is custom lowered as two split v2i64 vectors that then
669     // are lowered as a series of long multiplies(3), shifts(3) and adds(2)
670     // Because we believe v4i64 to be a legal type, we must also include the
671     // extract+insert in the cost table. Therefore, the cost here is 18
672     // instead of 8.
673     { ISD::MUL,     MVT::v4i64,     18 },
674 
675     { ISD::MUL,     MVT::v32i8,     26 }, // extend/pmullw/trunc sequence.
676 
677     { ISD::FDIV,    MVT::f32,       14 }, // SNB from http://www.agner.org/
678     { ISD::FDIV,    MVT::v4f32,     14 }, // SNB from http://www.agner.org/
679     { ISD::FDIV,    MVT::v8f32,     28 }, // SNB from http://www.agner.org/
680     { ISD::FDIV,    MVT::f64,       22 }, // SNB from http://www.agner.org/
681     { ISD::FDIV,    MVT::v2f64,     22 }, // SNB from http://www.agner.org/
682     { ISD::FDIV,    MVT::v4f64,     44 }, // SNB from http://www.agner.org/
683   };
684 
685   if (ST->hasAVX())
686     if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, LT.second))
687       return LT.first * Entry->Cost;
688 
689   static const CostTblEntry SSE42CostTable[] = {
690     { ISD::FADD, MVT::f64,     1 }, // Nehalem from http://www.agner.org/
691     { ISD::FADD, MVT::f32,     1 }, // Nehalem from http://www.agner.org/
692     { ISD::FADD, MVT::v2f64,   1 }, // Nehalem from http://www.agner.org/
693     { ISD::FADD, MVT::v4f32,   1 }, // Nehalem from http://www.agner.org/
694 
695     { ISD::FSUB, MVT::f64,     1 }, // Nehalem from http://www.agner.org/
696     { ISD::FSUB, MVT::f32 ,    1 }, // Nehalem from http://www.agner.org/
697     { ISD::FSUB, MVT::v2f64,   1 }, // Nehalem from http://www.agner.org/
698     { ISD::FSUB, MVT::v4f32,   1 }, // Nehalem from http://www.agner.org/
699 
700     { ISD::FMUL, MVT::f64,     1 }, // Nehalem from http://www.agner.org/
701     { ISD::FMUL, MVT::f32,     1 }, // Nehalem from http://www.agner.org/
702     { ISD::FMUL, MVT::v2f64,   1 }, // Nehalem from http://www.agner.org/
703     { ISD::FMUL, MVT::v4f32,   1 }, // Nehalem from http://www.agner.org/
704 
705     { ISD::FDIV,  MVT::f32,   14 }, // Nehalem from http://www.agner.org/
706     { ISD::FDIV,  MVT::v4f32, 14 }, // Nehalem from http://www.agner.org/
707     { ISD::FDIV,  MVT::f64,   22 }, // Nehalem from http://www.agner.org/
708     { ISD::FDIV,  MVT::v2f64, 22 }, // Nehalem from http://www.agner.org/
709   };
710 
711   if (ST->hasSSE42())
712     if (const auto *Entry = CostTableLookup(SSE42CostTable, ISD, LT.second))
713       return LT.first * Entry->Cost;
714 
715   static const CostTblEntry SSE41CostTable[] = {
716     { ISD::SHL,  MVT::v16i8,      11 }, // pblendvb sequence.
717     { ISD::SHL,  MVT::v32i8,  2*11+2 }, // pblendvb sequence + split.
718     { ISD::SHL,  MVT::v8i16,      14 }, // pblendvb sequence.
719     { ISD::SHL,  MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
720     { ISD::SHL,  MVT::v4i32,       4 }, // pslld/paddd/cvttps2dq/pmulld
721     { ISD::SHL,  MVT::v8i32,   2*4+2 }, // pslld/paddd/cvttps2dq/pmulld + split
722 
723     { ISD::SRL,  MVT::v16i8,      12 }, // pblendvb sequence.
724     { ISD::SRL,  MVT::v32i8,  2*12+2 }, // pblendvb sequence + split.
725     { ISD::SRL,  MVT::v8i16,      14 }, // pblendvb sequence.
726     { ISD::SRL,  MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
727     { ISD::SRL,  MVT::v4i32,      11 }, // Shift each lane + blend.
728     { ISD::SRL,  MVT::v8i32,  2*11+2 }, // Shift each lane + blend + split.
729 
730     { ISD::SRA,  MVT::v16i8,      24 }, // pblendvb sequence.
731     { ISD::SRA,  MVT::v32i8,  2*24+2 }, // pblendvb sequence + split.
732     { ISD::SRA,  MVT::v8i16,      14 }, // pblendvb sequence.
733     { ISD::SRA,  MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
734     { ISD::SRA,  MVT::v4i32,      12 }, // Shift each lane + blend.
735     { ISD::SRA,  MVT::v8i32,  2*12+2 }, // Shift each lane + blend + split.
736 
737     { ISD::MUL,  MVT::v4i32,       2 }  // pmulld (Nehalem from agner.org)
738   };
739 
740   if (ST->hasSSE41())
741     if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second))
742       return LT.first * Entry->Cost;
743 
744   static const CostTblEntry SSE2CostTable[] = {
745     // We don't correctly identify costs of casts because they are marked as
746     // custom.
747     { ISD::SHL,  MVT::v16i8,      26 }, // cmpgtb sequence.
748     { ISD::SHL,  MVT::v8i16,      32 }, // cmpgtb sequence.
749     { ISD::SHL,  MVT::v4i32,     2*5 }, // We optimized this using mul.
750     { ISD::SHL,  MVT::v2i64,       4 }, // splat+shuffle sequence.
751     { ISD::SHL,  MVT::v4i64,   2*4+2 }, // splat+shuffle sequence + split.
752 
753     { ISD::SRL,  MVT::v16i8,      26 }, // cmpgtb sequence.
754     { ISD::SRL,  MVT::v8i16,      32 }, // cmpgtb sequence.
755     { ISD::SRL,  MVT::v4i32,      16 }, // Shift each lane + blend.
756     { ISD::SRL,  MVT::v2i64,       4 }, // splat+shuffle sequence.
757     { ISD::SRL,  MVT::v4i64,   2*4+2 }, // splat+shuffle sequence + split.
758 
759     { ISD::SRA,  MVT::v16i8,      54 }, // unpacked cmpgtb sequence.
760     { ISD::SRA,  MVT::v8i16,      32 }, // cmpgtb sequence.
761     { ISD::SRA,  MVT::v4i32,      16 }, // Shift each lane + blend.
762     { ISD::SRA,  MVT::v2i64,      12 }, // srl/xor/sub sequence.
763     { ISD::SRA,  MVT::v4i64,  2*12+2 }, // srl/xor/sub sequence+split.
764 
765     { ISD::MUL,  MVT::v16i8,      12 }, // extend/pmullw/trunc sequence.
766     { ISD::MUL,  MVT::v8i16,       1 }, // pmullw
767     { ISD::MUL,  MVT::v4i32,       6 }, // 3*pmuludq/4*shuffle
768     { ISD::MUL,  MVT::v2i64,       8 }, // 3*pmuludq/3*shift/2*add
769 
770     { ISD::FDIV, MVT::f32,        23 }, // Pentium IV from http://www.agner.org/
771     { ISD::FDIV, MVT::v4f32,      39 }, // Pentium IV from http://www.agner.org/
772     { ISD::FDIV, MVT::f64,        38 }, // Pentium IV from http://www.agner.org/
773     { ISD::FDIV, MVT::v2f64,      69 }, // Pentium IV from http://www.agner.org/
774   };
775 
776   if (ST->hasSSE2())
777     if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second))
778       return LT.first * Entry->Cost;
779 
780   static const CostTblEntry SSE1CostTable[] = {
781     { ISD::FDIV, MVT::f32,   17 }, // Pentium III from http://www.agner.org/
782     { ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/
783   };
784 
785   if (ST->hasSSE1())
786     if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second))
787       return LT.first * Entry->Cost;
788 
789   // It is not a good idea to vectorize division. We have to scalarize it and
790   // in the process we will often end up having to spilling regular
791   // registers. The overhead of division is going to dominate most kernels
792   // anyways so try hard to prevent vectorization of division - it is
793   // generally a bad idea. Assume somewhat arbitrarily that we have to be able
794   // to hide "20 cycles" for each lane.
795   if (LT.second.isVector() && (ISD == ISD::SDIV || ISD == ISD::SREM ||
796                                ISD == ISD::UDIV || ISD == ISD::UREM)) {
797     int ScalarCost = getArithmeticInstrCost(
798         Opcode, Ty->getScalarType(), Op1Info, Op2Info,
799         TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
800     return 20 * LT.first * LT.second.getVectorNumElements() * ScalarCost;
801   }
802 
803   // Fallback to the default implementation.
804   return BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
805 }
806 
807 int X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
808                                Type *SubTp) {
809   // 64-bit packed float vectors (v2f32) are widened to type v4f32.
810   // 64-bit packed integer vectors (v2i32) are promoted to type v2i64.
811   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
812 
813   // For Broadcasts we are splatting the first element from the first input
814   // register, so only need to reference that input and all the output
815   // registers are the same.
816   if (Kind == TTI::SK_Broadcast)
817     LT.first = 1;
818 
819   // We are going to permute multiple sources and the result will be in multiple
820   // destinations. Providing an accurate cost only for splits where the element
821   // type remains the same.
822   if (Kind == TTI::SK_PermuteSingleSrc && LT.first != 1) {
823     MVT LegalVT = LT.second;
824     if (LegalVT.isVector() &&
825         LegalVT.getVectorElementType().getSizeInBits() ==
826             Tp->getVectorElementType()->getPrimitiveSizeInBits() &&
827         LegalVT.getVectorNumElements() < Tp->getVectorNumElements()) {
828 
829       unsigned VecTySize = DL.getTypeStoreSize(Tp);
830       unsigned LegalVTSize = LegalVT.getStoreSize();
831       // Number of source vectors after legalization:
832       unsigned NumOfSrcs = (VecTySize + LegalVTSize - 1) / LegalVTSize;
833       // Number of destination vectors after legalization:
834       unsigned NumOfDests = LT.first;
835 
836       Type *SingleOpTy = VectorType::get(Tp->getVectorElementType(),
837                                          LegalVT.getVectorNumElements());
838 
839       unsigned NumOfShuffles = (NumOfSrcs - 1) * NumOfDests;
840       return NumOfShuffles *
841              getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy, 0, nullptr);
842     }
843 
844     return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
845   }
846 
847   // For 2-input shuffles, we must account for splitting the 2 inputs into many.
848   if (Kind == TTI::SK_PermuteTwoSrc && LT.first != 1) {
849     // We assume that source and destination have the same vector type.
850     int NumOfDests = LT.first;
851     int NumOfShufflesPerDest = LT.first * 2 - 1;
852     LT.first = NumOfDests * NumOfShufflesPerDest;
853   }
854 
855   static const CostTblEntry AVX512VBMIShuffleTbl[] = {
856     { TTI::SK_Reverse,          MVT::v64i8,  1 }, // vpermb
857     { TTI::SK_Reverse,          MVT::v32i8,  1 }, // vpermb
858 
859     { TTI::SK_PermuteSingleSrc, MVT::v64i8,  1 }, // vpermb
860     { TTI::SK_PermuteSingleSrc, MVT::v32i8,  1 }, // vpermb
861 
862     { TTI::SK_PermuteTwoSrc,    MVT::v64i8,  1 }, // vpermt2b
863     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,  1 }, // vpermt2b
864     { TTI::SK_PermuteTwoSrc,    MVT::v16i8,  1 }  // vpermt2b
865   };
866 
867   if (ST->hasVBMI())
868     if (const auto *Entry =
869             CostTableLookup(AVX512VBMIShuffleTbl, Kind, LT.second))
870       return LT.first * Entry->Cost;
871 
872   static const CostTblEntry AVX512BWShuffleTbl[] = {
873     { TTI::SK_Broadcast,        MVT::v32i16, 1 }, // vpbroadcastw
874     { TTI::SK_Broadcast,        MVT::v64i8,  1 }, // vpbroadcastb
875 
876     { TTI::SK_Reverse,          MVT::v32i16, 1 }, // vpermw
877     { TTI::SK_Reverse,          MVT::v16i16, 1 }, // vpermw
878     { TTI::SK_Reverse,          MVT::v64i8,  2 }, // pshufb + vshufi64x2
879 
880     { TTI::SK_PermuteSingleSrc, MVT::v32i16, 1 }, // vpermw
881     { TTI::SK_PermuteSingleSrc, MVT::v16i16, 1 }, // vpermw
882     { TTI::SK_PermuteSingleSrc, MVT::v8i16,  1 }, // vpermw
883     { TTI::SK_PermuteSingleSrc, MVT::v64i8,  8 }, // extend to v32i16
884     { TTI::SK_PermuteSingleSrc, MVT::v32i8,  3 }, // vpermw + zext/trunc
885 
886     { TTI::SK_PermuteTwoSrc,    MVT::v32i16, 1 }, // vpermt2w
887     { TTI::SK_PermuteTwoSrc,    MVT::v16i16, 1 }, // vpermt2w
888     { TTI::SK_PermuteTwoSrc,    MVT::v8i16,  1 }, // vpermt2w
889     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,  3 }, // zext + vpermt2w + trunc
890     { TTI::SK_PermuteTwoSrc,    MVT::v64i8, 19 }, // 6 * v32i8 + 1
891     { TTI::SK_PermuteTwoSrc,    MVT::v16i8,  3 }  // zext + vpermt2w + trunc
892   };
893 
894   if (ST->hasBWI())
895     if (const auto *Entry =
896             CostTableLookup(AVX512BWShuffleTbl, Kind, LT.second))
897       return LT.first * Entry->Cost;
898 
899   static const CostTblEntry AVX512ShuffleTbl[] = {
900     { TTI::SK_Broadcast,        MVT::v8f64,  1 }, // vbroadcastpd
901     { TTI::SK_Broadcast,        MVT::v16f32, 1 }, // vbroadcastps
902     { TTI::SK_Broadcast,        MVT::v8i64,  1 }, // vpbroadcastq
903     { TTI::SK_Broadcast,        MVT::v16i32, 1 }, // vpbroadcastd
904 
905     { TTI::SK_Reverse,          MVT::v8f64,  1 }, // vpermpd
906     { TTI::SK_Reverse,          MVT::v16f32, 1 }, // vpermps
907     { TTI::SK_Reverse,          MVT::v8i64,  1 }, // vpermq
908     { TTI::SK_Reverse,          MVT::v16i32, 1 }, // vpermd
909 
910     { TTI::SK_PermuteSingleSrc, MVT::v8f64,  1 }, // vpermpd
911     { TTI::SK_PermuteSingleSrc, MVT::v4f64,  1 }, // vpermpd
912     { TTI::SK_PermuteSingleSrc, MVT::v2f64,  1 }, // vpermpd
913     { TTI::SK_PermuteSingleSrc, MVT::v16f32, 1 }, // vpermps
914     { TTI::SK_PermuteSingleSrc, MVT::v8f32,  1 }, // vpermps
915     { TTI::SK_PermuteSingleSrc, MVT::v4f32,  1 }, // vpermps
916     { TTI::SK_PermuteSingleSrc, MVT::v8i64,  1 }, // vpermq
917     { TTI::SK_PermuteSingleSrc, MVT::v4i64,  1 }, // vpermq
918     { TTI::SK_PermuteSingleSrc, MVT::v2i64,  1 }, // vpermq
919     { TTI::SK_PermuteSingleSrc, MVT::v16i32, 1 }, // vpermd
920     { TTI::SK_PermuteSingleSrc, MVT::v8i32,  1 }, // vpermd
921     { TTI::SK_PermuteSingleSrc, MVT::v4i32,  1 }, // vpermd
922     { TTI::SK_PermuteSingleSrc, MVT::v16i8,  1 }, // pshufb
923 
924     { TTI::SK_PermuteTwoSrc,    MVT::v8f64,  1 }, // vpermt2pd
925     { TTI::SK_PermuteTwoSrc,    MVT::v16f32, 1 }, // vpermt2ps
926     { TTI::SK_PermuteTwoSrc,    MVT::v8i64,  1 }, // vpermt2q
927     { TTI::SK_PermuteTwoSrc,    MVT::v16i32, 1 }, // vpermt2d
928     { TTI::SK_PermuteTwoSrc,    MVT::v4f64,  1 }, // vpermt2pd
929     { TTI::SK_PermuteTwoSrc,    MVT::v8f32,  1 }, // vpermt2ps
930     { TTI::SK_PermuteTwoSrc,    MVT::v4i64,  1 }, // vpermt2q
931     { TTI::SK_PermuteTwoSrc,    MVT::v8i32,  1 }, // vpermt2d
932     { TTI::SK_PermuteTwoSrc,    MVT::v2f64,  1 }, // vpermt2pd
933     { TTI::SK_PermuteTwoSrc,    MVT::v4f32,  1 }, // vpermt2ps
934     { TTI::SK_PermuteTwoSrc,    MVT::v2i64,  1 }, // vpermt2q
935     { TTI::SK_PermuteTwoSrc,    MVT::v4i32,  1 }  // vpermt2d
936   };
937 
938   if (ST->hasAVX512())
939     if (const auto *Entry = CostTableLookup(AVX512ShuffleTbl, Kind, LT.second))
940       return LT.first * Entry->Cost;
941 
942   static const CostTblEntry AVX2ShuffleTbl[] = {
943     { TTI::SK_Broadcast, MVT::v4f64,  1 }, // vbroadcastpd
944     { TTI::SK_Broadcast, MVT::v8f32,  1 }, // vbroadcastps
945     { TTI::SK_Broadcast, MVT::v4i64,  1 }, // vpbroadcastq
946     { TTI::SK_Broadcast, MVT::v8i32,  1 }, // vpbroadcastd
947     { TTI::SK_Broadcast, MVT::v16i16, 1 }, // vpbroadcastw
948     { TTI::SK_Broadcast, MVT::v32i8,  1 }, // vpbroadcastb
949 
950     { TTI::SK_Reverse,   MVT::v4f64,  1 }, // vpermpd
951     { TTI::SK_Reverse,   MVT::v8f32,  1 }, // vpermps
952     { TTI::SK_Reverse,   MVT::v4i64,  1 }, // vpermq
953     { TTI::SK_Reverse,   MVT::v8i32,  1 }, // vpermd
954     { TTI::SK_Reverse,   MVT::v16i16, 2 }, // vperm2i128 + pshufb
955     { TTI::SK_Reverse,   MVT::v32i8,  2 }, // vperm2i128 + pshufb
956 
957     { TTI::SK_Select,    MVT::v16i16, 1 }, // vpblendvb
958     { TTI::SK_Select,    MVT::v32i8,  1 }, // vpblendvb
959 
960     { TTI::SK_PermuteSingleSrc, MVT::v4f64,  1 }, // vpermpd
961     { TTI::SK_PermuteSingleSrc, MVT::v8f32,  1 }, // vpermps
962     { TTI::SK_PermuteSingleSrc, MVT::v4i64,  1 }, // vpermq
963     { TTI::SK_PermuteSingleSrc, MVT::v8i32,  1 }, // vpermd
964     { TTI::SK_PermuteSingleSrc, MVT::v16i16, 4 }, // vperm2i128 + 2*vpshufb
965                                                   // + vpblendvb
966     { TTI::SK_PermuteSingleSrc, MVT::v32i8,  4 }, // vperm2i128 + 2*vpshufb
967                                                   // + vpblendvb
968 
969     { TTI::SK_PermuteTwoSrc,    MVT::v4f64,  3 }, // 2*vpermpd + vblendpd
970     { TTI::SK_PermuteTwoSrc,    MVT::v8f32,  3 }, // 2*vpermps + vblendps
971     { TTI::SK_PermuteTwoSrc,    MVT::v4i64,  3 }, // 2*vpermq + vpblendd
972     { TTI::SK_PermuteTwoSrc,    MVT::v8i32,  3 }, // 2*vpermd + vpblendd
973     { TTI::SK_PermuteTwoSrc,    MVT::v16i16, 7 }, // 2*vperm2i128 + 4*vpshufb
974                                                   // + vpblendvb
975     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,  7 }, // 2*vperm2i128 + 4*vpshufb
976                                                   // + vpblendvb
977   };
978 
979   if (ST->hasAVX2())
980     if (const auto *Entry = CostTableLookup(AVX2ShuffleTbl, Kind, LT.second))
981       return LT.first * Entry->Cost;
982 
983   static const CostTblEntry XOPShuffleTbl[] = {
984     { TTI::SK_PermuteSingleSrc, MVT::v4f64,   2 }, // vperm2f128 + vpermil2pd
985     { TTI::SK_PermuteSingleSrc, MVT::v8f32,   2 }, // vperm2f128 + vpermil2ps
986     { TTI::SK_PermuteSingleSrc, MVT::v4i64,   2 }, // vperm2f128 + vpermil2pd
987     { TTI::SK_PermuteSingleSrc, MVT::v8i32,   2 }, // vperm2f128 + vpermil2ps
988     { TTI::SK_PermuteSingleSrc, MVT::v16i16,  4 }, // vextractf128 + 2*vpperm
989                                                    // + vinsertf128
990     { TTI::SK_PermuteSingleSrc, MVT::v32i8,   4 }, // vextractf128 + 2*vpperm
991                                                    // + vinsertf128
992 
993     { TTI::SK_PermuteTwoSrc,    MVT::v16i16,  9 }, // 2*vextractf128 + 6*vpperm
994                                                    // + vinsertf128
995     { TTI::SK_PermuteTwoSrc,    MVT::v8i16,   1 }, // vpperm
996     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,   9 }, // 2*vextractf128 + 6*vpperm
997                                                    // + vinsertf128
998     { TTI::SK_PermuteTwoSrc,    MVT::v16i8,   1 }, // vpperm
999   };
1000 
1001   if (ST->hasXOP())
1002     if (const auto *Entry = CostTableLookup(XOPShuffleTbl, Kind, LT.second))
1003       return LT.first * Entry->Cost;
1004 
1005   static const CostTblEntry AVX1ShuffleTbl[] = {
1006     { TTI::SK_Broadcast, MVT::v4f64,  2 }, // vperm2f128 + vpermilpd
1007     { TTI::SK_Broadcast, MVT::v8f32,  2 }, // vperm2f128 + vpermilps
1008     { TTI::SK_Broadcast, MVT::v4i64,  2 }, // vperm2f128 + vpermilpd
1009     { TTI::SK_Broadcast, MVT::v8i32,  2 }, // vperm2f128 + vpermilps
1010     { TTI::SK_Broadcast, MVT::v16i16, 3 }, // vpshuflw + vpshufd + vinsertf128
1011     { TTI::SK_Broadcast, MVT::v32i8,  2 }, // vpshufb + vinsertf128
1012 
1013     { TTI::SK_Reverse,   MVT::v4f64,  2 }, // vperm2f128 + vpermilpd
1014     { TTI::SK_Reverse,   MVT::v8f32,  2 }, // vperm2f128 + vpermilps
1015     { TTI::SK_Reverse,   MVT::v4i64,  2 }, // vperm2f128 + vpermilpd
1016     { TTI::SK_Reverse,   MVT::v8i32,  2 }, // vperm2f128 + vpermilps
1017     { TTI::SK_Reverse,   MVT::v16i16, 4 }, // vextractf128 + 2*pshufb
1018                                            // + vinsertf128
1019     { TTI::SK_Reverse,   MVT::v32i8,  4 }, // vextractf128 + 2*pshufb
1020                                            // + vinsertf128
1021 
1022     { TTI::SK_Select,    MVT::v4i64,  1 }, // vblendpd
1023     { TTI::SK_Select,    MVT::v4f64,  1 }, // vblendpd
1024     { TTI::SK_Select,    MVT::v8i32,  1 }, // vblendps
1025     { TTI::SK_Select,    MVT::v8f32,  1 }, // vblendps
1026     { TTI::SK_Select,    MVT::v16i16, 3 }, // vpand + vpandn + vpor
1027     { TTI::SK_Select,    MVT::v32i8,  3 }, // vpand + vpandn + vpor
1028 
1029     { TTI::SK_PermuteSingleSrc, MVT::v4f64,  2 }, // vperm2f128 + vshufpd
1030     { TTI::SK_PermuteSingleSrc, MVT::v4i64,  2 }, // vperm2f128 + vshufpd
1031     { TTI::SK_PermuteSingleSrc, MVT::v8f32,  4 }, // 2*vperm2f128 + 2*vshufps
1032     { TTI::SK_PermuteSingleSrc, MVT::v8i32,  4 }, // 2*vperm2f128 + 2*vshufps
1033     { TTI::SK_PermuteSingleSrc, MVT::v16i16, 8 }, // vextractf128 + 4*pshufb
1034                                                   // + 2*por + vinsertf128
1035     { TTI::SK_PermuteSingleSrc, MVT::v32i8,  8 }, // vextractf128 + 4*pshufb
1036                                                   // + 2*por + vinsertf128
1037 
1038     { TTI::SK_PermuteTwoSrc,    MVT::v4f64,   3 }, // 2*vperm2f128 + vshufpd
1039     { TTI::SK_PermuteTwoSrc,    MVT::v4i64,   3 }, // 2*vperm2f128 + vshufpd
1040     { TTI::SK_PermuteTwoSrc,    MVT::v8f32,   4 }, // 2*vperm2f128 + 2*vshufps
1041     { TTI::SK_PermuteTwoSrc,    MVT::v8i32,   4 }, // 2*vperm2f128 + 2*vshufps
1042     { TTI::SK_PermuteTwoSrc,    MVT::v16i16, 15 }, // 2*vextractf128 + 8*pshufb
1043                                                    // + 4*por + vinsertf128
1044     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,  15 }, // 2*vextractf128 + 8*pshufb
1045                                                    // + 4*por + vinsertf128
1046   };
1047 
1048   if (ST->hasAVX())
1049     if (const auto *Entry = CostTableLookup(AVX1ShuffleTbl, Kind, LT.second))
1050       return LT.first * Entry->Cost;
1051 
1052   static const CostTblEntry SSE41ShuffleTbl[] = {
1053     { TTI::SK_Select,    MVT::v2i64,  1 }, // pblendw
1054     { TTI::SK_Select,    MVT::v2f64,  1 }, // movsd
1055     { TTI::SK_Select,    MVT::v4i32,  1 }, // pblendw
1056     { TTI::SK_Select,    MVT::v4f32,  1 }, // blendps
1057     { TTI::SK_Select,    MVT::v8i16,  1 }, // pblendw
1058     { TTI::SK_Select,    MVT::v16i8,  1 }  // pblendvb
1059   };
1060 
1061   if (ST->hasSSE41())
1062     if (const auto *Entry = CostTableLookup(SSE41ShuffleTbl, Kind, LT.second))
1063       return LT.first * Entry->Cost;
1064 
1065   static const CostTblEntry SSSE3ShuffleTbl[] = {
1066     { TTI::SK_Broadcast, MVT::v8i16,  1 }, // pshufb
1067     { TTI::SK_Broadcast, MVT::v16i8,  1 }, // pshufb
1068 
1069     { TTI::SK_Reverse,   MVT::v8i16,  1 }, // pshufb
1070     { TTI::SK_Reverse,   MVT::v16i8,  1 }, // pshufb
1071 
1072     { TTI::SK_Select,    MVT::v8i16,  3 }, // 2*pshufb + por
1073     { TTI::SK_Select,    MVT::v16i8,  3 }, // 2*pshufb + por
1074 
1075     { TTI::SK_PermuteSingleSrc, MVT::v8i16, 1 }, // pshufb
1076     { TTI::SK_PermuteSingleSrc, MVT::v16i8, 1 }, // pshufb
1077 
1078     { TTI::SK_PermuteTwoSrc,    MVT::v8i16, 3 }, // 2*pshufb + por
1079     { TTI::SK_PermuteTwoSrc,    MVT::v16i8, 3 }, // 2*pshufb + por
1080   };
1081 
1082   if (ST->hasSSSE3())
1083     if (const auto *Entry = CostTableLookup(SSSE3ShuffleTbl, Kind, LT.second))
1084       return LT.first * Entry->Cost;
1085 
1086   static const CostTblEntry SSE2ShuffleTbl[] = {
1087     { TTI::SK_Broadcast, MVT::v2f64,  1 }, // shufpd
1088     { TTI::SK_Broadcast, MVT::v2i64,  1 }, // pshufd
1089     { TTI::SK_Broadcast, MVT::v4i32,  1 }, // pshufd
1090     { TTI::SK_Broadcast, MVT::v8i16,  2 }, // pshuflw + pshufd
1091     { TTI::SK_Broadcast, MVT::v16i8,  3 }, // unpck + pshuflw + pshufd
1092 
1093     { TTI::SK_Reverse,   MVT::v2f64,  1 }, // shufpd
1094     { TTI::SK_Reverse,   MVT::v2i64,  1 }, // pshufd
1095     { TTI::SK_Reverse,   MVT::v4i32,  1 }, // pshufd
1096     { TTI::SK_Reverse,   MVT::v8i16,  3 }, // pshuflw + pshufhw + pshufd
1097     { TTI::SK_Reverse,   MVT::v16i8,  9 }, // 2*pshuflw + 2*pshufhw
1098                                            // + 2*pshufd + 2*unpck + packus
1099 
1100     { TTI::SK_Select,    MVT::v2i64,  1 }, // movsd
1101     { TTI::SK_Select,    MVT::v2f64,  1 }, // movsd
1102     { TTI::SK_Select,    MVT::v4i32,  2 }, // 2*shufps
1103     { TTI::SK_Select,    MVT::v8i16,  3 }, // pand + pandn + por
1104     { TTI::SK_Select,    MVT::v16i8,  3 }, // pand + pandn + por
1105 
1106     { TTI::SK_PermuteSingleSrc, MVT::v2f64,  1 }, // shufpd
1107     { TTI::SK_PermuteSingleSrc, MVT::v2i64,  1 }, // pshufd
1108     { TTI::SK_PermuteSingleSrc, MVT::v4i32,  1 }, // pshufd
1109     { TTI::SK_PermuteSingleSrc, MVT::v8i16,  5 }, // 2*pshuflw + 2*pshufhw
1110                                                   // + pshufd/unpck
1111     { TTI::SK_PermuteSingleSrc, MVT::v16i8, 10 }, // 2*pshuflw + 2*pshufhw
1112                                                   // + 2*pshufd + 2*unpck + 2*packus
1113 
1114     { TTI::SK_PermuteTwoSrc,    MVT::v2f64,  1 }, // shufpd
1115     { TTI::SK_PermuteTwoSrc,    MVT::v2i64,  1 }, // shufpd
1116     { TTI::SK_PermuteTwoSrc,    MVT::v4i32,  2 }, // 2*{unpck,movsd,pshufd}
1117     { TTI::SK_PermuteTwoSrc,    MVT::v8i16,  8 }, // blend+permute
1118     { TTI::SK_PermuteTwoSrc,    MVT::v16i8, 13 }, // blend+permute
1119   };
1120 
1121   if (ST->hasSSE2())
1122     if (const auto *Entry = CostTableLookup(SSE2ShuffleTbl, Kind, LT.second))
1123       return LT.first * Entry->Cost;
1124 
1125   static const CostTblEntry SSE1ShuffleTbl[] = {
1126     { TTI::SK_Broadcast,        MVT::v4f32, 1 }, // shufps
1127     { TTI::SK_Reverse,          MVT::v4f32, 1 }, // shufps
1128     { TTI::SK_Select,           MVT::v4f32, 2 }, // 2*shufps
1129     { TTI::SK_PermuteSingleSrc, MVT::v4f32, 1 }, // shufps
1130     { TTI::SK_PermuteTwoSrc,    MVT::v4f32, 2 }, // 2*shufps
1131   };
1132 
1133   if (ST->hasSSE1())
1134     if (const auto *Entry = CostTableLookup(SSE1ShuffleTbl, Kind, LT.second))
1135       return LT.first * Entry->Cost;
1136 
1137   return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
1138 }
1139 
1140 int X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
1141                                  const Instruction *I) {
1142   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1143   assert(ISD && "Invalid opcode");
1144 
1145   // FIXME: Need a better design of the cost table to handle non-simple types of
1146   // potential massive combinations (elem_num x src_type x dst_type).
1147 
1148   static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = {
1149     { ISD::SINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
1150     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
1151     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
1152     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
1153     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
1154     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
1155 
1156     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
1157     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
1158     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
1159     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
1160     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
1161     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
1162 
1163     { ISD::FP_TO_SINT,  MVT::v2i64,  MVT::v2f32,  1 },
1164     { ISD::FP_TO_SINT,  MVT::v4i64,  MVT::v4f32,  1 },
1165     { ISD::FP_TO_SINT,  MVT::v8i64,  MVT::v8f32,  1 },
1166     { ISD::FP_TO_SINT,  MVT::v2i64,  MVT::v2f64,  1 },
1167     { ISD::FP_TO_SINT,  MVT::v4i64,  MVT::v4f64,  1 },
1168     { ISD::FP_TO_SINT,  MVT::v8i64,  MVT::v8f64,  1 },
1169 
1170     { ISD::FP_TO_UINT,  MVT::v2i64,  MVT::v2f32,  1 },
1171     { ISD::FP_TO_UINT,  MVT::v4i64,  MVT::v4f32,  1 },
1172     { ISD::FP_TO_UINT,  MVT::v8i64,  MVT::v8f32,  1 },
1173     { ISD::FP_TO_UINT,  MVT::v2i64,  MVT::v2f64,  1 },
1174     { ISD::FP_TO_UINT,  MVT::v4i64,  MVT::v4f64,  1 },
1175     { ISD::FP_TO_UINT,  MVT::v8i64,  MVT::v8f64,  1 },
1176   };
1177 
1178   // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and
1179   // 256-bit wide vectors.
1180 
1181   static const TypeConversionCostTblEntry AVX512FConversionTbl[] = {
1182     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v8f32,  1 },
1183     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v16f32, 3 },
1184     { ISD::FP_ROUND,  MVT::v8f32,   MVT::v8f64,  1 },
1185 
1186     { ISD::TRUNCATE,  MVT::v16i8,   MVT::v16i32, 1 },
1187     { ISD::TRUNCATE,  MVT::v16i16,  MVT::v16i32, 1 },
1188     { ISD::TRUNCATE,  MVT::v8i16,   MVT::v8i64,  1 },
1189     { ISD::TRUNCATE,  MVT::v8i32,   MVT::v8i64,  1 },
1190 
1191     // v16i1 -> v16i32 - load + broadcast
1192     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
1193     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
1194     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
1195     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
1196     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1197     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1198     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
1199     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
1200     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
1201     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
1202 
1203     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
1204     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
1205     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i8,   2 },
1206     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i8,  2 },
1207     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
1208     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i16, 2 },
1209     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
1210     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
1211     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64, 26 },
1212     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64, 26 },
1213 
1214     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
1215     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
1216     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i8,   2 },
1217     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i8,   2 },
1218     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i8,   2 },
1219     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i8,   2 },
1220     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i8,  2 },
1221     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i16,  5 },
1222     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i16,  2 },
1223     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  2 },
1224     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
1225     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i16, 2 },
1226     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  2 },
1227     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i32,  1 },
1228     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  1 },
1229     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  1 },
1230     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  1 },
1231     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
1232     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
1233     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  5 },
1234     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  5 },
1235     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64, 12 },
1236     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64, 26 },
1237 
1238     { ISD::FP_TO_UINT,  MVT::v2i32,  MVT::v2f32,  1 },
1239     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  1 },
1240     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f64,  1 },
1241     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f32,  1 },
1242     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v8f64,  2 },
1243     { ISD::FP_TO_UINT,  MVT::v8i8,   MVT::v8f64,  2 },
1244     { ISD::FP_TO_UINT,  MVT::v16i32, MVT::v16f32, 1 },
1245     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v16f32, 2 },
1246     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v16f32, 2 },
1247   };
1248 
1249   static const TypeConversionCostTblEntry AVX2ConversionTbl[] = {
1250     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
1251     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
1252     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
1253     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
1254     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,   3 },
1255     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,   3 },
1256     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   3 },
1257     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   3 },
1258     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
1259     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
1260     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
1261     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
1262     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
1263     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
1264     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
1265     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
1266 
1267     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i64,  2 },
1268     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i64,  2 },
1269     { ISD::TRUNCATE,    MVT::v4i32,  MVT::v4i64,  2 },
1270     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  2 },
1271     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  2 },
1272     { ISD::TRUNCATE,    MVT::v8i32,  MVT::v8i64,  4 },
1273 
1274     { ISD::FP_EXTEND,   MVT::v8f64,  MVT::v8f32,  3 },
1275     { ISD::FP_ROUND,    MVT::v8f32,  MVT::v8f64,  3 },
1276 
1277     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  8 },
1278   };
1279 
1280   static const TypeConversionCostTblEntry AVXConversionTbl[] = {
1281     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,  6 },
1282     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,  4 },
1283     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,  7 },
1284     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,  4 },
1285     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,  6 },
1286     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,  4 },
1287     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,  7 },
1288     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,  4 },
1289     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
1290     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
1291     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16, 6 },
1292     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16, 3 },
1293     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16, 4 },
1294     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16, 4 },
1295     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32, 4 },
1296     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32, 4 },
1297 
1298     { ISD::TRUNCATE,    MVT::v16i8, MVT::v16i16, 4 },
1299     { ISD::TRUNCATE,    MVT::v8i8,  MVT::v8i32,  4 },
1300     { ISD::TRUNCATE,    MVT::v8i16, MVT::v8i32,  5 },
1301     { ISD::TRUNCATE,    MVT::v4i8,  MVT::v4i64,  4 },
1302     { ISD::TRUNCATE,    MVT::v4i16, MVT::v4i64,  4 },
1303     { ISD::TRUNCATE,    MVT::v4i32, MVT::v4i64,  4 },
1304     { ISD::TRUNCATE,    MVT::v8i32, MVT::v8i64,  9 },
1305 
1306     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i1,  3 },
1307     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i1,  3 },
1308     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i1,  8 },
1309     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i8,  3 },
1310     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i8,  3 },
1311     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i8,  8 },
1312     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i16, 3 },
1313     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i16, 3 },
1314     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
1315     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },
1316     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i32, 1 },
1317     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i32, 1 },
1318 
1319     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i1,  7 },
1320     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i1,  7 },
1321     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i1,  6 },
1322     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i8,  2 },
1323     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i8,  2 },
1324     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i8,  5 },
1325     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
1326     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i16, 2 },
1327     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
1328     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i32, 6 },
1329     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i32, 6 },
1330     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i32, 6 },
1331     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i32, 9 },
1332     // The generic code to compute the scalar overhead is currently broken.
1333     // Workaround this limitation by estimating the scalarization overhead
1334     // here. We have roughly 10 instructions per scalar element.
1335     // Multiply that by the vector width.
1336     // FIXME: remove that when PR19268 is fixed.
1337     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i64, 10 },
1338     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i64, 20 },
1339     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
1340     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
1341 
1342     { ISD::FP_TO_SINT,  MVT::v4i8,  MVT::v4f32, 1 },
1343     { ISD::FP_TO_SINT,  MVT::v8i8,  MVT::v8f32, 7 },
1344     // This node is expanded into scalarized operations but BasicTTI is overly
1345     // optimistic estimating its cost.  It computes 3 per element (one
1346     // vector-extract, one scalar conversion and one vector-insert).  The
1347     // problem is that the inserts form a read-modify-write chain so latency
1348     // should be factored in too.  Inflating the cost per element by 1.
1349     { ISD::FP_TO_UINT,  MVT::v8i32, MVT::v8f32, 8*4 },
1350     { ISD::FP_TO_UINT,  MVT::v4i32, MVT::v4f64, 4*4 },
1351 
1352     { ISD::FP_EXTEND,   MVT::v4f64,  MVT::v4f32,  1 },
1353     { ISD::FP_ROUND,    MVT::v4f32,  MVT::v4f64,  1 },
1354   };
1355 
1356   static const TypeConversionCostTblEntry SSE41ConversionTbl[] = {
1357     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8,    2 },
1358     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8,    2 },
1359     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16,   2 },
1360     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16,   2 },
1361     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32,   2 },
1362     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32,   2 },
1363 
1364     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i8,   1 },
1365     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i8,   2 },
1366     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i8,   1 },
1367     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i8,   1 },
1368     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
1369     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
1370     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   2 },
1371     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   2 },
1372     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
1373     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
1374     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  4 },
1375     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  4 },
1376     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
1377     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
1378     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
1379     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
1380     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 4 },
1381     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 4 },
1382 
1383     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  2 },
1384     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  1 },
1385     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i32,  1 },
1386     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i32,  1 },
1387     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  3 },
1388     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  3 },
1389     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 6 },
1390 
1391   };
1392 
1393   static const TypeConversionCostTblEntry SSE2ConversionTbl[] = {
1394     // These are somewhat magic numbers justified by looking at the output of
1395     // Intel's IACA, running some kernels and making sure when we take
1396     // legalization into account the throughput will be overestimated.
1397     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
1398     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
1399     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
1400     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
1401     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 },
1402     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
1403     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
1404     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
1405 
1406     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
1407     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
1408     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
1409     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
1410     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
1411     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 8 },
1412     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
1413     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
1414 
1415     { ISD::FP_TO_SINT,  MVT::v2i32,  MVT::v2f64,  3 },
1416 
1417     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i8,   1 },
1418     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i8,   6 },
1419     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i8,   2 },
1420     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i8,   3 },
1421     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,   4 },
1422     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,   8 },
1423     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
1424     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i8,   2 },
1425     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   6 },
1426     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   6 },
1427     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  3 },
1428     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  4 },
1429     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  9 },
1430     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  12 },
1431     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
1432     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i16,  2 },
1433     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
1434     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16,  10 },
1435     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  3 },
1436     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  4 },
1437     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 6 },
1438     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 8 },
1439     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  3 },
1440     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  5 },
1441 
1442     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  4 },
1443     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  2 },
1444     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 3 },
1445     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i32,  3 },
1446     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i32,  3 },
1447     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  4 },
1448     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i32, 7 },
1449     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  5 },
1450     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 10 },
1451   };
1452 
1453   std::pair<int, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src);
1454   std::pair<int, MVT> LTDest = TLI->getTypeLegalizationCost(DL, Dst);
1455 
1456   if (ST->hasSSE2() && !ST->hasAVX()) {
1457     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
1458                                                    LTDest.second, LTSrc.second))
1459       return LTSrc.first * Entry->Cost;
1460   }
1461 
1462   EVT SrcTy = TLI->getValueType(DL, Src);
1463   EVT DstTy = TLI->getValueType(DL, Dst);
1464 
1465   // The function getSimpleVT only handles simple value types.
1466   if (!SrcTy.isSimple() || !DstTy.isSimple())
1467     return BaseT::getCastInstrCost(Opcode, Dst, Src);
1468 
1469   if (ST->hasDQI())
1470     if (const auto *Entry = ConvertCostTableLookup(AVX512DQConversionTbl, ISD,
1471                                                    DstTy.getSimpleVT(),
1472                                                    SrcTy.getSimpleVT()))
1473       return Entry->Cost;
1474 
1475   if (ST->hasAVX512())
1476     if (const auto *Entry = ConvertCostTableLookup(AVX512FConversionTbl, ISD,
1477                                                    DstTy.getSimpleVT(),
1478                                                    SrcTy.getSimpleVT()))
1479       return Entry->Cost;
1480 
1481   if (ST->hasAVX2()) {
1482     if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
1483                                                    DstTy.getSimpleVT(),
1484                                                    SrcTy.getSimpleVT()))
1485       return Entry->Cost;
1486   }
1487 
1488   if (ST->hasAVX()) {
1489     if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
1490                                                    DstTy.getSimpleVT(),
1491                                                    SrcTy.getSimpleVT()))
1492       return Entry->Cost;
1493   }
1494 
1495   if (ST->hasSSE41()) {
1496     if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
1497                                                    DstTy.getSimpleVT(),
1498                                                    SrcTy.getSimpleVT()))
1499       return Entry->Cost;
1500   }
1501 
1502   if (ST->hasSSE2()) {
1503     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
1504                                                    DstTy.getSimpleVT(),
1505                                                    SrcTy.getSimpleVT()))
1506       return Entry->Cost;
1507   }
1508 
1509   return BaseT::getCastInstrCost(Opcode, Dst, Src, I);
1510 }
1511 
1512 int X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
1513                                    const Instruction *I) {
1514   // Legalize the type.
1515   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
1516 
1517   MVT MTy = LT.second;
1518 
1519   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1520   assert(ISD && "Invalid opcode");
1521 
1522   static const CostTblEntry SSE2CostTbl[] = {
1523     { ISD::SETCC,   MVT::v2i64,   8 },
1524     { ISD::SETCC,   MVT::v4i32,   1 },
1525     { ISD::SETCC,   MVT::v8i16,   1 },
1526     { ISD::SETCC,   MVT::v16i8,   1 },
1527   };
1528 
1529   static const CostTblEntry SSE42CostTbl[] = {
1530     { ISD::SETCC,   MVT::v2f64,   1 },
1531     { ISD::SETCC,   MVT::v4f32,   1 },
1532     { ISD::SETCC,   MVT::v2i64,   1 },
1533   };
1534 
1535   static const CostTblEntry AVX1CostTbl[] = {
1536     { ISD::SETCC,   MVT::v4f64,   1 },
1537     { ISD::SETCC,   MVT::v8f32,   1 },
1538     // AVX1 does not support 8-wide integer compare.
1539     { ISD::SETCC,   MVT::v4i64,   4 },
1540     { ISD::SETCC,   MVT::v8i32,   4 },
1541     { ISD::SETCC,   MVT::v16i16,  4 },
1542     { ISD::SETCC,   MVT::v32i8,   4 },
1543   };
1544 
1545   static const CostTblEntry AVX2CostTbl[] = {
1546     { ISD::SETCC,   MVT::v4i64,   1 },
1547     { ISD::SETCC,   MVT::v8i32,   1 },
1548     { ISD::SETCC,   MVT::v16i16,  1 },
1549     { ISD::SETCC,   MVT::v32i8,   1 },
1550   };
1551 
1552   static const CostTblEntry AVX512CostTbl[] = {
1553     { ISD::SETCC,   MVT::v8i64,   1 },
1554     { ISD::SETCC,   MVT::v16i32,  1 },
1555     { ISD::SETCC,   MVT::v8f64,   1 },
1556     { ISD::SETCC,   MVT::v16f32,  1 },
1557   };
1558 
1559   static const CostTblEntry AVX512BWCostTbl[] = {
1560     { ISD::SETCC,   MVT::v32i16,  1 },
1561     { ISD::SETCC,   MVT::v64i8,   1 },
1562   };
1563 
1564   if (ST->hasBWI())
1565     if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
1566       return LT.first * Entry->Cost;
1567 
1568   if (ST->hasAVX512())
1569     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
1570       return LT.first * Entry->Cost;
1571 
1572   if (ST->hasAVX2())
1573     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
1574       return LT.first * Entry->Cost;
1575 
1576   if (ST->hasAVX())
1577     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
1578       return LT.first * Entry->Cost;
1579 
1580   if (ST->hasSSE42())
1581     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
1582       return LT.first * Entry->Cost;
1583 
1584   if (ST->hasSSE2())
1585     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
1586       return LT.first * Entry->Cost;
1587 
1588   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, I);
1589 }
1590 
1591 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; }
1592 
1593 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1594                                       ArrayRef<Type *> Tys, FastMathFlags FMF,
1595                                       unsigned ScalarizationCostPassed) {
1596   // Costs should match the codegen from:
1597   // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll
1598   // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll
1599   // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll
1600   // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll
1601   // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll
1602   static const CostTblEntry AVX512CDCostTbl[] = {
1603     { ISD::CTLZ,       MVT::v8i64,   1 },
1604     { ISD::CTLZ,       MVT::v16i32,  1 },
1605     { ISD::CTLZ,       MVT::v32i16,  8 },
1606     { ISD::CTLZ,       MVT::v64i8,  20 },
1607     { ISD::CTLZ,       MVT::v4i64,   1 },
1608     { ISD::CTLZ,       MVT::v8i32,   1 },
1609     { ISD::CTLZ,       MVT::v16i16,  4 },
1610     { ISD::CTLZ,       MVT::v32i8,  10 },
1611     { ISD::CTLZ,       MVT::v2i64,   1 },
1612     { ISD::CTLZ,       MVT::v4i32,   1 },
1613     { ISD::CTLZ,       MVT::v8i16,   4 },
1614     { ISD::CTLZ,       MVT::v16i8,   4 },
1615   };
1616   static const CostTblEntry AVX512BWCostTbl[] = {
1617     { ISD::BITREVERSE, MVT::v8i64,   5 },
1618     { ISD::BITREVERSE, MVT::v16i32,  5 },
1619     { ISD::BITREVERSE, MVT::v32i16,  5 },
1620     { ISD::BITREVERSE, MVT::v64i8,   5 },
1621     { ISD::CTLZ,       MVT::v8i64,  23 },
1622     { ISD::CTLZ,       MVT::v16i32, 22 },
1623     { ISD::CTLZ,       MVT::v32i16, 18 },
1624     { ISD::CTLZ,       MVT::v64i8,  17 },
1625     { ISD::CTPOP,      MVT::v8i64,   7 },
1626     { ISD::CTPOP,      MVT::v16i32, 11 },
1627     { ISD::CTPOP,      MVT::v32i16,  9 },
1628     { ISD::CTPOP,      MVT::v64i8,   6 },
1629     { ISD::CTTZ,       MVT::v8i64,  10 },
1630     { ISD::CTTZ,       MVT::v16i32, 14 },
1631     { ISD::CTTZ,       MVT::v32i16, 12 },
1632     { ISD::CTTZ,       MVT::v64i8,   9 },
1633   };
1634   static const CostTblEntry AVX512CostTbl[] = {
1635     { ISD::BITREVERSE, MVT::v8i64,  36 },
1636     { ISD::BITREVERSE, MVT::v16i32, 24 },
1637     { ISD::CTLZ,       MVT::v8i64,  29 },
1638     { ISD::CTLZ,       MVT::v16i32, 35 },
1639     { ISD::CTPOP,      MVT::v8i64,  16 },
1640     { ISD::CTPOP,      MVT::v16i32, 24 },
1641     { ISD::CTTZ,       MVT::v8i64,  20 },
1642     { ISD::CTTZ,       MVT::v16i32, 28 },
1643   };
1644   static const CostTblEntry XOPCostTbl[] = {
1645     { ISD::BITREVERSE, MVT::v4i64,   4 },
1646     { ISD::BITREVERSE, MVT::v8i32,   4 },
1647     { ISD::BITREVERSE, MVT::v16i16,  4 },
1648     { ISD::BITREVERSE, MVT::v32i8,   4 },
1649     { ISD::BITREVERSE, MVT::v2i64,   1 },
1650     { ISD::BITREVERSE, MVT::v4i32,   1 },
1651     { ISD::BITREVERSE, MVT::v8i16,   1 },
1652     { ISD::BITREVERSE, MVT::v16i8,   1 },
1653     { ISD::BITREVERSE, MVT::i64,     3 },
1654     { ISD::BITREVERSE, MVT::i32,     3 },
1655     { ISD::BITREVERSE, MVT::i16,     3 },
1656     { ISD::BITREVERSE, MVT::i8,      3 }
1657   };
1658   static const CostTblEntry AVX2CostTbl[] = {
1659     { ISD::BITREVERSE, MVT::v4i64,   5 },
1660     { ISD::BITREVERSE, MVT::v8i32,   5 },
1661     { ISD::BITREVERSE, MVT::v16i16,  5 },
1662     { ISD::BITREVERSE, MVT::v32i8,   5 },
1663     { ISD::BSWAP,      MVT::v4i64,   1 },
1664     { ISD::BSWAP,      MVT::v8i32,   1 },
1665     { ISD::BSWAP,      MVT::v16i16,  1 },
1666     { ISD::CTLZ,       MVT::v4i64,  23 },
1667     { ISD::CTLZ,       MVT::v8i32,  18 },
1668     { ISD::CTLZ,       MVT::v16i16, 14 },
1669     { ISD::CTLZ,       MVT::v32i8,   9 },
1670     { ISD::CTPOP,      MVT::v4i64,   7 },
1671     { ISD::CTPOP,      MVT::v8i32,  11 },
1672     { ISD::CTPOP,      MVT::v16i16,  9 },
1673     { ISD::CTPOP,      MVT::v32i8,   6 },
1674     { ISD::CTTZ,       MVT::v4i64,  10 },
1675     { ISD::CTTZ,       MVT::v8i32,  14 },
1676     { ISD::CTTZ,       MVT::v16i16, 12 },
1677     { ISD::CTTZ,       MVT::v32i8,   9 },
1678     { ISD::FSQRT,      MVT::f32,     7 }, // Haswell from http://www.agner.org/
1679     { ISD::FSQRT,      MVT::v4f32,   7 }, // Haswell from http://www.agner.org/
1680     { ISD::FSQRT,      MVT::v8f32,  14 }, // Haswell from http://www.agner.org/
1681     { ISD::FSQRT,      MVT::f64,    14 }, // Haswell from http://www.agner.org/
1682     { ISD::FSQRT,      MVT::v2f64,  14 }, // Haswell from http://www.agner.org/
1683     { ISD::FSQRT,      MVT::v4f64,  28 }, // Haswell from http://www.agner.org/
1684   };
1685   static const CostTblEntry AVX1CostTbl[] = {
1686     { ISD::BITREVERSE, MVT::v4i64,  12 }, // 2 x 128-bit Op + extract/insert
1687     { ISD::BITREVERSE, MVT::v8i32,  12 }, // 2 x 128-bit Op + extract/insert
1688     { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert
1689     { ISD::BITREVERSE, MVT::v32i8,  12 }, // 2 x 128-bit Op + extract/insert
1690     { ISD::BSWAP,      MVT::v4i64,   4 },
1691     { ISD::BSWAP,      MVT::v8i32,   4 },
1692     { ISD::BSWAP,      MVT::v16i16,  4 },
1693     { ISD::CTLZ,       MVT::v4i64,  48 }, // 2 x 128-bit Op + extract/insert
1694     { ISD::CTLZ,       MVT::v8i32,  38 }, // 2 x 128-bit Op + extract/insert
1695     { ISD::CTLZ,       MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert
1696     { ISD::CTLZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
1697     { ISD::CTPOP,      MVT::v4i64,  16 }, // 2 x 128-bit Op + extract/insert
1698     { ISD::CTPOP,      MVT::v8i32,  24 }, // 2 x 128-bit Op + extract/insert
1699     { ISD::CTPOP,      MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert
1700     { ISD::CTPOP,      MVT::v32i8,  14 }, // 2 x 128-bit Op + extract/insert
1701     { ISD::CTTZ,       MVT::v4i64,  22 }, // 2 x 128-bit Op + extract/insert
1702     { ISD::CTTZ,       MVT::v8i32,  30 }, // 2 x 128-bit Op + extract/insert
1703     { ISD::CTTZ,       MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert
1704     { ISD::CTTZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
1705     { ISD::FSQRT,      MVT::f32,    14 }, // SNB from http://www.agner.org/
1706     { ISD::FSQRT,      MVT::v4f32,  14 }, // SNB from http://www.agner.org/
1707     { ISD::FSQRT,      MVT::v8f32,  28 }, // SNB from http://www.agner.org/
1708     { ISD::FSQRT,      MVT::f64,    21 }, // SNB from http://www.agner.org/
1709     { ISD::FSQRT,      MVT::v2f64,  21 }, // SNB from http://www.agner.org/
1710     { ISD::FSQRT,      MVT::v4f64,  43 }, // SNB from http://www.agner.org/
1711   };
1712   static const CostTblEntry GLMCostTbl[] = {
1713     { ISD::FSQRT, MVT::f32,   19 }, // sqrtss
1714     { ISD::FSQRT, MVT::v4f32, 37 }, // sqrtps
1715     { ISD::FSQRT, MVT::f64,   34 }, // sqrtsd
1716     { ISD::FSQRT, MVT::v2f64, 67 }, // sqrtpd
1717   };
1718   static const CostTblEntry SLMCostTbl[] = {
1719     { ISD::FSQRT, MVT::f32,   20 }, // sqrtss
1720     { ISD::FSQRT, MVT::v4f32, 40 }, // sqrtps
1721     { ISD::FSQRT, MVT::f64,   35 }, // sqrtsd
1722     { ISD::FSQRT, MVT::v2f64, 70 }, // sqrtpd
1723   };
1724   static const CostTblEntry SSE42CostTbl[] = {
1725     { ISD::FSQRT,      MVT::f32,    18 }, // Nehalem from http://www.agner.org/
1726     { ISD::FSQRT,      MVT::v4f32,  18 }, // Nehalem from http://www.agner.org/
1727   };
1728   static const CostTblEntry SSSE3CostTbl[] = {
1729     { ISD::BITREVERSE, MVT::v2i64,   5 },
1730     { ISD::BITREVERSE, MVT::v4i32,   5 },
1731     { ISD::BITREVERSE, MVT::v8i16,   5 },
1732     { ISD::BITREVERSE, MVT::v16i8,   5 },
1733     { ISD::BSWAP,      MVT::v2i64,   1 },
1734     { ISD::BSWAP,      MVT::v4i32,   1 },
1735     { ISD::BSWAP,      MVT::v8i16,   1 },
1736     { ISD::CTLZ,       MVT::v2i64,  23 },
1737     { ISD::CTLZ,       MVT::v4i32,  18 },
1738     { ISD::CTLZ,       MVT::v8i16,  14 },
1739     { ISD::CTLZ,       MVT::v16i8,   9 },
1740     { ISD::CTPOP,      MVT::v2i64,   7 },
1741     { ISD::CTPOP,      MVT::v4i32,  11 },
1742     { ISD::CTPOP,      MVT::v8i16,   9 },
1743     { ISD::CTPOP,      MVT::v16i8,   6 },
1744     { ISD::CTTZ,       MVT::v2i64,  10 },
1745     { ISD::CTTZ,       MVT::v4i32,  14 },
1746     { ISD::CTTZ,       MVT::v8i16,  12 },
1747     { ISD::CTTZ,       MVT::v16i8,   9 }
1748   };
1749   static const CostTblEntry SSE2CostTbl[] = {
1750     { ISD::BITREVERSE, MVT::v2i64,  29 },
1751     { ISD::BITREVERSE, MVT::v4i32,  27 },
1752     { ISD::BITREVERSE, MVT::v8i16,  27 },
1753     { ISD::BITREVERSE, MVT::v16i8,  20 },
1754     { ISD::BSWAP,      MVT::v2i64,   7 },
1755     { ISD::BSWAP,      MVT::v4i32,   7 },
1756     { ISD::BSWAP,      MVT::v8i16,   7 },
1757     { ISD::CTLZ,       MVT::v2i64,  25 },
1758     { ISD::CTLZ,       MVT::v4i32,  26 },
1759     { ISD::CTLZ,       MVT::v8i16,  20 },
1760     { ISD::CTLZ,       MVT::v16i8,  17 },
1761     { ISD::CTPOP,      MVT::v2i64,  12 },
1762     { ISD::CTPOP,      MVT::v4i32,  15 },
1763     { ISD::CTPOP,      MVT::v8i16,  13 },
1764     { ISD::CTPOP,      MVT::v16i8,  10 },
1765     { ISD::CTTZ,       MVT::v2i64,  14 },
1766     { ISD::CTTZ,       MVT::v4i32,  18 },
1767     { ISD::CTTZ,       MVT::v8i16,  16 },
1768     { ISD::CTTZ,       MVT::v16i8,  13 },
1769     { ISD::FSQRT,      MVT::f64,    32 }, // Nehalem from http://www.agner.org/
1770     { ISD::FSQRT,      MVT::v2f64,  32 }, // Nehalem from http://www.agner.org/
1771   };
1772   static const CostTblEntry SSE1CostTbl[] = {
1773     { ISD::FSQRT,      MVT::f32,    28 }, // Pentium III from http://www.agner.org/
1774     { ISD::FSQRT,      MVT::v4f32,  56 }, // Pentium III from http://www.agner.org/
1775   };
1776   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
1777     { ISD::BITREVERSE, MVT::i64,    14 }
1778   };
1779   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
1780     { ISD::BITREVERSE, MVT::i32,    14 },
1781     { ISD::BITREVERSE, MVT::i16,    14 },
1782     { ISD::BITREVERSE, MVT::i8,     11 }
1783   };
1784 
1785   unsigned ISD = ISD::DELETED_NODE;
1786   switch (IID) {
1787   default:
1788     break;
1789   case Intrinsic::bitreverse:
1790     ISD = ISD::BITREVERSE;
1791     break;
1792   case Intrinsic::bswap:
1793     ISD = ISD::BSWAP;
1794     break;
1795   case Intrinsic::ctlz:
1796     ISD = ISD::CTLZ;
1797     break;
1798   case Intrinsic::ctpop:
1799     ISD = ISD::CTPOP;
1800     break;
1801   case Intrinsic::cttz:
1802     ISD = ISD::CTTZ;
1803     break;
1804   case Intrinsic::sqrt:
1805     ISD = ISD::FSQRT;
1806     break;
1807   }
1808 
1809   // Legalize the type.
1810   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
1811   MVT MTy = LT.second;
1812 
1813   // Attempt to lookup cost.
1814   if (ST->isGLM())
1815     if (const auto *Entry = CostTableLookup(GLMCostTbl, ISD, MTy))
1816       return LT.first * Entry->Cost;
1817 
1818   if (ST->isSLM())
1819     if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy))
1820       return LT.first * Entry->Cost;
1821 
1822   if (ST->hasCDI())
1823     if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy))
1824       return LT.first * Entry->Cost;
1825 
1826   if (ST->hasBWI())
1827     if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
1828       return LT.first * Entry->Cost;
1829 
1830   if (ST->hasAVX512())
1831     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
1832       return LT.first * Entry->Cost;
1833 
1834   if (ST->hasXOP())
1835     if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
1836       return LT.first * Entry->Cost;
1837 
1838   if (ST->hasAVX2())
1839     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
1840       return LT.first * Entry->Cost;
1841 
1842   if (ST->hasAVX())
1843     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
1844       return LT.first * Entry->Cost;
1845 
1846   if (ST->hasSSE42())
1847     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
1848       return LT.first * Entry->Cost;
1849 
1850   if (ST->hasSSSE3())
1851     if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy))
1852       return LT.first * Entry->Cost;
1853 
1854   if (ST->hasSSE2())
1855     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
1856       return LT.first * Entry->Cost;
1857 
1858   if (ST->hasSSE1())
1859     if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
1860       return LT.first * Entry->Cost;
1861 
1862   if (ST->is64Bit())
1863     if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
1864       return LT.first * Entry->Cost;
1865 
1866   if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
1867     return LT.first * Entry->Cost;
1868 
1869   return BaseT::getIntrinsicInstrCost(IID, RetTy, Tys, FMF, ScalarizationCostPassed);
1870 }
1871 
1872 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1873                      ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) {
1874   return BaseT::getIntrinsicInstrCost(IID, RetTy, Args, FMF, VF);
1875 }
1876 
1877 int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
1878   assert(Val->isVectorTy() && "This must be a vector type");
1879 
1880   Type *ScalarType = Val->getScalarType();
1881 
1882   if (Index != -1U) {
1883     // Legalize the type.
1884     std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Val);
1885 
1886     // This type is legalized to a scalar type.
1887     if (!LT.second.isVector())
1888       return 0;
1889 
1890     // The type may be split. Normalize the index to the new type.
1891     unsigned Width = LT.second.getVectorNumElements();
1892     Index = Index % Width;
1893 
1894     // Floating point scalars are already located in index #0.
1895     if (ScalarType->isFloatingPointTy() && Index == 0)
1896       return 0;
1897   }
1898 
1899   // Add to the base cost if we know that the extracted element of a vector is
1900   // destined to be moved to and used in the integer register file.
1901   int RegisterFileMoveCost = 0;
1902   if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy())
1903     RegisterFileMoveCost = 1;
1904 
1905   return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost;
1906 }
1907 
1908 int X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
1909                                 unsigned AddressSpace, const Instruction *I) {
1910   // Handle non-power-of-two vectors such as <3 x float>
1911   if (VectorType *VTy = dyn_cast<VectorType>(Src)) {
1912     unsigned NumElem = VTy->getVectorNumElements();
1913 
1914     // Handle a few common cases:
1915     // <3 x float>
1916     if (NumElem == 3 && VTy->getScalarSizeInBits() == 32)
1917       // Cost = 64 bit store + extract + 32 bit store.
1918       return 3;
1919 
1920     // <3 x double>
1921     if (NumElem == 3 && VTy->getScalarSizeInBits() == 64)
1922       // Cost = 128 bit store + unpack + 64 bit store.
1923       return 3;
1924 
1925     // Assume that all other non-power-of-two numbers are scalarized.
1926     if (!isPowerOf2_32(NumElem)) {
1927       int Cost = BaseT::getMemoryOpCost(Opcode, VTy->getScalarType(), Alignment,
1928                                         AddressSpace);
1929       int SplitCost = getScalarizationOverhead(Src, Opcode == Instruction::Load,
1930                                                Opcode == Instruction::Store);
1931       return NumElem * Cost + SplitCost;
1932     }
1933   }
1934 
1935   // Legalize the type.
1936   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
1937   assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
1938          "Invalid Opcode");
1939 
1940   // Each load/store unit costs 1.
1941   int Cost = LT.first * 1;
1942 
1943   // This isn't exactly right. We're using slow unaligned 32-byte accesses as a
1944   // proxy for a double-pumped AVX memory interface such as on Sandybridge.
1945   if (LT.second.getStoreSize() == 32 && ST->isUnalignedMem32Slow())
1946     Cost *= 2;
1947 
1948   return Cost;
1949 }
1950 
1951 int X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy,
1952                                       unsigned Alignment,
1953                                       unsigned AddressSpace) {
1954   VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy);
1955   if (!SrcVTy)
1956     // To calculate scalar take the regular cost, without mask
1957     return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace);
1958 
1959   unsigned NumElem = SrcVTy->getVectorNumElements();
1960   VectorType *MaskTy =
1961     VectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem);
1962   if ((Opcode == Instruction::Load && !isLegalMaskedLoad(SrcVTy)) ||
1963       (Opcode == Instruction::Store && !isLegalMaskedStore(SrcVTy)) ||
1964       !isPowerOf2_32(NumElem)) {
1965     // Scalarization
1966     int MaskSplitCost = getScalarizationOverhead(MaskTy, false, true);
1967     int ScalarCompareCost = getCmpSelInstrCost(
1968         Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr);
1969     int BranchCost = getCFInstrCost(Instruction::Br);
1970     int MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost);
1971 
1972     int ValueSplitCost = getScalarizationOverhead(
1973         SrcVTy, Opcode == Instruction::Load, Opcode == Instruction::Store);
1974     int MemopCost =
1975         NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
1976                                          Alignment, AddressSpace);
1977     return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost;
1978   }
1979 
1980   // Legalize the type.
1981   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy);
1982   auto VT = TLI->getValueType(DL, SrcVTy);
1983   int Cost = 0;
1984   if (VT.isSimple() && LT.second != VT.getSimpleVT() &&
1985       LT.second.getVectorNumElements() == NumElem)
1986     // Promotion requires expand/truncate for data and a shuffle for mask.
1987     Cost += getShuffleCost(TTI::SK_Select, SrcVTy, 0, nullptr) +
1988             getShuffleCost(TTI::SK_Select, MaskTy, 0, nullptr);
1989 
1990   else if (LT.second.getVectorNumElements() > NumElem) {
1991     VectorType *NewMaskTy = VectorType::get(MaskTy->getVectorElementType(),
1992                                             LT.second.getVectorNumElements());
1993     // Expanding requires fill mask with zeroes
1994     Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, 0, MaskTy);
1995   }
1996   if (!ST->hasAVX512())
1997     return Cost + LT.first*4; // Each maskmov costs 4
1998 
1999   // AVX-512 masked load/store is cheapper
2000   return Cost+LT.first;
2001 }
2002 
2003 int X86TTIImpl::getAddressComputationCost(Type *Ty, ScalarEvolution *SE,
2004                                           const SCEV *Ptr) {
2005   // Address computations in vectorized code with non-consecutive addresses will
2006   // likely result in more instructions compared to scalar code where the
2007   // computation can more often be merged into the index mode. The resulting
2008   // extra micro-ops can significantly decrease throughput.
2009   unsigned NumVectorInstToHideOverhead = 10;
2010 
2011   // Cost modeling of Strided Access Computation is hidden by the indexing
2012   // modes of X86 regardless of the stride value. We dont believe that there
2013   // is a difference between constant strided access in gerenal and constant
2014   // strided value which is less than or equal to 64.
2015   // Even in the case of (loop invariant) stride whose value is not known at
2016   // compile time, the address computation will not incur more than one extra
2017   // ADD instruction.
2018   if (Ty->isVectorTy() && SE) {
2019     if (!BaseT::isStridedAccess(Ptr))
2020       return NumVectorInstToHideOverhead;
2021     if (!BaseT::getConstantStrideStep(SE, Ptr))
2022       return 1;
2023   }
2024 
2025   return BaseT::getAddressComputationCost(Ty, SE, Ptr);
2026 }
2027 
2028 int X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, Type *ValTy,
2029                                            bool IsPairwise) {
2030 
2031   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
2032 
2033   MVT MTy = LT.second;
2034 
2035   int ISD = TLI->InstructionOpcodeToISD(Opcode);
2036   assert(ISD && "Invalid opcode");
2037 
2038   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
2039   // and make it as the cost.
2040 
2041   static const CostTblEntry SSE42CostTblPairWise[] = {
2042     { ISD::FADD,  MVT::v2f64,   2 },
2043     { ISD::FADD,  MVT::v4f32,   4 },
2044     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
2045     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.5".
2046     { ISD::ADD,   MVT::v8i16,   5 },
2047   };
2048 
2049   static const CostTblEntry AVX1CostTblPairWise[] = {
2050     { ISD::FADD,  MVT::v4f32,   4 },
2051     { ISD::FADD,  MVT::v4f64,   5 },
2052     { ISD::FADD,  MVT::v8f32,   7 },
2053     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
2054     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.5".
2055     { ISD::ADD,   MVT::v4i64,   5 },      // The data reported by the IACA tool is "4.8".
2056     { ISD::ADD,   MVT::v8i16,   5 },
2057     { ISD::ADD,   MVT::v8i32,   5 },
2058   };
2059 
2060   static const CostTblEntry SSE42CostTblNoPairWise[] = {
2061     { ISD::FADD,  MVT::v2f64,   2 },
2062     { ISD::FADD,  MVT::v4f32,   4 },
2063     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
2064     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.3".
2065     { ISD::ADD,   MVT::v8i16,   4 },      // The data reported by the IACA tool is "4.3".
2066   };
2067 
2068   static const CostTblEntry AVX1CostTblNoPairWise[] = {
2069     { ISD::FADD,  MVT::v4f32,   3 },
2070     { ISD::FADD,  MVT::v4f64,   3 },
2071     { ISD::FADD,  MVT::v8f32,   4 },
2072     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
2073     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "2.8".
2074     { ISD::ADD,   MVT::v4i64,   3 },
2075     { ISD::ADD,   MVT::v8i16,   4 },
2076     { ISD::ADD,   MVT::v8i32,   5 },
2077   };
2078 
2079   if (IsPairwise) {
2080     if (ST->hasAVX())
2081       if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy))
2082         return LT.first * Entry->Cost;
2083 
2084     if (ST->hasSSE42())
2085       if (const auto *Entry = CostTableLookup(SSE42CostTblPairWise, ISD, MTy))
2086         return LT.first * Entry->Cost;
2087   } else {
2088     if (ST->hasAVX())
2089       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
2090         return LT.first * Entry->Cost;
2091 
2092     if (ST->hasSSE42())
2093       if (const auto *Entry = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy))
2094         return LT.first * Entry->Cost;
2095   }
2096 
2097   return BaseT::getArithmeticReductionCost(Opcode, ValTy, IsPairwise);
2098 }
2099 
2100 int X86TTIImpl::getMinMaxReductionCost(Type *ValTy, Type *CondTy,
2101                                        bool IsPairwise, bool IsUnsigned) {
2102   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
2103 
2104   MVT MTy = LT.second;
2105 
2106   int ISD;
2107   if (ValTy->isIntOrIntVectorTy()) {
2108     ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN;
2109   } else {
2110     assert(ValTy->isFPOrFPVectorTy() &&
2111            "Expected float point or integer vector type.");
2112     ISD = ISD::FMINNUM;
2113   }
2114 
2115   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
2116   // and make it as the cost.
2117 
2118   static const CostTblEntry SSE42CostTblPairWise[] = {
2119       {ISD::FMINNUM, MVT::v2f64, 3},
2120       {ISD::FMINNUM, MVT::v4f32, 2},
2121       {ISD::SMIN, MVT::v2i64, 7}, // The data reported by the IACA is "6.8"
2122       {ISD::UMIN, MVT::v2i64, 8}, // The data reported by the IACA is "8.6"
2123       {ISD::SMIN, MVT::v4i32, 1}, // The data reported by the IACA is "1.5"
2124       {ISD::UMIN, MVT::v4i32, 2}, // The data reported by the IACA is "1.8"
2125       {ISD::SMIN, MVT::v8i16, 2},
2126       {ISD::UMIN, MVT::v8i16, 2},
2127   };
2128 
2129   static const CostTblEntry AVX1CostTblPairWise[] = {
2130       {ISD::FMINNUM, MVT::v4f32, 1},
2131       {ISD::FMINNUM, MVT::v4f64, 1},
2132       {ISD::FMINNUM, MVT::v8f32, 2},
2133       {ISD::SMIN, MVT::v2i64, 3},
2134       {ISD::UMIN, MVT::v2i64, 3},
2135       {ISD::SMIN, MVT::v4i32, 1},
2136       {ISD::UMIN, MVT::v4i32, 1},
2137       {ISD::SMIN, MVT::v8i16, 1},
2138       {ISD::UMIN, MVT::v8i16, 1},
2139       {ISD::SMIN, MVT::v8i32, 3},
2140       {ISD::UMIN, MVT::v8i32, 3},
2141   };
2142 
2143   static const CostTblEntry AVX2CostTblPairWise[] = {
2144       {ISD::SMIN, MVT::v4i64, 2},
2145       {ISD::UMIN, MVT::v4i64, 2},
2146       {ISD::SMIN, MVT::v8i32, 1},
2147       {ISD::UMIN, MVT::v8i32, 1},
2148       {ISD::SMIN, MVT::v16i16, 1},
2149       {ISD::UMIN, MVT::v16i16, 1},
2150       {ISD::SMIN, MVT::v32i8, 2},
2151       {ISD::UMIN, MVT::v32i8, 2},
2152   };
2153 
2154   static const CostTblEntry AVX512CostTblPairWise[] = {
2155       {ISD::FMINNUM, MVT::v8f64, 1},
2156       {ISD::FMINNUM, MVT::v16f32, 2},
2157       {ISD::SMIN, MVT::v8i64, 2},
2158       {ISD::UMIN, MVT::v8i64, 2},
2159       {ISD::SMIN, MVT::v16i32, 1},
2160       {ISD::UMIN, MVT::v16i32, 1},
2161   };
2162 
2163   static const CostTblEntry SSE42CostTblNoPairWise[] = {
2164       {ISD::FMINNUM, MVT::v2f64, 3},
2165       {ISD::FMINNUM, MVT::v4f32, 3},
2166       {ISD::SMIN, MVT::v2i64, 7}, // The data reported by the IACA is "6.8"
2167       {ISD::UMIN, MVT::v2i64, 9}, // The data reported by the IACA is "8.6"
2168       {ISD::SMIN, MVT::v4i32, 1}, // The data reported by the IACA is "1.5"
2169       {ISD::UMIN, MVT::v4i32, 2}, // The data reported by the IACA is "1.8"
2170       {ISD::SMIN, MVT::v8i16, 1}, // The data reported by the IACA is "1.5"
2171       {ISD::UMIN, MVT::v8i16, 2}, // The data reported by the IACA is "1.8"
2172   };
2173 
2174   static const CostTblEntry AVX1CostTblNoPairWise[] = {
2175       {ISD::FMINNUM, MVT::v4f32, 1},
2176       {ISD::FMINNUM, MVT::v4f64, 1},
2177       {ISD::FMINNUM, MVT::v8f32, 1},
2178       {ISD::SMIN, MVT::v2i64, 3},
2179       {ISD::UMIN, MVT::v2i64, 3},
2180       {ISD::SMIN, MVT::v4i32, 1},
2181       {ISD::UMIN, MVT::v4i32, 1},
2182       {ISD::SMIN, MVT::v8i16, 1},
2183       {ISD::UMIN, MVT::v8i16, 1},
2184       {ISD::SMIN, MVT::v8i32, 2},
2185       {ISD::UMIN, MVT::v8i32, 2},
2186   };
2187 
2188   static const CostTblEntry AVX2CostTblNoPairWise[] = {
2189       {ISD::SMIN, MVT::v4i64, 1},
2190       {ISD::UMIN, MVT::v4i64, 1},
2191       {ISD::SMIN, MVT::v8i32, 1},
2192       {ISD::UMIN, MVT::v8i32, 1},
2193       {ISD::SMIN, MVT::v16i16, 1},
2194       {ISD::UMIN, MVT::v16i16, 1},
2195       {ISD::SMIN, MVT::v32i8, 1},
2196       {ISD::UMIN, MVT::v32i8, 1},
2197   };
2198 
2199   static const CostTblEntry AVX512CostTblNoPairWise[] = {
2200       {ISD::FMINNUM, MVT::v8f64, 1},
2201       {ISD::FMINNUM, MVT::v16f32, 2},
2202       {ISD::SMIN, MVT::v8i64, 1},
2203       {ISD::UMIN, MVT::v8i64, 1},
2204       {ISD::SMIN, MVT::v16i32, 1},
2205       {ISD::UMIN, MVT::v16i32, 1},
2206   };
2207 
2208   if (IsPairwise) {
2209     if (ST->hasAVX512())
2210       if (const auto *Entry = CostTableLookup(AVX512CostTblPairWise, ISD, MTy))
2211         return LT.first * Entry->Cost;
2212 
2213     if (ST->hasAVX2())
2214       if (const auto *Entry = CostTableLookup(AVX2CostTblPairWise, ISD, MTy))
2215         return LT.first * Entry->Cost;
2216 
2217     if (ST->hasAVX())
2218       if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy))
2219         return LT.first * Entry->Cost;
2220 
2221     if (ST->hasSSE42())
2222       if (const auto *Entry = CostTableLookup(SSE42CostTblPairWise, ISD, MTy))
2223         return LT.first * Entry->Cost;
2224   } else {
2225     if (ST->hasAVX512())
2226       if (const auto *Entry =
2227               CostTableLookup(AVX512CostTblNoPairWise, ISD, MTy))
2228         return LT.first * Entry->Cost;
2229 
2230     if (ST->hasAVX2())
2231       if (const auto *Entry = CostTableLookup(AVX2CostTblNoPairWise, ISD, MTy))
2232         return LT.first * Entry->Cost;
2233 
2234     if (ST->hasAVX())
2235       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
2236         return LT.first * Entry->Cost;
2237 
2238     if (ST->hasSSE42())
2239       if (const auto *Entry = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy))
2240         return LT.first * Entry->Cost;
2241   }
2242 
2243   return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsPairwise, IsUnsigned);
2244 }
2245 
2246 /// Calculate the cost of materializing a 64-bit value. This helper
2247 /// method might only calculate a fraction of a larger immediate. Therefore it
2248 /// is valid to return a cost of ZERO.
2249 int X86TTIImpl::getIntImmCost(int64_t Val) {
2250   if (Val == 0)
2251     return TTI::TCC_Free;
2252 
2253   if (isInt<32>(Val))
2254     return TTI::TCC_Basic;
2255 
2256   return 2 * TTI::TCC_Basic;
2257 }
2258 
2259 int X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) {
2260   assert(Ty->isIntegerTy());
2261 
2262   unsigned BitSize = Ty->getPrimitiveSizeInBits();
2263   if (BitSize == 0)
2264     return ~0U;
2265 
2266   // Never hoist constants larger than 128bit, because this might lead to
2267   // incorrect code generation or assertions in codegen.
2268   // Fixme: Create a cost model for types larger than i128 once the codegen
2269   // issues have been fixed.
2270   if (BitSize > 128)
2271     return TTI::TCC_Free;
2272 
2273   if (Imm == 0)
2274     return TTI::TCC_Free;
2275 
2276   // Sign-extend all constants to a multiple of 64-bit.
2277   APInt ImmVal = Imm;
2278   if (BitSize % 64 != 0)
2279     ImmVal = Imm.sext(alignTo(BitSize, 64));
2280 
2281   // Split the constant into 64-bit chunks and calculate the cost for each
2282   // chunk.
2283   int Cost = 0;
2284   for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
2285     APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
2286     int64_t Val = Tmp.getSExtValue();
2287     Cost += getIntImmCost(Val);
2288   }
2289   // We need at least one instruction to materialize the constant.
2290   return std::max(1, Cost);
2291 }
2292 
2293 int X86TTIImpl::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
2294                               Type *Ty) {
2295   assert(Ty->isIntegerTy());
2296 
2297   unsigned BitSize = Ty->getPrimitiveSizeInBits();
2298   // There is no cost model for constants with a bit size of 0. Return TCC_Free
2299   // here, so that constant hoisting will ignore this constant.
2300   if (BitSize == 0)
2301     return TTI::TCC_Free;
2302 
2303   unsigned ImmIdx = ~0U;
2304   switch (Opcode) {
2305   default:
2306     return TTI::TCC_Free;
2307   case Instruction::GetElementPtr:
2308     // Always hoist the base address of a GetElementPtr. This prevents the
2309     // creation of new constants for every base constant that gets constant
2310     // folded with the offset.
2311     if (Idx == 0)
2312       return 2 * TTI::TCC_Basic;
2313     return TTI::TCC_Free;
2314   case Instruction::Store:
2315     ImmIdx = 0;
2316     break;
2317   case Instruction::ICmp:
2318     // This is an imperfect hack to prevent constant hoisting of
2319     // compares that might be trying to check if a 64-bit value fits in
2320     // 32-bits. The backend can optimize these cases using a right shift by 32.
2321     // Ideally we would check the compare predicate here. There also other
2322     // similar immediates the backend can use shifts for.
2323     if (Idx == 1 && Imm.getBitWidth() == 64) {
2324       uint64_t ImmVal = Imm.getZExtValue();
2325       if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff)
2326         return TTI::TCC_Free;
2327     }
2328     ImmIdx = 1;
2329     break;
2330   case Instruction::And:
2331     // We support 64-bit ANDs with immediates with 32-bits of leading zeroes
2332     // by using a 32-bit operation with implicit zero extension. Detect such
2333     // immediates here as the normal path expects bit 31 to be sign extended.
2334     if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue()))
2335       return TTI::TCC_Free;
2336     ImmIdx = 1;
2337     break;
2338   case Instruction::Add:
2339   case Instruction::Sub:
2340     // For add/sub, we can use the opposite instruction for INT32_MIN.
2341     if (Idx == 1 && Imm.getBitWidth() == 64 && Imm.getZExtValue() == 0x80000000)
2342       return TTI::TCC_Free;
2343     ImmIdx = 1;
2344     break;
2345   case Instruction::Mul:
2346   case Instruction::UDiv:
2347   case Instruction::SDiv:
2348   case Instruction::URem:
2349   case Instruction::SRem:
2350   case Instruction::Or:
2351   case Instruction::Xor:
2352     ImmIdx = 1;
2353     break;
2354   // Always return TCC_Free for the shift value of a shift instruction.
2355   case Instruction::Shl:
2356   case Instruction::LShr:
2357   case Instruction::AShr:
2358     if (Idx == 1)
2359       return TTI::TCC_Free;
2360     break;
2361   case Instruction::Trunc:
2362   case Instruction::ZExt:
2363   case Instruction::SExt:
2364   case Instruction::IntToPtr:
2365   case Instruction::PtrToInt:
2366   case Instruction::BitCast:
2367   case Instruction::PHI:
2368   case Instruction::Call:
2369   case Instruction::Select:
2370   case Instruction::Ret:
2371   case Instruction::Load:
2372     break;
2373   }
2374 
2375   if (Idx == ImmIdx) {
2376     int NumConstants = divideCeil(BitSize, 64);
2377     int Cost = X86TTIImpl::getIntImmCost(Imm, Ty);
2378     return (Cost <= NumConstants * TTI::TCC_Basic)
2379                ? static_cast<int>(TTI::TCC_Free)
2380                : Cost;
2381   }
2382 
2383   return X86TTIImpl::getIntImmCost(Imm, Ty);
2384 }
2385 
2386 int X86TTIImpl::getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
2387                               Type *Ty) {
2388   assert(Ty->isIntegerTy());
2389 
2390   unsigned BitSize = Ty->getPrimitiveSizeInBits();
2391   // There is no cost model for constants with a bit size of 0. Return TCC_Free
2392   // here, so that constant hoisting will ignore this constant.
2393   if (BitSize == 0)
2394     return TTI::TCC_Free;
2395 
2396   switch (IID) {
2397   default:
2398     return TTI::TCC_Free;
2399   case Intrinsic::sadd_with_overflow:
2400   case Intrinsic::uadd_with_overflow:
2401   case Intrinsic::ssub_with_overflow:
2402   case Intrinsic::usub_with_overflow:
2403   case Intrinsic::smul_with_overflow:
2404   case Intrinsic::umul_with_overflow:
2405     if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
2406       return TTI::TCC_Free;
2407     break;
2408   case Intrinsic::experimental_stackmap:
2409     if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
2410       return TTI::TCC_Free;
2411     break;
2412   case Intrinsic::experimental_patchpoint_void:
2413   case Intrinsic::experimental_patchpoint_i64:
2414     if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
2415       return TTI::TCC_Free;
2416     break;
2417   }
2418   return X86TTIImpl::getIntImmCost(Imm, Ty);
2419 }
2420 
2421 unsigned X86TTIImpl::getUserCost(const User *U,
2422                                  ArrayRef<const Value *> Operands) {
2423   if (isa<StoreInst>(U)) {
2424     Value *Ptr = U->getOperand(1);
2425     // Store instruction with index and scale costs 2 Uops.
2426     // Check the preceding GEP to identify non-const indices.
2427     if (auto GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
2428       if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); }))
2429         return TTI::TCC_Basic * 2;
2430     }
2431     return TTI::TCC_Basic;
2432   }
2433   return BaseT::getUserCost(U, Operands);
2434 }
2435 
2436 // Return an average cost of Gather / Scatter instruction, maybe improved later
2437 int X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy, Value *Ptr,
2438                                 unsigned Alignment, unsigned AddressSpace) {
2439 
2440   assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost");
2441   unsigned VF = SrcVTy->getVectorNumElements();
2442 
2443   // Try to reduce index size from 64 bit (default for GEP)
2444   // to 32. It is essential for VF 16. If the index can't be reduced to 32, the
2445   // operation will use 16 x 64 indices which do not fit in a zmm and needs
2446   // to split. Also check that the base pointer is the same for all lanes,
2447   // and that there's at most one variable index.
2448   auto getIndexSizeInBits = [](Value *Ptr, const DataLayout& DL) {
2449     unsigned IndexSize = DL.getPointerSizeInBits();
2450     GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
2451     if (IndexSize < 64 || !GEP)
2452       return IndexSize;
2453 
2454     unsigned NumOfVarIndices = 0;
2455     Value *Ptrs = GEP->getPointerOperand();
2456     if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs))
2457       return IndexSize;
2458     for (unsigned i = 1; i < GEP->getNumOperands(); ++i) {
2459       if (isa<Constant>(GEP->getOperand(i)))
2460         continue;
2461       Type *IndxTy = GEP->getOperand(i)->getType();
2462       if (IndxTy->isVectorTy())
2463         IndxTy = IndxTy->getVectorElementType();
2464       if ((IndxTy->getPrimitiveSizeInBits() == 64 &&
2465           !isa<SExtInst>(GEP->getOperand(i))) ||
2466          ++NumOfVarIndices > 1)
2467         return IndexSize; // 64
2468     }
2469     return (unsigned)32;
2470   };
2471 
2472 
2473   // Trying to reduce IndexSize to 32 bits for vector 16.
2474   // By default the IndexSize is equal to pointer size.
2475   unsigned IndexSize = (ST->hasAVX512() && VF >= 16)
2476                            ? getIndexSizeInBits(Ptr, DL)
2477                            : DL.getPointerSizeInBits();
2478 
2479   Type *IndexVTy = VectorType::get(IntegerType::get(SrcVTy->getContext(),
2480                                                     IndexSize), VF);
2481   std::pair<int, MVT> IdxsLT = TLI->getTypeLegalizationCost(DL, IndexVTy);
2482   std::pair<int, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, SrcVTy);
2483   int SplitFactor = std::max(IdxsLT.first, SrcLT.first);
2484   if (SplitFactor > 1) {
2485     // Handle splitting of vector of pointers
2486     Type *SplitSrcTy = VectorType::get(SrcVTy->getScalarType(), VF / SplitFactor);
2487     return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment,
2488                                          AddressSpace);
2489   }
2490 
2491   // The gather / scatter cost is given by Intel architects. It is a rough
2492   // number since we are looking at one instruction in a time.
2493   const int GSOverhead = (Opcode == Instruction::Load)
2494                              ? ST->getGatherOverhead()
2495                              : ST->getScatterOverhead();
2496   return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
2497                                            Alignment, AddressSpace);
2498 }
2499 
2500 /// Return the cost of full scalarization of gather / scatter operation.
2501 ///
2502 /// Opcode - Load or Store instruction.
2503 /// SrcVTy - The type of the data vector that should be gathered or scattered.
2504 /// VariableMask - The mask is non-constant at compile time.
2505 /// Alignment - Alignment for one element.
2506 /// AddressSpace - pointer[s] address space.
2507 ///
2508 int X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy,
2509                                 bool VariableMask, unsigned Alignment,
2510                                 unsigned AddressSpace) {
2511   unsigned VF = SrcVTy->getVectorNumElements();
2512 
2513   int MaskUnpackCost = 0;
2514   if (VariableMask) {
2515     VectorType *MaskTy =
2516       VectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF);
2517     MaskUnpackCost = getScalarizationOverhead(MaskTy, false, true);
2518     int ScalarCompareCost =
2519       getCmpSelInstrCost(Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()),
2520                          nullptr);
2521     int BranchCost = getCFInstrCost(Instruction::Br);
2522     MaskUnpackCost += VF * (BranchCost + ScalarCompareCost);
2523   }
2524 
2525   // The cost of the scalar loads/stores.
2526   int MemoryOpCost = VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
2527                                           Alignment, AddressSpace);
2528 
2529   int InsertExtractCost = 0;
2530   if (Opcode == Instruction::Load)
2531     for (unsigned i = 0; i < VF; ++i)
2532       // Add the cost of inserting each scalar load into the vector
2533       InsertExtractCost +=
2534         getVectorInstrCost(Instruction::InsertElement, SrcVTy, i);
2535   else
2536     for (unsigned i = 0; i < VF; ++i)
2537       // Add the cost of extracting each element out of the data vector
2538       InsertExtractCost +=
2539         getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i);
2540 
2541   return MemoryOpCost + MaskUnpackCost + InsertExtractCost;
2542 }
2543 
2544 /// Calculate the cost of Gather / Scatter operation
2545 int X86TTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *SrcVTy,
2546                                        Value *Ptr, bool VariableMask,
2547                                        unsigned Alignment) {
2548   assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter");
2549   unsigned VF = SrcVTy->getVectorNumElements();
2550   PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
2551   if (!PtrTy && Ptr->getType()->isVectorTy())
2552     PtrTy = dyn_cast<PointerType>(Ptr->getType()->getVectorElementType());
2553   assert(PtrTy && "Unexpected type for Ptr argument");
2554   unsigned AddressSpace = PtrTy->getAddressSpace();
2555 
2556   bool Scalarize = false;
2557   if ((Opcode == Instruction::Load && !isLegalMaskedGather(SrcVTy)) ||
2558       (Opcode == Instruction::Store && !isLegalMaskedScatter(SrcVTy)))
2559     Scalarize = true;
2560   // Gather / Scatter for vector 2 is not profitable on KNL / SKX
2561   // Vector-4 of gather/scatter instruction does not exist on KNL.
2562   // We can extend it to 8 elements, but zeroing upper bits of
2563   // the mask vector will add more instructions. Right now we give the scalar
2564   // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter instruction
2565   // is better in the VariableMask case.
2566   if (ST->hasAVX512() && (VF == 2 || (VF == 4 && !ST->hasVLX())))
2567     Scalarize = true;
2568 
2569   if (Scalarize)
2570     return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment,
2571                            AddressSpace);
2572 
2573   return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace);
2574 }
2575 
2576 bool X86TTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1,
2577                                TargetTransformInfo::LSRCost &C2) {
2578     // X86 specific here are "instruction number 1st priority".
2579     return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost,
2580                     C1.NumIVMuls, C1.NumBaseAdds,
2581                     C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
2582            std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost,
2583                     C2.NumIVMuls, C2.NumBaseAdds,
2584                     C2.ScaleCost, C2.ImmCost, C2.SetupCost);
2585 }
2586 
2587 bool X86TTIImpl::canMacroFuseCmp() {
2588   return ST->hasMacroFusion();
2589 }
2590 
2591 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy) {
2592   // The backend can't handle a single element vector.
2593   if (isa<VectorType>(DataTy) && DataTy->getVectorNumElements() == 1)
2594     return false;
2595   Type *ScalarTy = DataTy->getScalarType();
2596   int DataWidth = isa<PointerType>(ScalarTy) ?
2597     DL.getPointerSizeInBits() : ScalarTy->getPrimitiveSizeInBits();
2598 
2599   return ((DataWidth == 32 || DataWidth == 64) && ST->hasAVX()) ||
2600          ((DataWidth == 8 || DataWidth == 16) && ST->hasBWI());
2601 }
2602 
2603 bool X86TTIImpl::isLegalMaskedStore(Type *DataType) {
2604   return isLegalMaskedLoad(DataType);
2605 }
2606 
2607 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy) {
2608   // This function is called now in two cases: from the Loop Vectorizer
2609   // and from the Scalarizer.
2610   // When the Loop Vectorizer asks about legality of the feature,
2611   // the vectorization factor is not calculated yet. The Loop Vectorizer
2612   // sends a scalar type and the decision is based on the width of the
2613   // scalar element.
2614   // Later on, the cost model will estimate usage this intrinsic based on
2615   // the vector type.
2616   // The Scalarizer asks again about legality. It sends a vector type.
2617   // In this case we can reject non-power-of-2 vectors.
2618   // We also reject single element vectors as the type legalizer can't
2619   // scalarize it.
2620   if (isa<VectorType>(DataTy)) {
2621     unsigned NumElts = DataTy->getVectorNumElements();
2622     if (NumElts == 1 || !isPowerOf2_32(NumElts))
2623       return false;
2624   }
2625   Type *ScalarTy = DataTy->getScalarType();
2626   int DataWidth = isa<PointerType>(ScalarTy) ?
2627     DL.getPointerSizeInBits() : ScalarTy->getPrimitiveSizeInBits();
2628 
2629   // Some CPUs have better gather performance than others.
2630   // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only
2631   // enable gather with a -march.
2632   return (DataWidth == 32 || DataWidth == 64) &&
2633          (ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2()));
2634 }
2635 
2636 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType) {
2637   // AVX2 doesn't support scatter
2638   if (!ST->hasAVX512())
2639     return false;
2640   return isLegalMaskedGather(DataType);
2641 }
2642 
2643 bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) {
2644   EVT VT = TLI->getValueType(DL, DataType);
2645   return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT);
2646 }
2647 
2648 bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
2649   return false;
2650 }
2651 
2652 bool X86TTIImpl::areInlineCompatible(const Function *Caller,
2653                                      const Function *Callee) const {
2654   const TargetMachine &TM = getTLI()->getTargetMachine();
2655 
2656   // Work this as a subsetting of subtarget features.
2657   const FeatureBitset &CallerBits =
2658       TM.getSubtargetImpl(*Caller)->getFeatureBits();
2659   const FeatureBitset &CalleeBits =
2660       TM.getSubtargetImpl(*Callee)->getFeatureBits();
2661 
2662   // FIXME: This is likely too limiting as it will include subtarget features
2663   // that we might not care about for inlining, but it is conservatively
2664   // correct.
2665   return (CallerBits & CalleeBits) == CalleeBits;
2666 }
2667 
2668 const X86TTIImpl::TTI::MemCmpExpansionOptions *
2669 X86TTIImpl::enableMemCmpExpansion(bool IsZeroCmp) const {
2670   // Only enable vector loads for equality comparison.
2671   // Right now the vector version is not as fast, see #33329.
2672   static const auto ThreeWayOptions = [this]() {
2673     TTI::MemCmpExpansionOptions Options;
2674     if (ST->is64Bit()) {
2675       Options.LoadSizes.push_back(8);
2676     }
2677     Options.LoadSizes.push_back(4);
2678     Options.LoadSizes.push_back(2);
2679     Options.LoadSizes.push_back(1);
2680     return Options;
2681   }();
2682   static const auto EqZeroOptions = [this]() {
2683     TTI::MemCmpExpansionOptions Options;
2684     // TODO: enable AVX512 when the DAG is ready.
2685     // if (ST->hasAVX512()) Options.LoadSizes.push_back(64);
2686     if (ST->hasAVX2()) Options.LoadSizes.push_back(32);
2687     if (ST->hasSSE2()) Options.LoadSizes.push_back(16);
2688     if (ST->is64Bit()) {
2689       Options.LoadSizes.push_back(8);
2690     }
2691     Options.LoadSizes.push_back(4);
2692     Options.LoadSizes.push_back(2);
2693     Options.LoadSizes.push_back(1);
2694     return Options;
2695   }();
2696   return IsZeroCmp ? &EqZeroOptions : &ThreeWayOptions;
2697 }
2698 
2699 bool X86TTIImpl::enableInterleavedAccessVectorization() {
2700   // TODO: We expect this to be beneficial regardless of arch,
2701   // but there are currently some unexplained performance artifacts on Atom.
2702   // As a temporary solution, disable on Atom.
2703   return !(ST->isAtom());
2704 }
2705 
2706 // Get estimation for interleaved load/store operations for AVX2.
2707 // \p Factor is the interleaved-access factor (stride) - number of
2708 // (interleaved) elements in the group.
2709 // \p Indices contains the indices for a strided load: when the
2710 // interleaved load has gaps they indicate which elements are used.
2711 // If Indices is empty (or if the number of indices is equal to the size
2712 // of the interleaved-access as given in \p Factor) the access has no gaps.
2713 //
2714 // As opposed to AVX-512, AVX2 does not have generic shuffles that allow
2715 // computing the cost using a generic formula as a function of generic
2716 // shuffles. We therefore use a lookup table instead, filled according to
2717 // the instruction sequences that codegen currently generates.
2718 int X86TTIImpl::getInterleavedMemoryOpCostAVX2(unsigned Opcode, Type *VecTy,
2719                                                unsigned Factor,
2720                                                ArrayRef<unsigned> Indices,
2721                                                unsigned Alignment,
2722                                                unsigned AddressSpace) {
2723 
2724   // We currently Support only fully-interleaved groups, with no gaps.
2725   // TODO: Support also strided loads (interleaved-groups with gaps).
2726   if (Indices.size() && Indices.size() != Factor)
2727     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2728                                              Alignment, AddressSpace);
2729 
2730   // VecTy for interleave memop is <VF*Factor x Elt>.
2731   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
2732   // VecTy = <12 x i32>.
2733   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
2734 
2735   // This function can be called with VecTy=<6xi128>, Factor=3, in which case
2736   // the VF=2, while v2i128 is an unsupported MVT vector type
2737   // (see MachineValueType.h::getVectorVT()).
2738   if (!LegalVT.isVector())
2739     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2740                                              Alignment, AddressSpace);
2741 
2742   unsigned VF = VecTy->getVectorNumElements() / Factor;
2743   Type *ScalarTy = VecTy->getVectorElementType();
2744 
2745   // Calculate the number of memory operations (NumOfMemOps), required
2746   // for load/store the VecTy.
2747   unsigned VecTySize = DL.getTypeStoreSize(VecTy);
2748   unsigned LegalVTSize = LegalVT.getStoreSize();
2749   unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize;
2750 
2751   // Get the cost of one memory operation.
2752   Type *SingleMemOpTy = VectorType::get(VecTy->getVectorElementType(),
2753                                         LegalVT.getVectorNumElements());
2754   unsigned MemOpCost =
2755       getMemoryOpCost(Opcode, SingleMemOpTy, Alignment, AddressSpace);
2756 
2757   VectorType *VT = VectorType::get(ScalarTy, VF);
2758   EVT ETy = TLI->getValueType(DL, VT);
2759   if (!ETy.isSimple())
2760     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2761                                              Alignment, AddressSpace);
2762 
2763   // TODO: Complete for other data-types and strides.
2764   // Each combination of Stride, ElementTy and VF results in a different
2765   // sequence; The cost tables are therefore accessed with:
2766   // Factor (stride) and VectorType=VFxElemType.
2767   // The Cost accounts only for the shuffle sequence;
2768   // The cost of the loads/stores is accounted for separately.
2769   //
2770   static const CostTblEntry AVX2InterleavedLoadTbl[] = {
2771     { 2, MVT::v4i64, 6 }, //(load 8i64 and) deinterleave into 2 x 4i64
2772     { 2, MVT::v4f64, 6 }, //(load 8f64 and) deinterleave into 2 x 4f64
2773 
2774     { 3, MVT::v2i8,  10 }, //(load 6i8 and)  deinterleave into 3 x 2i8
2775     { 3, MVT::v4i8,  4 },  //(load 12i8 and) deinterleave into 3 x 4i8
2776     { 3, MVT::v8i8,  9 },  //(load 24i8 and) deinterleave into 3 x 8i8
2777     { 3, MVT::v16i8, 11},  //(load 48i8 and) deinterleave into 3 x 16i8
2778     { 3, MVT::v32i8, 13},  //(load 96i8 and) deinterleave into 3 x 32i8
2779     { 3, MVT::v8f32, 17 }, //(load 24f32 and)deinterleave into 3 x 8f32
2780 
2781     { 4, MVT::v2i8,  12 }, //(load 8i8 and)   deinterleave into 4 x 2i8
2782     { 4, MVT::v4i8,  4 },  //(load 16i8 and)  deinterleave into 4 x 4i8
2783     { 4, MVT::v8i8,  20 }, //(load 32i8 and)  deinterleave into 4 x 8i8
2784     { 4, MVT::v16i8, 39 }, //(load 64i8 and)  deinterleave into 4 x 16i8
2785     { 4, MVT::v32i8, 80 }, //(load 128i8 and) deinterleave into 4 x 32i8
2786 
2787     { 8, MVT::v8f32, 40 }  //(load 64f32 and)deinterleave into 8 x 8f32
2788   };
2789 
2790   static const CostTblEntry AVX2InterleavedStoreTbl[] = {
2791     { 2, MVT::v4i64, 6 }, //interleave into 2 x 4i64 into 8i64 (and store)
2792     { 2, MVT::v4f64, 6 }, //interleave into 2 x 4f64 into 8f64 (and store)
2793 
2794     { 3, MVT::v2i8,  7 },  //interleave 3 x 2i8  into 6i8 (and store)
2795     { 3, MVT::v4i8,  8 },  //interleave 3 x 4i8  into 12i8 (and store)
2796     { 3, MVT::v8i8,  11 }, //interleave 3 x 8i8  into 24i8 (and store)
2797     { 3, MVT::v16i8, 11 }, //interleave 3 x 16i8 into 48i8 (and store)
2798     { 3, MVT::v32i8, 13 }, //interleave 3 x 32i8 into 96i8 (and store)
2799 
2800     { 4, MVT::v2i8,  12 }, //interleave 4 x 2i8  into 8i8 (and store)
2801     { 4, MVT::v4i8,  9 },  //interleave 4 x 4i8  into 16i8 (and store)
2802     { 4, MVT::v8i8,  10 }, //interleave 4 x 8i8  into 32i8 (and store)
2803     { 4, MVT::v16i8, 10 }, //interleave 4 x 16i8 into 64i8 (and store)
2804     { 4, MVT::v32i8, 12 }  //interleave 4 x 32i8 into 128i8 (and store)
2805   };
2806 
2807   if (Opcode == Instruction::Load) {
2808     if (const auto *Entry =
2809             CostTableLookup(AVX2InterleavedLoadTbl, Factor, ETy.getSimpleVT()))
2810       return NumOfMemOps * MemOpCost + Entry->Cost;
2811   } else {
2812     assert(Opcode == Instruction::Store &&
2813            "Expected Store Instruction at this  point");
2814     if (const auto *Entry =
2815             CostTableLookup(AVX2InterleavedStoreTbl, Factor, ETy.getSimpleVT()))
2816       return NumOfMemOps * MemOpCost + Entry->Cost;
2817   }
2818 
2819   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2820                                            Alignment, AddressSpace);
2821 }
2822 
2823 // Get estimation for interleaved load/store operations and strided load.
2824 // \p Indices contains indices for strided load.
2825 // \p Factor - the factor of interleaving.
2826 // AVX-512 provides 3-src shuffles that significantly reduces the cost.
2827 int X86TTIImpl::getInterleavedMemoryOpCostAVX512(unsigned Opcode, Type *VecTy,
2828                                                  unsigned Factor,
2829                                                  ArrayRef<unsigned> Indices,
2830                                                  unsigned Alignment,
2831                                                  unsigned AddressSpace) {
2832 
2833   // VecTy for interleave memop is <VF*Factor x Elt>.
2834   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
2835   // VecTy = <12 x i32>.
2836 
2837   // Calculate the number of memory operations (NumOfMemOps), required
2838   // for load/store the VecTy.
2839   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
2840   unsigned VecTySize = DL.getTypeStoreSize(VecTy);
2841   unsigned LegalVTSize = LegalVT.getStoreSize();
2842   unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize;
2843 
2844   // Get the cost of one memory operation.
2845   Type *SingleMemOpTy = VectorType::get(VecTy->getVectorElementType(),
2846                                         LegalVT.getVectorNumElements());
2847   unsigned MemOpCost =
2848       getMemoryOpCost(Opcode, SingleMemOpTy, Alignment, AddressSpace);
2849 
2850   unsigned VF = VecTy->getVectorNumElements() / Factor;
2851   MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF);
2852 
2853   if (Opcode == Instruction::Load) {
2854     // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl)
2855     // contain the cost of the optimized shuffle sequence that the
2856     // X86InterleavedAccess pass will generate.
2857     // The cost of loads and stores are computed separately from the table.
2858 
2859     // X86InterleavedAccess support only the following interleaved-access group.
2860     static const CostTblEntry AVX512InterleavedLoadTbl[] = {
2861         {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8
2862         {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8
2863         {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8
2864     };
2865 
2866     if (const auto *Entry =
2867             CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT))
2868       return NumOfMemOps * MemOpCost + Entry->Cost;
2869     //If an entry does not exist, fallback to the default implementation.
2870 
2871     // Kind of shuffle depends on number of loaded values.
2872     // If we load the entire data in one register, we can use a 1-src shuffle.
2873     // Otherwise, we'll merge 2 sources in each operation.
2874     TTI::ShuffleKind ShuffleKind =
2875         (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc;
2876 
2877     unsigned ShuffleCost =
2878         getShuffleCost(ShuffleKind, SingleMemOpTy, 0, nullptr);
2879 
2880     unsigned NumOfLoadsInInterleaveGrp =
2881         Indices.size() ? Indices.size() : Factor;
2882     Type *ResultTy = VectorType::get(VecTy->getVectorElementType(),
2883                                      VecTy->getVectorNumElements() / Factor);
2884     unsigned NumOfResults =
2885         getTLI()->getTypeLegalizationCost(DL, ResultTy).first *
2886         NumOfLoadsInInterleaveGrp;
2887 
2888     // About a half of the loads may be folded in shuffles when we have only
2889     // one result. If we have more than one result, we do not fold loads at all.
2890     unsigned NumOfUnfoldedLoads =
2891         NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2;
2892 
2893     // Get a number of shuffle operations per result.
2894     unsigned NumOfShufflesPerResult =
2895         std::max((unsigned)1, (unsigned)(NumOfMemOps - 1));
2896 
2897     // The SK_MergeTwoSrc shuffle clobbers one of src operands.
2898     // When we have more than one destination, we need additional instructions
2899     // to keep sources.
2900     unsigned NumOfMoves = 0;
2901     if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc)
2902       NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2;
2903 
2904     int Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost +
2905                NumOfUnfoldedLoads * MemOpCost + NumOfMoves;
2906 
2907     return Cost;
2908   }
2909 
2910   // Store.
2911   assert(Opcode == Instruction::Store &&
2912          "Expected Store Instruction at this  point");
2913   // X86InterleavedAccess support only the following interleaved-access group.
2914   static const CostTblEntry AVX512InterleavedStoreTbl[] = {
2915       {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store)
2916       {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store)
2917       {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store)
2918 
2919       {4, MVT::v8i8, 10},  // interleave 4 x 8i8  into 32i8  (and store)
2920       {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8  (and store)
2921       {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store)
2922       {4, MVT::v64i8, 24}  // interleave 4 x 32i8 into 256i8 (and store)
2923   };
2924 
2925   if (const auto *Entry =
2926           CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT))
2927     return NumOfMemOps * MemOpCost + Entry->Cost;
2928   //If an entry does not exist, fallback to the default implementation.
2929 
2930   // There is no strided stores meanwhile. And store can't be folded in
2931   // shuffle.
2932   unsigned NumOfSources = Factor; // The number of values to be merged.
2933   unsigned ShuffleCost =
2934       getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, 0, nullptr);
2935   unsigned NumOfShufflesPerStore = NumOfSources - 1;
2936 
2937   // The SK_MergeTwoSrc shuffle clobbers one of src operands.
2938   // We need additional instructions to keep sources.
2939   unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2;
2940   int Cost = NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) +
2941              NumOfMoves;
2942   return Cost;
2943 }
2944 
2945 int X86TTIImpl::getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
2946                                            unsigned Factor,
2947                                            ArrayRef<unsigned> Indices,
2948                                            unsigned Alignment,
2949                                            unsigned AddressSpace) {
2950   auto isSupportedOnAVX512 = [](Type *VecTy, bool HasBW) {
2951     Type *EltTy = VecTy->getVectorElementType();
2952     if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) ||
2953         EltTy->isIntegerTy(32) || EltTy->isPointerTy())
2954       return true;
2955     if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8))
2956       return HasBW;
2957     return false;
2958   };
2959   if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI()))
2960     return getInterleavedMemoryOpCostAVX512(Opcode, VecTy, Factor, Indices,
2961                                             Alignment, AddressSpace);
2962   if (ST->hasAVX2())
2963     return getInterleavedMemoryOpCostAVX2(Opcode, VecTy, Factor, Indices,
2964                                           Alignment, AddressSpace);
2965 
2966   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2967                                            Alignment, AddressSpace);
2968 }
2969