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   if (Vector) {
134     if (ST->hasAVX512())
135       return 512;
136     if (ST->hasAVX())
137       return 256;
138     if (ST->hasSSE1())
139       return 128;
140     return 0;
141   }
142 
143   if (ST->is64Bit())
144     return 64;
145 
146   return 32;
147 }
148 
149 unsigned X86TTIImpl::getLoadStoreVecRegBitWidth(unsigned) const {
150   return getRegisterBitWidth(true);
151 }
152 
153 unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) {
154   // If the loop will not be vectorized, don't interleave the loop.
155   // Let regular unroll to unroll the loop, which saves the overflow
156   // check and memory check cost.
157   if (VF == 1)
158     return 1;
159 
160   if (ST->isAtom())
161     return 1;
162 
163   // Sandybridge and Haswell have multiple execution ports and pipelined
164   // vector units.
165   if (ST->hasAVX())
166     return 4;
167 
168   return 2;
169 }
170 
171 int X86TTIImpl::getArithmeticInstrCost(
172     unsigned Opcode, Type *Ty,
173     TTI::OperandValueKind Op1Info, TTI::OperandValueKind Op2Info,
174     TTI::OperandValueProperties Opd1PropInfo,
175     TTI::OperandValueProperties Opd2PropInfo,
176     ArrayRef<const Value *> Args) {
177   // Legalize the type.
178   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
179 
180   int ISD = TLI->InstructionOpcodeToISD(Opcode);
181   assert(ISD && "Invalid opcode");
182 
183   static const CostTblEntry SLMCostTable[] = {
184     { ISD::MUL,  MVT::v4i32, 11 }, // pmulld
185     { ISD::MUL,  MVT::v8i16, 2  }, // pmullw
186     { ISD::MUL,  MVT::v16i8, 14 }, // extend/pmullw/trunc sequence.
187     { ISD::FMUL, MVT::f64,   2  }, // mulsd
188     { ISD::FMUL, MVT::v2f64, 4  }, // mulpd
189     { ISD::FMUL, MVT::v4f32, 2  }, // mulps
190     { ISD::FDIV, MVT::f32,   17 }, // divss
191     { ISD::FDIV, MVT::v4f32, 39 }, // divps
192     { ISD::FDIV, MVT::f64,   32 }, // divsd
193     { ISD::FDIV, MVT::v2f64, 69 }, // divpd
194     { ISD::FADD, MVT::v2f64, 2  }, // addpd
195     { ISD::FSUB, MVT::v2f64, 2  }, // subpd
196     // v2i64/v4i64 mul is custom lowered as a series of long:
197     // multiplies(3), shifts(3) and adds(2)
198     // slm muldq version throughput is 2 and addq throughput 4
199     // thus: 3X2 (muldq throughput) + 3X1 (shift throuput) +
200     //       3X4 (addq throughput) = 17
201     { ISD::MUL,  MVT::v2i64, 17 },
202     // slm addq\subq throughput is 4
203     { ISD::ADD,  MVT::v2i64, 4  },
204     { ISD::SUB,  MVT::v2i64, 4  },
205   };
206 
207   if (ST->isSLM()) {
208     if (Args.size() == 2 && ISD == ISD::MUL && LT.second == MVT::v4i32) {
209       // Check if the operands can be shrinked into a smaller datatype.
210       bool Op1Signed = false;
211       unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed);
212       bool Op2Signed = false;
213       unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed);
214 
215       bool signedMode = Op1Signed | Op2Signed;
216       unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize);
217 
218       if (OpMinSize <= 7)
219         return LT.first * 3; // pmullw/sext
220       if (!signedMode && OpMinSize <= 8)
221         return LT.first * 3; // pmullw/zext
222       if (OpMinSize <= 15)
223         return LT.first * 5; // pmullw/pmulhw/pshuf
224       if (!signedMode && OpMinSize <= 16)
225         return LT.first * 5; // pmullw/pmulhw/pshuf
226     }
227     if (const auto *Entry = CostTableLookup(SLMCostTable, ISD,
228                                             LT.second)) {
229       return LT.first * Entry->Cost;
230     }
231   }
232 
233   if (ISD == ISD::SDIV &&
234       Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
235       Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
236     // On X86, vector signed division by constants power-of-two are
237     // normally expanded to the sequence SRA + SRL + ADD + SRA.
238     // The OperandValue properties many not be same as that of previous
239     // operation;conservatively assume OP_None.
240     int Cost = 2 * getArithmeticInstrCost(Instruction::AShr, Ty, Op1Info,
241                                           Op2Info, TargetTransformInfo::OP_None,
242                                           TargetTransformInfo::OP_None);
243     Cost += getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info,
244                                    TargetTransformInfo::OP_None,
245                                    TargetTransformInfo::OP_None);
246     Cost += getArithmeticInstrCost(Instruction::Add, Ty, Op1Info, Op2Info,
247                                    TargetTransformInfo::OP_None,
248                                    TargetTransformInfo::OP_None);
249 
250     return Cost;
251   }
252 
253   static const CostTblEntry AVX512BWUniformConstCostTable[] = {
254     { ISD::SHL,  MVT::v64i8,   2 }, // psllw + pand.
255     { ISD::SRL,  MVT::v64i8,   2 }, // psrlw + pand.
256     { ISD::SRA,  MVT::v64i8,   4 }, // psrlw, pand, pxor, psubb.
257 
258     { ISD::SDIV, MVT::v32i16,  6 }, // vpmulhw sequence
259     { ISD::UDIV, MVT::v32i16,  6 }, // vpmulhuw sequence
260   };
261 
262   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
263       ST->hasBWI()) {
264     if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD,
265                                             LT.second))
266       return LT.first * Entry->Cost;
267   }
268 
269   static const CostTblEntry AVX512UniformConstCostTable[] = {
270     { ISD::SRA,  MVT::v2i64,   1 },
271     { ISD::SRA,  MVT::v4i64,   1 },
272     { ISD::SRA,  MVT::v8i64,   1 },
273 
274     { ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence
275     { ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence
276   };
277 
278   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
279       ST->hasAVX512()) {
280     if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD,
281                                             LT.second))
282       return LT.first * Entry->Cost;
283   }
284 
285   static const CostTblEntry AVX2UniformConstCostTable[] = {
286     { ISD::SHL,  MVT::v32i8,   2 }, // psllw + pand.
287     { ISD::SRL,  MVT::v32i8,   2 }, // psrlw + pand.
288     { ISD::SRA,  MVT::v32i8,   4 }, // psrlw, pand, pxor, psubb.
289 
290     { ISD::SRA,  MVT::v4i64,   4 }, // 2 x psrad + shuffle.
291 
292     { ISD::SDIV, MVT::v16i16,  6 }, // vpmulhw sequence
293     { ISD::UDIV, MVT::v16i16,  6 }, // vpmulhuw sequence
294     { ISD::SDIV, MVT::v8i32,  15 }, // vpmuldq sequence
295     { ISD::UDIV, MVT::v8i32,  15 }, // vpmuludq sequence
296   };
297 
298   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
299       ST->hasAVX2()) {
300     if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD,
301                                             LT.second))
302       return LT.first * Entry->Cost;
303   }
304 
305   static const CostTblEntry SSE2UniformConstCostTable[] = {
306     { ISD::SHL,  MVT::v16i8,     2 }, // psllw + pand.
307     { ISD::SRL,  MVT::v16i8,     2 }, // psrlw + pand.
308     { ISD::SRA,  MVT::v16i8,     4 }, // psrlw, pand, pxor, psubb.
309 
310     { ISD::SHL,  MVT::v32i8,   4+2 }, // 2*(psllw + pand) + split.
311     { ISD::SRL,  MVT::v32i8,   4+2 }, // 2*(psrlw + pand) + split.
312     { ISD::SRA,  MVT::v32i8,   8+2 }, // 2*(psrlw, pand, pxor, psubb) + split.
313 
314     { ISD::SDIV, MVT::v16i16, 12+2 }, // 2*pmulhw sequence + split.
315     { ISD::SDIV, MVT::v8i16,     6 }, // pmulhw sequence
316     { ISD::UDIV, MVT::v16i16, 12+2 }, // 2*pmulhuw sequence + split.
317     { ISD::UDIV, MVT::v8i16,     6 }, // pmulhuw sequence
318     { ISD::SDIV, MVT::v8i32,  38+2 }, // 2*pmuludq sequence + split.
319     { ISD::SDIV, MVT::v4i32,    19 }, // pmuludq sequence
320     { ISD::UDIV, MVT::v8i32,  30+2 }, // 2*pmuludq sequence + split.
321     { ISD::UDIV, MVT::v4i32,    15 }, // pmuludq sequence
322   };
323 
324   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
325       ST->hasSSE2()) {
326     // pmuldq sequence.
327     if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX())
328       return LT.first * 32;
329     if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
330       return LT.first * 15;
331 
332     // XOP has faster vXi8 shifts.
333     if ((ISD != ISD::SHL && ISD != ISD::SRL && ISD != ISD::SRA) ||
334         !ST->hasXOP())
335       if (const auto *Entry =
336               CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second))
337         return LT.first * Entry->Cost;
338   }
339 
340   static const CostTblEntry AVX2UniformCostTable[] = {
341     // Uniform splats are cheaper for the following instructions.
342     { ISD::SHL,  MVT::v16i16, 1 }, // psllw.
343     { ISD::SRL,  MVT::v16i16, 1 }, // psrlw.
344     { ISD::SRA,  MVT::v16i16, 1 }, // psraw.
345   };
346 
347   if (ST->hasAVX2() &&
348       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
349        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
350     if (const auto *Entry =
351             CostTableLookup(AVX2UniformCostTable, ISD, LT.second))
352       return LT.first * Entry->Cost;
353   }
354 
355   static const CostTblEntry SSE2UniformCostTable[] = {
356     // Uniform splats are cheaper for the following instructions.
357     { ISD::SHL,  MVT::v8i16,  1 }, // psllw.
358     { ISD::SHL,  MVT::v4i32,  1 }, // pslld
359     { ISD::SHL,  MVT::v2i64,  1 }, // psllq.
360 
361     { ISD::SRL,  MVT::v8i16,  1 }, // psrlw.
362     { ISD::SRL,  MVT::v4i32,  1 }, // psrld.
363     { ISD::SRL,  MVT::v2i64,  1 }, // psrlq.
364 
365     { ISD::SRA,  MVT::v8i16,  1 }, // psraw.
366     { ISD::SRA,  MVT::v4i32,  1 }, // psrad.
367   };
368 
369   if (ST->hasSSE2() &&
370       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
371        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
372     if (const auto *Entry =
373             CostTableLookup(SSE2UniformCostTable, ISD, LT.second))
374       return LT.first * Entry->Cost;
375   }
376 
377   static const CostTblEntry AVX512DQCostTable[] = {
378     { ISD::MUL,  MVT::v2i64, 1 },
379     { ISD::MUL,  MVT::v4i64, 1 },
380     { ISD::MUL,  MVT::v8i64, 1 }
381   };
382 
383   // Look for AVX512DQ lowering tricks for custom cases.
384   if (ST->hasDQI())
385     if (const auto *Entry = CostTableLookup(AVX512DQCostTable, ISD, LT.second))
386       return LT.first * Entry->Cost;
387 
388   static const CostTblEntry AVX512BWCostTable[] = {
389     { ISD::SHL,   MVT::v8i16,      1 }, // vpsllvw
390     { ISD::SRL,   MVT::v8i16,      1 }, // vpsrlvw
391     { ISD::SRA,   MVT::v8i16,      1 }, // vpsravw
392 
393     { ISD::SHL,   MVT::v16i16,     1 }, // vpsllvw
394     { ISD::SRL,   MVT::v16i16,     1 }, // vpsrlvw
395     { ISD::SRA,   MVT::v16i16,     1 }, // vpsravw
396 
397     { ISD::SHL,   MVT::v32i16,     1 }, // vpsllvw
398     { ISD::SRL,   MVT::v32i16,     1 }, // vpsrlvw
399     { ISD::SRA,   MVT::v32i16,     1 }, // vpsravw
400 
401     { ISD::SHL,   MVT::v64i8,     11 }, // vpblendvb sequence.
402     { ISD::SRL,   MVT::v64i8,     11 }, // vpblendvb sequence.
403     { ISD::SRA,   MVT::v64i8,     24 }, // vpblendvb sequence.
404 
405     { ISD::MUL,   MVT::v64i8,     11 }, // extend/pmullw/trunc sequence.
406     { ISD::MUL,   MVT::v32i8,      4 }, // extend/pmullw/trunc sequence.
407     { ISD::MUL,   MVT::v16i8,      4 }, // extend/pmullw/trunc sequence.
408 
409     // Vectorizing division is a bad idea. See the SSE2 table for more comments.
410     { ISD::SDIV,  MVT::v64i8,  64*20 },
411     { ISD::SDIV,  MVT::v32i16, 32*20 },
412     { ISD::UDIV,  MVT::v64i8,  64*20 },
413     { ISD::UDIV,  MVT::v32i16, 32*20 }
414   };
415 
416   // Look for AVX512BW lowering tricks for custom cases.
417   if (ST->hasBWI())
418     if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD, LT.second))
419       return LT.first * Entry->Cost;
420 
421   static const CostTblEntry AVX512CostTable[] = {
422     { ISD::SHL,     MVT::v16i32,     1 },
423     { ISD::SRL,     MVT::v16i32,     1 },
424     { ISD::SRA,     MVT::v16i32,     1 },
425 
426     { ISD::SHL,     MVT::v8i64,      1 },
427     { ISD::SRL,     MVT::v8i64,      1 },
428 
429     { ISD::SRA,     MVT::v2i64,      1 },
430     { ISD::SRA,     MVT::v4i64,      1 },
431     { ISD::SRA,     MVT::v8i64,      1 },
432 
433     { ISD::MUL,     MVT::v32i8,     13 }, // extend/pmullw/trunc sequence.
434     { ISD::MUL,     MVT::v16i8,      5 }, // extend/pmullw/trunc sequence.
435     { ISD::MUL,     MVT::v16i32,     1 }, // pmulld
436     { ISD::MUL,     MVT::v8i64,      8 }, // 3*pmuludq/3*shift/2*add
437 
438     // Vectorizing division is a bad idea. See the SSE2 table for more comments.
439     { ISD::SDIV,    MVT::v16i32, 16*20 },
440     { ISD::SDIV,    MVT::v8i64,   8*20 },
441     { ISD::UDIV,    MVT::v16i32, 16*20 },
442     { ISD::UDIV,    MVT::v8i64,   8*20 }
443   };
444 
445   if (ST->hasAVX512())
446     if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second))
447       return LT.first * Entry->Cost;
448 
449   static const CostTblEntry AVX2ShiftCostTable[] = {
450     // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
451     // customize them to detect the cases where shift amount is a scalar one.
452     { ISD::SHL,     MVT::v4i32,    1 },
453     { ISD::SRL,     MVT::v4i32,    1 },
454     { ISD::SRA,     MVT::v4i32,    1 },
455     { ISD::SHL,     MVT::v8i32,    1 },
456     { ISD::SRL,     MVT::v8i32,    1 },
457     { ISD::SRA,     MVT::v8i32,    1 },
458     { ISD::SHL,     MVT::v2i64,    1 },
459     { ISD::SRL,     MVT::v2i64,    1 },
460     { ISD::SHL,     MVT::v4i64,    1 },
461     { ISD::SRL,     MVT::v4i64,    1 },
462   };
463 
464   // Look for AVX2 lowering tricks.
465   if (ST->hasAVX2()) {
466     if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
467         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
468          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
469       // On AVX2, a packed v16i16 shift left by a constant build_vector
470       // is lowered into a vector multiply (vpmullw).
471       return LT.first;
472 
473     if (const auto *Entry = CostTableLookup(AVX2ShiftCostTable, ISD, LT.second))
474       return LT.first * Entry->Cost;
475   }
476 
477   static const CostTblEntry XOPShiftCostTable[] = {
478     // 128bit shifts take 1cy, but right shifts require negation beforehand.
479     { ISD::SHL,     MVT::v16i8,    1 },
480     { ISD::SRL,     MVT::v16i8,    2 },
481     { ISD::SRA,     MVT::v16i8,    2 },
482     { ISD::SHL,     MVT::v8i16,    1 },
483     { ISD::SRL,     MVT::v8i16,    2 },
484     { ISD::SRA,     MVT::v8i16,    2 },
485     { ISD::SHL,     MVT::v4i32,    1 },
486     { ISD::SRL,     MVT::v4i32,    2 },
487     { ISD::SRA,     MVT::v4i32,    2 },
488     { ISD::SHL,     MVT::v2i64,    1 },
489     { ISD::SRL,     MVT::v2i64,    2 },
490     { ISD::SRA,     MVT::v2i64,    2 },
491     // 256bit shifts require splitting if AVX2 didn't catch them above.
492     { ISD::SHL,     MVT::v32i8,  2+2 },
493     { ISD::SRL,     MVT::v32i8,  4+2 },
494     { ISD::SRA,     MVT::v32i8,  4+2 },
495     { ISD::SHL,     MVT::v16i16, 2+2 },
496     { ISD::SRL,     MVT::v16i16, 4+2 },
497     { ISD::SRA,     MVT::v16i16, 4+2 },
498     { ISD::SHL,     MVT::v8i32,  2+2 },
499     { ISD::SRL,     MVT::v8i32,  4+2 },
500     { ISD::SRA,     MVT::v8i32,  4+2 },
501     { ISD::SHL,     MVT::v4i64,  2+2 },
502     { ISD::SRL,     MVT::v4i64,  4+2 },
503     { ISD::SRA,     MVT::v4i64,  4+2 },
504   };
505 
506   // Look for XOP lowering tricks.
507   if (ST->hasXOP())
508     if (const auto *Entry = CostTableLookup(XOPShiftCostTable, ISD, LT.second))
509       return LT.first * Entry->Cost;
510 
511   static const CostTblEntry SSE2UniformShiftCostTable[] = {
512     // Uniform splats are cheaper for the following instructions.
513     { ISD::SHL,  MVT::v16i16, 2+2 }, // 2*psllw + split.
514     { ISD::SHL,  MVT::v8i32,  2+2 }, // 2*pslld + split.
515     { ISD::SHL,  MVT::v4i64,  2+2 }, // 2*psllq + split.
516 
517     { ISD::SRL,  MVT::v16i16, 2+2 }, // 2*psrlw + split.
518     { ISD::SRL,  MVT::v8i32,  2+2 }, // 2*psrld + split.
519     { ISD::SRL,  MVT::v4i64,  2+2 }, // 2*psrlq + split.
520 
521     { ISD::SRA,  MVT::v16i16, 2+2 }, // 2*psraw + split.
522     { ISD::SRA,  MVT::v8i32,  2+2 }, // 2*psrad + split.
523     { ISD::SRA,  MVT::v2i64,    4 }, // 2*psrad + shuffle.
524     { ISD::SRA,  MVT::v4i64,  8+2 }, // 2*(2*psrad + shuffle) + split.
525   };
526 
527   if (ST->hasSSE2() &&
528       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
529        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
530 
531     // Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table.
532     if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2())
533       return LT.first * 4; // 2*psrad + shuffle.
534 
535     if (const auto *Entry =
536             CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second))
537       return LT.first * Entry->Cost;
538   }
539 
540   if (ISD == ISD::SHL &&
541       Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
542     MVT VT = LT.second;
543     // Vector shift left by non uniform constant can be lowered
544     // into vector multiply.
545     if (((VT == MVT::v8i16 || VT == MVT::v4i32) && ST->hasSSE2()) ||
546         ((VT == MVT::v16i16 || VT == MVT::v8i32) && ST->hasAVX()))
547       ISD = ISD::MUL;
548   }
549 
550   static const CostTblEntry AVX2CostTable[] = {
551     { ISD::SHL,  MVT::v32i8,     11 }, // vpblendvb sequence.
552     { ISD::SHL,  MVT::v16i16,    10 }, // extend/vpsrlvd/pack sequence.
553 
554     { ISD::SRL,  MVT::v32i8,     11 }, // vpblendvb sequence.
555     { ISD::SRL,  MVT::v16i16,    10 }, // extend/vpsrlvd/pack sequence.
556 
557     { ISD::SRA,  MVT::v32i8,     24 }, // vpblendvb sequence.
558     { ISD::SRA,  MVT::v16i16,    10 }, // extend/vpsravd/pack sequence.
559     { ISD::SRA,  MVT::v2i64,      4 }, // srl/xor/sub sequence.
560     { ISD::SRA,  MVT::v4i64,      4 }, // srl/xor/sub sequence.
561 
562     { ISD::SUB,  MVT::v32i8,      1 }, // psubb
563     { ISD::ADD,  MVT::v32i8,      1 }, // paddb
564     { ISD::SUB,  MVT::v16i16,     1 }, // psubw
565     { ISD::ADD,  MVT::v16i16,     1 }, // paddw
566     { ISD::SUB,  MVT::v8i32,      1 }, // psubd
567     { ISD::ADD,  MVT::v8i32,      1 }, // paddd
568     { ISD::SUB,  MVT::v4i64,      1 }, // psubq
569     { ISD::ADD,  MVT::v4i64,      1 }, // paddq
570 
571     { ISD::MUL,  MVT::v32i8,     17 }, // extend/pmullw/trunc sequence.
572     { ISD::MUL,  MVT::v16i8,      7 }, // extend/pmullw/trunc sequence.
573     { ISD::MUL,  MVT::v16i16,     1 }, // pmullw
574     { ISD::MUL,  MVT::v8i32,      1 }, // pmulld
575     { ISD::MUL,  MVT::v4i64,      8 }, // 3*pmuludq/3*shift/2*add
576 
577     { ISD::FDIV, MVT::f32,        7 }, // Haswell from http://www.agner.org/
578     { ISD::FDIV, MVT::v4f32,      7 }, // Haswell from http://www.agner.org/
579     { ISD::FDIV, MVT::v8f32,     14 }, // Haswell from http://www.agner.org/
580     { ISD::FDIV, MVT::f64,       14 }, // Haswell from http://www.agner.org/
581     { ISD::FDIV, MVT::v2f64,     14 }, // Haswell from http://www.agner.org/
582     { ISD::FDIV, MVT::v4f64,     28 }, // Haswell from http://www.agner.org/
583   };
584 
585   // Look for AVX2 lowering tricks for custom cases.
586   if (ST->hasAVX2())
587     if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second))
588       return LT.first * Entry->Cost;
589 
590   static const CostTblEntry AVX1CostTable[] = {
591     // We don't have to scalarize unsupported ops. We can issue two half-sized
592     // operations and we only need to extract the upper YMM half.
593     // Two ops + 1 extract + 1 insert = 4.
594     { ISD::MUL,     MVT::v16i16,     4 },
595     { ISD::MUL,     MVT::v8i32,      4 },
596     { ISD::SUB,     MVT::v32i8,      4 },
597     { ISD::ADD,     MVT::v32i8,      4 },
598     { ISD::SUB,     MVT::v16i16,     4 },
599     { ISD::ADD,     MVT::v16i16,     4 },
600     { ISD::SUB,     MVT::v8i32,      4 },
601     { ISD::ADD,     MVT::v8i32,      4 },
602     { ISD::SUB,     MVT::v4i64,      4 },
603     { ISD::ADD,     MVT::v4i64,      4 },
604 
605     // A v4i64 multiply is custom lowered as two split v2i64 vectors that then
606     // are lowered as a series of long multiplies(3), shifts(3) and adds(2)
607     // Because we believe v4i64 to be a legal type, we must also include the
608     // extract+insert in the cost table. Therefore, the cost here is 18
609     // instead of 8.
610     { ISD::MUL,     MVT::v4i64,     18 },
611 
612     { ISD::MUL,     MVT::v32i8,     26 }, // extend/pmullw/trunc sequence.
613 
614     { ISD::FDIV,    MVT::f32,       14 }, // SNB from http://www.agner.org/
615     { ISD::FDIV,    MVT::v4f32,     14 }, // SNB from http://www.agner.org/
616     { ISD::FDIV,    MVT::v8f32,     28 }, // SNB from http://www.agner.org/
617     { ISD::FDIV,    MVT::f64,       22 }, // SNB from http://www.agner.org/
618     { ISD::FDIV,    MVT::v2f64,     22 }, // SNB from http://www.agner.org/
619     { ISD::FDIV,    MVT::v4f64,     44 }, // SNB from http://www.agner.org/
620 
621     // Vectorizing division is a bad idea. See the SSE2 table for more comments.
622     { ISD::SDIV,    MVT::v32i8,  32*20 },
623     { ISD::SDIV,    MVT::v16i16, 16*20 },
624     { ISD::SDIV,    MVT::v8i32,   8*20 },
625     { ISD::SDIV,    MVT::v4i64,   4*20 },
626     { ISD::UDIV,    MVT::v32i8,  32*20 },
627     { ISD::UDIV,    MVT::v16i16, 16*20 },
628     { ISD::UDIV,    MVT::v8i32,   8*20 },
629     { ISD::UDIV,    MVT::v4i64,   4*20 },
630   };
631 
632   if (ST->hasAVX())
633     if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, LT.second))
634       return LT.first * Entry->Cost;
635 
636   static const CostTblEntry SSE42CostTable[] = {
637     { ISD::FDIV,  MVT::f32,   14 }, // Nehalem from http://www.agner.org/
638     { ISD::FDIV,  MVT::v4f32, 14 }, // Nehalem from http://www.agner.org/
639     { ISD::FDIV,  MVT::f64,   22 }, // Nehalem from http://www.agner.org/
640     { ISD::FDIV,  MVT::v2f64, 22 }, // Nehalem from http://www.agner.org/
641   };
642 
643   if (ST->hasSSE42())
644     if (const auto *Entry = CostTableLookup(SSE42CostTable, ISD, LT.second))
645       return LT.first * Entry->Cost;
646 
647   static const CostTblEntry SSE41CostTable[] = {
648     { ISD::SHL,  MVT::v16i8,      11 }, // pblendvb sequence.
649     { ISD::SHL,  MVT::v32i8,  2*11+2 }, // pblendvb sequence + split.
650     { ISD::SHL,  MVT::v8i16,      14 }, // pblendvb sequence.
651     { ISD::SHL,  MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
652     { ISD::SHL,  MVT::v4i32,       4 }, // pslld/paddd/cvttps2dq/pmulld
653     { ISD::SHL,  MVT::v8i32,   2*4+2 }, // pslld/paddd/cvttps2dq/pmulld + split
654 
655     { ISD::SRL,  MVT::v16i8,      12 }, // pblendvb sequence.
656     { ISD::SRL,  MVT::v32i8,  2*12+2 }, // pblendvb sequence + split.
657     { ISD::SRL,  MVT::v8i16,      14 }, // pblendvb sequence.
658     { ISD::SRL,  MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
659     { ISD::SRL,  MVT::v4i32,      11 }, // Shift each lane + blend.
660     { ISD::SRL,  MVT::v8i32,  2*11+2 }, // Shift each lane + blend + split.
661 
662     { ISD::SRA,  MVT::v16i8,      24 }, // pblendvb sequence.
663     { ISD::SRA,  MVT::v32i8,  2*24+2 }, // pblendvb sequence + split.
664     { ISD::SRA,  MVT::v8i16,      14 }, // pblendvb sequence.
665     { ISD::SRA,  MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
666     { ISD::SRA,  MVT::v4i32,      12 }, // Shift each lane + blend.
667     { ISD::SRA,  MVT::v8i32,  2*12+2 }, // Shift each lane + blend + split.
668 
669     { ISD::MUL,  MVT::v4i32,       1 }  // pmulld
670   };
671 
672   if (ST->hasSSE41())
673     if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second))
674       return LT.first * Entry->Cost;
675 
676   static const CostTblEntry SSE2CostTable[] = {
677     // We don't correctly identify costs of casts because they are marked as
678     // custom.
679     { ISD::SHL,  MVT::v16i8,      26 }, // cmpgtb sequence.
680     { ISD::SHL,  MVT::v8i16,      32 }, // cmpgtb sequence.
681     { ISD::SHL,  MVT::v4i32,     2*5 }, // We optimized this using mul.
682     { ISD::SHL,  MVT::v2i64,       4 }, // splat+shuffle sequence.
683     { ISD::SHL,  MVT::v4i64,   2*4+2 }, // splat+shuffle sequence + split.
684 
685     { ISD::SRL,  MVT::v16i8,      26 }, // cmpgtb sequence.
686     { ISD::SRL,  MVT::v8i16,      32 }, // cmpgtb sequence.
687     { ISD::SRL,  MVT::v4i32,      16 }, // Shift each lane + blend.
688     { ISD::SRL,  MVT::v2i64,       4 }, // splat+shuffle sequence.
689     { ISD::SRL,  MVT::v4i64,   2*4+2 }, // splat+shuffle sequence + split.
690 
691     { ISD::SRA,  MVT::v16i8,      54 }, // unpacked cmpgtb sequence.
692     { ISD::SRA,  MVT::v8i16,      32 }, // cmpgtb sequence.
693     { ISD::SRA,  MVT::v4i32,      16 }, // Shift each lane + blend.
694     { ISD::SRA,  MVT::v2i64,      12 }, // srl/xor/sub sequence.
695     { ISD::SRA,  MVT::v4i64,  2*12+2 }, // srl/xor/sub sequence+split.
696 
697     { ISD::MUL,  MVT::v16i8,      12 }, // extend/pmullw/trunc sequence.
698     { ISD::MUL,  MVT::v8i16,       1 }, // pmullw
699     { ISD::MUL,  MVT::v4i32,       6 }, // 3*pmuludq/4*shuffle
700     { ISD::MUL,  MVT::v2i64,       8 }, // 3*pmuludq/3*shift/2*add
701 
702     { ISD::FDIV, MVT::f32,        23 }, // Pentium IV from http://www.agner.org/
703     { ISD::FDIV, MVT::v4f32,      39 }, // Pentium IV from http://www.agner.org/
704     { ISD::FDIV, MVT::f64,        38 }, // Pentium IV from http://www.agner.org/
705     { ISD::FDIV, MVT::v2f64,      69 }, // Pentium IV from http://www.agner.org/
706 
707     // It is not a good idea to vectorize division. We have to scalarize it and
708     // in the process we will often end up having to spilling regular
709     // registers. The overhead of division is going to dominate most kernels
710     // anyways so try hard to prevent vectorization of division - it is
711     // generally a bad idea. Assume somewhat arbitrarily that we have to be able
712     // to hide "20 cycles" for each lane.
713     { ISD::SDIV,  MVT::v16i8,  16*20 },
714     { ISD::SDIV,  MVT::v8i16,   8*20 },
715     { ISD::SDIV,  MVT::v4i32,   4*20 },
716     { ISD::SDIV,  MVT::v2i64,   2*20 },
717     { ISD::UDIV,  MVT::v16i8,  16*20 },
718     { ISD::UDIV,  MVT::v8i16,   8*20 },
719     { ISD::UDIV,  MVT::v4i32,   4*20 },
720     { ISD::UDIV,  MVT::v2i64,   2*20 },
721   };
722 
723   if (ST->hasSSE2())
724     if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second))
725       return LT.first * Entry->Cost;
726 
727   static const CostTblEntry SSE1CostTable[] = {
728     { ISD::FDIV, MVT::f32,   17 }, // Pentium III from http://www.agner.org/
729     { ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/
730   };
731 
732   if (ST->hasSSE1())
733     if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second))
734       return LT.first * Entry->Cost;
735 
736   // Fallback to the default implementation.
737   return BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
738 }
739 
740 int X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
741                                Type *SubTp) {
742   // 64-bit packed float vectors (v2f32) are widened to type v4f32.
743   // 64-bit packed integer vectors (v2i32) are promoted to type v2i64.
744   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
745 
746   // For Broadcasts we are splatting the first element from the first input
747   // register, so only need to reference that input and all the output
748   // registers are the same.
749   if (Kind == TTI::SK_Broadcast)
750     LT.first = 1;
751 
752   // We are going to permute multiple sources and the result will be in multiple
753   // destinations. Providing an accurate cost only for splits where the element
754   // type remains the same.
755   if (Kind == TTI::SK_PermuteSingleSrc && LT.first != 1) {
756     MVT LegalVT = LT.second;
757     if (LegalVT.getVectorElementType().getSizeInBits() ==
758             Tp->getVectorElementType()->getPrimitiveSizeInBits() &&
759         LegalVT.getVectorNumElements() < Tp->getVectorNumElements()) {
760 
761       unsigned VecTySize = DL.getTypeStoreSize(Tp);
762       unsigned LegalVTSize = LegalVT.getStoreSize();
763       // Number of source vectors after legalization:
764       unsigned NumOfSrcs = (VecTySize + LegalVTSize - 1) / LegalVTSize;
765       // Number of destination vectors after legalization:
766       unsigned NumOfDests = LT.first;
767 
768       Type *SingleOpTy = VectorType::get(Tp->getVectorElementType(),
769                                          LegalVT.getVectorNumElements());
770 
771       unsigned NumOfShuffles = (NumOfSrcs - 1) * NumOfDests;
772       return NumOfShuffles *
773              getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy, 0, nullptr);
774     }
775 
776     return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
777   }
778 
779   // For 2-input shuffles, we must account for splitting the 2 inputs into many.
780   if (Kind == TTI::SK_PermuteTwoSrc && LT.first != 1) {
781     // We assume that source and destination have the same vector type.
782     int NumOfDests = LT.first;
783     int NumOfShufflesPerDest = LT.first * 2 - 1;
784     LT.first = NumOfDests * NumOfShufflesPerDest;
785   }
786 
787   static const CostTblEntry AVX512VBMIShuffleTbl[] = {
788     { TTI::SK_Reverse,          MVT::v64i8,  1 }, // vpermb
789     { TTI::SK_Reverse,          MVT::v32i8,  1 }, // vpermb
790 
791     { TTI::SK_PermuteSingleSrc, MVT::v64i8,  1 }, // vpermb
792     { TTI::SK_PermuteSingleSrc, MVT::v32i8,  1 }, // vpermb
793 
794     { TTI::SK_PermuteTwoSrc,    MVT::v64i8,  1 }, // vpermt2b
795     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,  1 }, // vpermt2b
796     { TTI::SK_PermuteTwoSrc,    MVT::v16i8,  1 }  // vpermt2b
797   };
798 
799   if (ST->hasVBMI())
800     if (const auto *Entry =
801             CostTableLookup(AVX512VBMIShuffleTbl, Kind, LT.second))
802       return LT.first * Entry->Cost;
803 
804   static const CostTblEntry AVX512BWShuffleTbl[] = {
805     { TTI::SK_Broadcast,        MVT::v32i16, 1 }, // vpbroadcastw
806     { TTI::SK_Broadcast,        MVT::v64i8,  1 }, // vpbroadcastb
807 
808     { TTI::SK_Reverse,          MVT::v32i16, 1 }, // vpermw
809     { TTI::SK_Reverse,          MVT::v16i16, 1 }, // vpermw
810     { TTI::SK_Reverse,          MVT::v64i8,  2 }, // pshufb + vshufi64x2
811 
812     { TTI::SK_PermuteSingleSrc, MVT::v32i16, 1 }, // vpermw
813     { TTI::SK_PermuteSingleSrc, MVT::v16i16, 1 }, // vpermw
814     { TTI::SK_PermuteSingleSrc, MVT::v8i16,  1 }, // vpermw
815     { TTI::SK_PermuteSingleSrc, MVT::v64i8,  8 }, // extend to v32i16
816     { TTI::SK_PermuteSingleSrc, MVT::v32i8,  3 }, // vpermw + zext/trunc
817 
818     { TTI::SK_PermuteTwoSrc,    MVT::v32i16, 1 }, // vpermt2w
819     { TTI::SK_PermuteTwoSrc,    MVT::v16i16, 1 }, // vpermt2w
820     { TTI::SK_PermuteTwoSrc,    MVT::v8i16,  1 }, // vpermt2w
821     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,  3 }, // zext + vpermt2w + trunc
822     { TTI::SK_PermuteTwoSrc,    MVT::v64i8, 19 }, // 6 * v32i8 + 1
823     { TTI::SK_PermuteTwoSrc,    MVT::v16i8,  3 }  // zext + vpermt2w + trunc
824   };
825 
826   if (ST->hasBWI())
827     if (const auto *Entry =
828             CostTableLookup(AVX512BWShuffleTbl, Kind, LT.second))
829       return LT.first * Entry->Cost;
830 
831   static const CostTblEntry AVX512ShuffleTbl[] = {
832     { TTI::SK_Broadcast,        MVT::v8f64,  1 }, // vbroadcastpd
833     { TTI::SK_Broadcast,        MVT::v16f32, 1 }, // vbroadcastps
834     { TTI::SK_Broadcast,        MVT::v8i64,  1 }, // vpbroadcastq
835     { TTI::SK_Broadcast,        MVT::v16i32, 1 }, // vpbroadcastd
836 
837     { TTI::SK_Reverse,          MVT::v8f64,  1 }, // vpermpd
838     { TTI::SK_Reverse,          MVT::v16f32, 1 }, // vpermps
839     { TTI::SK_Reverse,          MVT::v8i64,  1 }, // vpermq
840     { TTI::SK_Reverse,          MVT::v16i32, 1 }, // vpermd
841 
842     { TTI::SK_PermuteSingleSrc, MVT::v8f64,  1 }, // vpermpd
843     { TTI::SK_PermuteSingleSrc, MVT::v4f64,  1 }, // vpermpd
844     { TTI::SK_PermuteSingleSrc, MVT::v2f64,  1 }, // vpermpd
845     { TTI::SK_PermuteSingleSrc, MVT::v16f32, 1 }, // vpermps
846     { TTI::SK_PermuteSingleSrc, MVT::v8f32,  1 }, // vpermps
847     { TTI::SK_PermuteSingleSrc, MVT::v4f32,  1 }, // vpermps
848     { TTI::SK_PermuteSingleSrc, MVT::v8i64,  1 }, // vpermq
849     { TTI::SK_PermuteSingleSrc, MVT::v4i64,  1 }, // vpermq
850     { TTI::SK_PermuteSingleSrc, MVT::v2i64,  1 }, // vpermq
851     { TTI::SK_PermuteSingleSrc, MVT::v16i32, 1 }, // vpermd
852     { TTI::SK_PermuteSingleSrc, MVT::v8i32,  1 }, // vpermd
853     { TTI::SK_PermuteSingleSrc, MVT::v4i32,  1 }, // vpermd
854     { TTI::SK_PermuteSingleSrc, MVT::v16i8,  1 }, // pshufb
855 
856     { TTI::SK_PermuteTwoSrc,    MVT::v8f64,  1 }, // vpermt2pd
857     { TTI::SK_PermuteTwoSrc,    MVT::v16f32, 1 }, // vpermt2ps
858     { TTI::SK_PermuteTwoSrc,    MVT::v8i64,  1 }, // vpermt2q
859     { TTI::SK_PermuteTwoSrc,    MVT::v16i32, 1 }, // vpermt2d
860     { TTI::SK_PermuteTwoSrc,    MVT::v4f64,  1 }, // vpermt2pd
861     { TTI::SK_PermuteTwoSrc,    MVT::v8f32,  1 }, // vpermt2ps
862     { TTI::SK_PermuteTwoSrc,    MVT::v4i64,  1 }, // vpermt2q
863     { TTI::SK_PermuteTwoSrc,    MVT::v8i32,  1 }, // vpermt2d
864     { TTI::SK_PermuteTwoSrc,    MVT::v2f64,  1 }, // vpermt2pd
865     { TTI::SK_PermuteTwoSrc,    MVT::v4f32,  1 }, // vpermt2ps
866     { TTI::SK_PermuteTwoSrc,    MVT::v2i64,  1 }, // vpermt2q
867     { TTI::SK_PermuteTwoSrc,    MVT::v4i32,  1 }  // vpermt2d
868   };
869 
870   if (ST->hasAVX512())
871     if (const auto *Entry = CostTableLookup(AVX512ShuffleTbl, Kind, LT.second))
872       return LT.first * Entry->Cost;
873 
874   static const CostTblEntry AVX2ShuffleTbl[] = {
875     { TTI::SK_Broadcast, MVT::v4f64,  1 }, // vbroadcastpd
876     { TTI::SK_Broadcast, MVT::v8f32,  1 }, // vbroadcastps
877     { TTI::SK_Broadcast, MVT::v4i64,  1 }, // vpbroadcastq
878     { TTI::SK_Broadcast, MVT::v8i32,  1 }, // vpbroadcastd
879     { TTI::SK_Broadcast, MVT::v16i16, 1 }, // vpbroadcastw
880     { TTI::SK_Broadcast, MVT::v32i8,  1 }, // vpbroadcastb
881 
882     { TTI::SK_Reverse,   MVT::v4f64,  1 }, // vpermpd
883     { TTI::SK_Reverse,   MVT::v8f32,  1 }, // vpermps
884     { TTI::SK_Reverse,   MVT::v4i64,  1 }, // vpermq
885     { TTI::SK_Reverse,   MVT::v8i32,  1 }, // vpermd
886     { TTI::SK_Reverse,   MVT::v16i16, 2 }, // vperm2i128 + pshufb
887     { TTI::SK_Reverse,   MVT::v32i8,  2 }, // vperm2i128 + pshufb
888 
889     { TTI::SK_Alternate, MVT::v16i16, 1 }, // vpblendw
890     { TTI::SK_Alternate, MVT::v32i8,  1 }, // vpblendvb
891 
892     { TTI::SK_PermuteSingleSrc, MVT::v4f64,  1 }, // vpermpd
893     { TTI::SK_PermuteSingleSrc, MVT::v8f32,  1 }, // vpermps
894     { TTI::SK_PermuteSingleSrc, MVT::v4i64,  1 }, // vpermq
895     { TTI::SK_PermuteSingleSrc, MVT::v8i32,  1 }, // vpermd
896     { TTI::SK_PermuteSingleSrc, MVT::v16i16, 4 }, // vperm2i128 + 2*vpshufb
897                                                   // + vpblendvb
898     { TTI::SK_PermuteSingleSrc, MVT::v32i8,  4 }, // vperm2i128 + 2*vpshufb
899                                                   // + vpblendvb
900 
901     { TTI::SK_PermuteTwoSrc,    MVT::v4f64,  3 }, // 2*vpermpd + vblendpd
902     { TTI::SK_PermuteTwoSrc,    MVT::v8f32,  3 }, // 2*vpermps + vblendps
903     { TTI::SK_PermuteTwoSrc,    MVT::v4i64,  3 }, // 2*vpermq + vpblendd
904     { TTI::SK_PermuteTwoSrc,    MVT::v8i32,  3 }, // 2*vpermd + vpblendd
905     { TTI::SK_PermuteTwoSrc,    MVT::v16i16, 7 }, // 2*vperm2i128 + 4*vpshufb
906                                                   // + vpblendvb
907     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,  7 }, // 2*vperm2i128 + 4*vpshufb
908                                                   // + vpblendvb
909   };
910 
911   if (ST->hasAVX2())
912     if (const auto *Entry = CostTableLookup(AVX2ShuffleTbl, Kind, LT.second))
913       return LT.first * Entry->Cost;
914 
915   static const CostTblEntry XOPShuffleTbl[] = {
916     { TTI::SK_PermuteSingleSrc, MVT::v4f64,   2 }, // vperm2f128 + vpermil2pd
917     { TTI::SK_PermuteSingleSrc, MVT::v8f32,   2 }, // vperm2f128 + vpermil2ps
918     { TTI::SK_PermuteSingleSrc, MVT::v4i64,   2 }, // vperm2f128 + vpermil2pd
919     { TTI::SK_PermuteSingleSrc, MVT::v8i32,   2 }, // vperm2f128 + vpermil2ps
920     { TTI::SK_PermuteSingleSrc, MVT::v16i16,  4 }, // vextractf128 + 2*vpperm
921                                                    // + vinsertf128
922     { TTI::SK_PermuteSingleSrc, MVT::v32i8,   4 }, // vextractf128 + 2*vpperm
923                                                    // + vinsertf128
924 
925     { TTI::SK_PermuteTwoSrc,    MVT::v16i16,  9 }, // 2*vextractf128 + 6*vpperm
926                                                    // + vinsertf128
927     { TTI::SK_PermuteTwoSrc,    MVT::v8i16,   1 }, // vpperm
928     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,   9 }, // 2*vextractf128 + 6*vpperm
929                                                    // + vinsertf128
930     { TTI::SK_PermuteTwoSrc,    MVT::v16i8,   1 }, // vpperm
931   };
932 
933   if (ST->hasXOP())
934     if (const auto *Entry = CostTableLookup(XOPShuffleTbl, Kind, LT.second))
935       return LT.first * Entry->Cost;
936 
937   static const CostTblEntry AVX1ShuffleTbl[] = {
938     { TTI::SK_Broadcast, MVT::v4f64,  2 }, // vperm2f128 + vpermilpd
939     { TTI::SK_Broadcast, MVT::v8f32,  2 }, // vperm2f128 + vpermilps
940     { TTI::SK_Broadcast, MVT::v4i64,  2 }, // vperm2f128 + vpermilpd
941     { TTI::SK_Broadcast, MVT::v8i32,  2 }, // vperm2f128 + vpermilps
942     { TTI::SK_Broadcast, MVT::v16i16, 3 }, // vpshuflw + vpshufd + vinsertf128
943     { TTI::SK_Broadcast, MVT::v32i8,  2 }, // vpshufb + vinsertf128
944 
945     { TTI::SK_Reverse,   MVT::v4f64,  2 }, // vperm2f128 + vpermilpd
946     { TTI::SK_Reverse,   MVT::v8f32,  2 }, // vperm2f128 + vpermilps
947     { TTI::SK_Reverse,   MVT::v4i64,  2 }, // vperm2f128 + vpermilpd
948     { TTI::SK_Reverse,   MVT::v8i32,  2 }, // vperm2f128 + vpermilps
949     { TTI::SK_Reverse,   MVT::v16i16, 4 }, // vextractf128 + 2*pshufb
950                                            // + vinsertf128
951     { TTI::SK_Reverse,   MVT::v32i8,  4 }, // vextractf128 + 2*pshufb
952                                            // + vinsertf128
953 
954     { TTI::SK_Alternate, MVT::v4i64,  1 }, // vblendpd
955     { TTI::SK_Alternate, MVT::v4f64,  1 }, // vblendpd
956     { TTI::SK_Alternate, MVT::v8i32,  1 }, // vblendps
957     { TTI::SK_Alternate, MVT::v8f32,  1 }, // vblendps
958     { TTI::SK_Alternate, MVT::v16i16, 3 }, // vpand + vpandn + vpor
959     { TTI::SK_Alternate, MVT::v32i8,  3 }, // vpand + vpandn + vpor
960 
961     { TTI::SK_PermuteSingleSrc, MVT::v4f64,  3 }, // 2*vperm2f128 + vshufpd
962     { TTI::SK_PermuteSingleSrc, MVT::v4i64,  3 }, // 2*vperm2f128 + vshufpd
963     { TTI::SK_PermuteSingleSrc, MVT::v8f32,  4 }, // 2*vperm2f128 + 2*vshufps
964     { TTI::SK_PermuteSingleSrc, MVT::v8i32,  4 }, // 2*vperm2f128 + 2*vshufps
965     { TTI::SK_PermuteSingleSrc, MVT::v16i16, 8 }, // vextractf128 + 4*pshufb
966                                                   // + 2*por + vinsertf128
967     { TTI::SK_PermuteSingleSrc, MVT::v32i8,  8 }, // vextractf128 + 4*pshufb
968                                                   // + 2*por + vinsertf128
969 
970     { TTI::SK_PermuteTwoSrc,    MVT::v4f64,   4 }, // 2*vperm2f128 + 2*vshufpd
971     { TTI::SK_PermuteTwoSrc,    MVT::v8f32,   4 }, // 2*vperm2f128 + 2*vshufps
972     { TTI::SK_PermuteTwoSrc,    MVT::v4i64,   4 }, // 2*vperm2f128 + 2*vshufpd
973     { TTI::SK_PermuteTwoSrc,    MVT::v8i32,   4 }, // 2*vperm2f128 + 2*vshufps
974     { TTI::SK_PermuteTwoSrc,    MVT::v16i16, 15 }, // 2*vextractf128 + 8*pshufb
975                                                    // + 4*por + vinsertf128
976     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,  15 }, // 2*vextractf128 + 8*pshufb
977                                                    // + 4*por + vinsertf128
978   };
979 
980   if (ST->hasAVX())
981     if (const auto *Entry = CostTableLookup(AVX1ShuffleTbl, Kind, LT.second))
982       return LT.first * Entry->Cost;
983 
984   static const CostTblEntry SSE41ShuffleTbl[] = {
985     { TTI::SK_Alternate, MVT::v2i64,  1 }, // pblendw
986     { TTI::SK_Alternate, MVT::v2f64,  1 }, // movsd
987     { TTI::SK_Alternate, MVT::v4i32,  1 }, // pblendw
988     { TTI::SK_Alternate, MVT::v4f32,  1 }, // blendps
989     { TTI::SK_Alternate, MVT::v8i16,  1 }, // pblendw
990     { TTI::SK_Alternate, MVT::v16i8,  1 }  // pblendvb
991   };
992 
993   if (ST->hasSSE41())
994     if (const auto *Entry = CostTableLookup(SSE41ShuffleTbl, Kind, LT.second))
995       return LT.first * Entry->Cost;
996 
997   static const CostTblEntry SSSE3ShuffleTbl[] = {
998     { TTI::SK_Broadcast, MVT::v8i16,  1 }, // pshufb
999     { TTI::SK_Broadcast, MVT::v16i8,  1 }, // pshufb
1000 
1001     { TTI::SK_Reverse,   MVT::v8i16,  1 }, // pshufb
1002     { TTI::SK_Reverse,   MVT::v16i8,  1 }, // pshufb
1003 
1004     { TTI::SK_Alternate, MVT::v8i16,  3 }, // 2*pshufb + por
1005     { TTI::SK_Alternate, MVT::v16i8,  3 }, // 2*pshufb + por
1006 
1007     { TTI::SK_PermuteSingleSrc, MVT::v8i16, 1 }, // pshufb
1008     { TTI::SK_PermuteSingleSrc, MVT::v16i8, 1 }, // pshufb
1009 
1010     { TTI::SK_PermuteTwoSrc,    MVT::v8i16, 3 }, // 2*pshufb + por
1011     { TTI::SK_PermuteTwoSrc,    MVT::v16i8, 3 }, // 2*pshufb + por
1012   };
1013 
1014   if (ST->hasSSSE3())
1015     if (const auto *Entry = CostTableLookup(SSSE3ShuffleTbl, Kind, LT.second))
1016       return LT.first * Entry->Cost;
1017 
1018   static const CostTblEntry SSE2ShuffleTbl[] = {
1019     { TTI::SK_Broadcast, MVT::v2f64,  1 }, // shufpd
1020     { TTI::SK_Broadcast, MVT::v2i64,  1 }, // pshufd
1021     { TTI::SK_Broadcast, MVT::v4i32,  1 }, // pshufd
1022     { TTI::SK_Broadcast, MVT::v8i16,  2 }, // pshuflw + pshufd
1023     { TTI::SK_Broadcast, MVT::v16i8,  3 }, // unpck + pshuflw + pshufd
1024 
1025     { TTI::SK_Reverse,   MVT::v2f64,  1 }, // shufpd
1026     { TTI::SK_Reverse,   MVT::v2i64,  1 }, // pshufd
1027     { TTI::SK_Reverse,   MVT::v4i32,  1 }, // pshufd
1028     { TTI::SK_Reverse,   MVT::v8i16,  3 }, // pshuflw + pshufhw + pshufd
1029     { TTI::SK_Reverse,   MVT::v16i8,  9 }, // 2*pshuflw + 2*pshufhw
1030                                            // + 2*pshufd + 2*unpck + packus
1031 
1032     { TTI::SK_Alternate, MVT::v2i64,  1 }, // movsd
1033     { TTI::SK_Alternate, MVT::v2f64,  1 }, // movsd
1034     { TTI::SK_Alternate, MVT::v4i32,  2 }, // 2*shufps
1035     { TTI::SK_Alternate, MVT::v8i16,  3 }, // pand + pandn + por
1036     { TTI::SK_Alternate, MVT::v16i8,  3 }, // pand + pandn + por
1037 
1038     { TTI::SK_PermuteSingleSrc, MVT::v2f64,  1 }, // shufpd
1039     { TTI::SK_PermuteSingleSrc, MVT::v2i64,  1 }, // pshufd
1040     { TTI::SK_PermuteSingleSrc, MVT::v4i32,  1 }, // pshufd
1041     { TTI::SK_PermuteSingleSrc, MVT::v8i16,  5 }, // 2*pshuflw + 2*pshufhw
1042                                                   // + pshufd/unpck
1043     { TTI::SK_PermuteSingleSrc, MVT::v16i8, 10 }, // 2*pshuflw + 2*pshufhw
1044                                                   // + 2*pshufd + 2*unpck + 2*packus
1045 
1046     { TTI::SK_PermuteTwoSrc,    MVT::v2f64,  1 }, // shufpd
1047     { TTI::SK_PermuteTwoSrc,    MVT::v2i64,  1 }, // shufpd
1048     { TTI::SK_PermuteTwoSrc,    MVT::v4i32,  2 }, // 2*{unpck,movsd,pshufd}
1049     { TTI::SK_PermuteTwoSrc,    MVT::v8i16,  8 }, // blend+permute
1050     { TTI::SK_PermuteTwoSrc,    MVT::v16i8, 13 }, // blend+permute
1051   };
1052 
1053   if (ST->hasSSE2())
1054     if (const auto *Entry = CostTableLookup(SSE2ShuffleTbl, Kind, LT.second))
1055       return LT.first * Entry->Cost;
1056 
1057   static const CostTblEntry SSE1ShuffleTbl[] = {
1058     { TTI::SK_Broadcast,        MVT::v4f32, 1 }, // shufps
1059     { TTI::SK_Reverse,          MVT::v4f32, 1 }, // shufps
1060     { TTI::SK_Alternate,        MVT::v4f32, 2 }, // 2*shufps
1061     { TTI::SK_PermuteSingleSrc, MVT::v4f32, 1 }, // shufps
1062     { TTI::SK_PermuteTwoSrc,    MVT::v4f32, 2 }, // 2*shufps
1063   };
1064 
1065   if (ST->hasSSE1())
1066     if (const auto *Entry = CostTableLookup(SSE1ShuffleTbl, Kind, LT.second))
1067       return LT.first * Entry->Cost;
1068 
1069   return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
1070 }
1071 
1072 int X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
1073                                  const Instruction *I) {
1074   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1075   assert(ISD && "Invalid opcode");
1076 
1077   // FIXME: Need a better design of the cost table to handle non-simple types of
1078   // potential massive combinations (elem_num x src_type x dst_type).
1079 
1080   static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = {
1081     { ISD::SINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
1082     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
1083     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
1084     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
1085     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
1086     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
1087 
1088     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
1089     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
1090     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
1091     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
1092     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
1093     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
1094 
1095     { ISD::FP_TO_SINT,  MVT::v2i64,  MVT::v2f32,  1 },
1096     { ISD::FP_TO_SINT,  MVT::v4i64,  MVT::v4f32,  1 },
1097     { ISD::FP_TO_SINT,  MVT::v8i64,  MVT::v8f32,  1 },
1098     { ISD::FP_TO_SINT,  MVT::v2i64,  MVT::v2f64,  1 },
1099     { ISD::FP_TO_SINT,  MVT::v4i64,  MVT::v4f64,  1 },
1100     { ISD::FP_TO_SINT,  MVT::v8i64,  MVT::v8f64,  1 },
1101 
1102     { ISD::FP_TO_UINT,  MVT::v2i64,  MVT::v2f32,  1 },
1103     { ISD::FP_TO_UINT,  MVT::v4i64,  MVT::v4f32,  1 },
1104     { ISD::FP_TO_UINT,  MVT::v8i64,  MVT::v8f32,  1 },
1105     { ISD::FP_TO_UINT,  MVT::v2i64,  MVT::v2f64,  1 },
1106     { ISD::FP_TO_UINT,  MVT::v4i64,  MVT::v4f64,  1 },
1107     { ISD::FP_TO_UINT,  MVT::v8i64,  MVT::v8f64,  1 },
1108   };
1109 
1110   // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and
1111   // 256-bit wide vectors.
1112 
1113   static const TypeConversionCostTblEntry AVX512FConversionTbl[] = {
1114     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v8f32,  1 },
1115     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v16f32, 3 },
1116     { ISD::FP_ROUND,  MVT::v8f32,   MVT::v8f64,  1 },
1117 
1118     { ISD::TRUNCATE,  MVT::v16i8,   MVT::v16i32, 1 },
1119     { ISD::TRUNCATE,  MVT::v16i16,  MVT::v16i32, 1 },
1120     { ISD::TRUNCATE,  MVT::v8i16,   MVT::v8i64,  1 },
1121     { ISD::TRUNCATE,  MVT::v8i32,   MVT::v8i64,  1 },
1122 
1123     // v16i1 -> v16i32 - load + broadcast
1124     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
1125     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
1126     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
1127     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
1128     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1129     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1130     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
1131     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
1132     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
1133     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
1134 
1135     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
1136     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
1137     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i8,   2 },
1138     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i8,  2 },
1139     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
1140     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i16, 2 },
1141     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
1142     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
1143     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64, 26 },
1144     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64, 26 },
1145 
1146     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
1147     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
1148     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i8,   2 },
1149     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i8,   2 },
1150     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i8,   2 },
1151     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i8,   2 },
1152     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i8,  2 },
1153     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i16,  5 },
1154     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i16,  2 },
1155     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  2 },
1156     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
1157     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i16, 2 },
1158     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  2 },
1159     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i32,  1 },
1160     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  1 },
1161     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  1 },
1162     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  1 },
1163     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
1164     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
1165     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  5 },
1166     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  5 },
1167     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64, 12 },
1168     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64, 26 },
1169 
1170     { ISD::FP_TO_UINT,  MVT::v2i32,  MVT::v2f32,  1 },
1171     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  1 },
1172     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f32,  1 },
1173     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v8f64,  2 },
1174     { ISD::FP_TO_UINT,  MVT::v8i8,   MVT::v8f64,  2 },
1175     { ISD::FP_TO_UINT,  MVT::v16i32, MVT::v16f32, 1 },
1176     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v16f32, 2 },
1177     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v16f32, 2 },
1178   };
1179 
1180   static const TypeConversionCostTblEntry AVX2ConversionTbl[] = {
1181     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
1182     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
1183     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
1184     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
1185     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,   3 },
1186     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,   3 },
1187     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   3 },
1188     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   3 },
1189     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
1190     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
1191     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
1192     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
1193     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
1194     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
1195     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
1196     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
1197 
1198     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i64,  2 },
1199     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i64,  2 },
1200     { ISD::TRUNCATE,    MVT::v4i32,  MVT::v4i64,  2 },
1201     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  2 },
1202     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  2 },
1203     { ISD::TRUNCATE,    MVT::v8i32,  MVT::v8i64,  4 },
1204 
1205     { ISD::FP_EXTEND,   MVT::v8f64,  MVT::v8f32,  3 },
1206     { ISD::FP_ROUND,    MVT::v8f32,  MVT::v8f64,  3 },
1207 
1208     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  8 },
1209   };
1210 
1211   static const TypeConversionCostTblEntry AVXConversionTbl[] = {
1212     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,  6 },
1213     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,  4 },
1214     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,  7 },
1215     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,  4 },
1216     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,  6 },
1217     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,  4 },
1218     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,  7 },
1219     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,  4 },
1220     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
1221     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
1222     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16, 6 },
1223     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16, 3 },
1224     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16, 4 },
1225     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16, 4 },
1226     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32, 4 },
1227     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32, 4 },
1228 
1229     { ISD::TRUNCATE,    MVT::v16i8, MVT::v16i16, 4 },
1230     { ISD::TRUNCATE,    MVT::v8i8,  MVT::v8i32,  4 },
1231     { ISD::TRUNCATE,    MVT::v8i16, MVT::v8i32,  5 },
1232     { ISD::TRUNCATE,    MVT::v4i8,  MVT::v4i64,  4 },
1233     { ISD::TRUNCATE,    MVT::v4i16, MVT::v4i64,  4 },
1234     { ISD::TRUNCATE,    MVT::v4i32, MVT::v4i64,  4 },
1235     { ISD::TRUNCATE,    MVT::v8i32, MVT::v8i64,  9 },
1236 
1237     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i1,  3 },
1238     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i1,  3 },
1239     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i1,  8 },
1240     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i8,  3 },
1241     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i8,  3 },
1242     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i8,  8 },
1243     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i16, 3 },
1244     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i16, 3 },
1245     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
1246     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },
1247     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i32, 1 },
1248     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i32, 1 },
1249 
1250     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i1,  7 },
1251     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i1,  7 },
1252     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i1,  6 },
1253     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i8,  2 },
1254     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i8,  2 },
1255     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i8,  5 },
1256     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
1257     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i16, 2 },
1258     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
1259     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i32, 6 },
1260     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i32, 6 },
1261     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i32, 6 },
1262     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i32, 9 },
1263     // The generic code to compute the scalar overhead is currently broken.
1264     // Workaround this limitation by estimating the scalarization overhead
1265     // here. We have roughly 10 instructions per scalar element.
1266     // Multiply that by the vector width.
1267     // FIXME: remove that when PR19268 is fixed.
1268     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i64, 10 },
1269     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i64, 20 },
1270     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
1271     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
1272 
1273     { ISD::FP_TO_SINT,  MVT::v4i8,  MVT::v4f32, 1 },
1274     { ISD::FP_TO_SINT,  MVT::v8i8,  MVT::v8f32, 7 },
1275     // This node is expanded into scalarized operations but BasicTTI is overly
1276     // optimistic estimating its cost.  It computes 3 per element (one
1277     // vector-extract, one scalar conversion and one vector-insert).  The
1278     // problem is that the inserts form a read-modify-write chain so latency
1279     // should be factored in too.  Inflating the cost per element by 1.
1280     { ISD::FP_TO_UINT,  MVT::v8i32, MVT::v8f32, 8*4 },
1281     { ISD::FP_TO_UINT,  MVT::v4i32, MVT::v4f64, 4*4 },
1282 
1283     { ISD::FP_EXTEND,   MVT::v4f64,  MVT::v4f32,  1 },
1284     { ISD::FP_ROUND,    MVT::v4f32,  MVT::v4f64,  1 },
1285   };
1286 
1287   static const TypeConversionCostTblEntry SSE41ConversionTbl[] = {
1288     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8,    2 },
1289     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8,    2 },
1290     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16,   2 },
1291     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16,   2 },
1292     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32,   2 },
1293     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32,   2 },
1294 
1295     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i8,   1 },
1296     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i8,   2 },
1297     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i8,   1 },
1298     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i8,   1 },
1299     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
1300     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
1301     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   2 },
1302     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   2 },
1303     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
1304     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
1305     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  4 },
1306     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  4 },
1307     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
1308     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
1309     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
1310     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
1311     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 4 },
1312     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 4 },
1313 
1314     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  2 },
1315     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  1 },
1316     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i32,  1 },
1317     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i32,  1 },
1318     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  3 },
1319     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  3 },
1320     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 6 },
1321 
1322   };
1323 
1324   static const TypeConversionCostTblEntry SSE2ConversionTbl[] = {
1325     // These are somewhat magic numbers justified by looking at the output of
1326     // Intel's IACA, running some kernels and making sure when we take
1327     // legalization into account the throughput will be overestimated.
1328     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
1329     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
1330     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
1331     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
1332     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 },
1333     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
1334     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
1335     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
1336 
1337     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
1338     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
1339     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
1340     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
1341     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
1342     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 8 },
1343     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
1344     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
1345 
1346     { ISD::FP_TO_SINT,  MVT::v2i32,  MVT::v2f64,  3 },
1347 
1348     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i8,   1 },
1349     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i8,   6 },
1350     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i8,   2 },
1351     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i8,   3 },
1352     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,   4 },
1353     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,   8 },
1354     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
1355     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i8,   2 },
1356     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   6 },
1357     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   6 },
1358     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  3 },
1359     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  4 },
1360     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  9 },
1361     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  12 },
1362     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
1363     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i16,  2 },
1364     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
1365     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16,  10 },
1366     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  3 },
1367     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  4 },
1368     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 6 },
1369     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 8 },
1370     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  3 },
1371     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  5 },
1372 
1373     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  4 },
1374     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  2 },
1375     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 3 },
1376     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i32,  3 },
1377     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i32,  3 },
1378     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  4 },
1379     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i32, 7 },
1380     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  5 },
1381     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 10 },
1382   };
1383 
1384   std::pair<int, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src);
1385   std::pair<int, MVT> LTDest = TLI->getTypeLegalizationCost(DL, Dst);
1386 
1387   if (ST->hasSSE2() && !ST->hasAVX()) {
1388     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
1389                                                    LTDest.second, LTSrc.second))
1390       return LTSrc.first * Entry->Cost;
1391   }
1392 
1393   EVT SrcTy = TLI->getValueType(DL, Src);
1394   EVT DstTy = TLI->getValueType(DL, Dst);
1395 
1396   // The function getSimpleVT only handles simple value types.
1397   if (!SrcTy.isSimple() || !DstTy.isSimple())
1398     return BaseT::getCastInstrCost(Opcode, Dst, Src);
1399 
1400   if (ST->hasDQI())
1401     if (const auto *Entry = ConvertCostTableLookup(AVX512DQConversionTbl, ISD,
1402                                                    DstTy.getSimpleVT(),
1403                                                    SrcTy.getSimpleVT()))
1404       return Entry->Cost;
1405 
1406   if (ST->hasAVX512())
1407     if (const auto *Entry = ConvertCostTableLookup(AVX512FConversionTbl, ISD,
1408                                                    DstTy.getSimpleVT(),
1409                                                    SrcTy.getSimpleVT()))
1410       return Entry->Cost;
1411 
1412   if (ST->hasAVX2()) {
1413     if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
1414                                                    DstTy.getSimpleVT(),
1415                                                    SrcTy.getSimpleVT()))
1416       return Entry->Cost;
1417   }
1418 
1419   if (ST->hasAVX()) {
1420     if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
1421                                                    DstTy.getSimpleVT(),
1422                                                    SrcTy.getSimpleVT()))
1423       return Entry->Cost;
1424   }
1425 
1426   if (ST->hasSSE41()) {
1427     if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
1428                                                    DstTy.getSimpleVT(),
1429                                                    SrcTy.getSimpleVT()))
1430       return Entry->Cost;
1431   }
1432 
1433   if (ST->hasSSE2()) {
1434     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
1435                                                    DstTy.getSimpleVT(),
1436                                                    SrcTy.getSimpleVT()))
1437       return Entry->Cost;
1438   }
1439 
1440   return BaseT::getCastInstrCost(Opcode, Dst, Src, I);
1441 }
1442 
1443 int X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
1444                                    const Instruction *I) {
1445   // Legalize the type.
1446   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
1447 
1448   MVT MTy = LT.second;
1449 
1450   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1451   assert(ISD && "Invalid opcode");
1452 
1453   static const CostTblEntry SSE2CostTbl[] = {
1454     { ISD::SETCC,   MVT::v2i64,   8 },
1455     { ISD::SETCC,   MVT::v4i32,   1 },
1456     { ISD::SETCC,   MVT::v8i16,   1 },
1457     { ISD::SETCC,   MVT::v16i8,   1 },
1458   };
1459 
1460   static const CostTblEntry SSE42CostTbl[] = {
1461     { ISD::SETCC,   MVT::v2f64,   1 },
1462     { ISD::SETCC,   MVT::v4f32,   1 },
1463     { ISD::SETCC,   MVT::v2i64,   1 },
1464   };
1465 
1466   static const CostTblEntry AVX1CostTbl[] = {
1467     { ISD::SETCC,   MVT::v4f64,   1 },
1468     { ISD::SETCC,   MVT::v8f32,   1 },
1469     // AVX1 does not support 8-wide integer compare.
1470     { ISD::SETCC,   MVT::v4i64,   4 },
1471     { ISD::SETCC,   MVT::v8i32,   4 },
1472     { ISD::SETCC,   MVT::v16i16,  4 },
1473     { ISD::SETCC,   MVT::v32i8,   4 },
1474   };
1475 
1476   static const CostTblEntry AVX2CostTbl[] = {
1477     { ISD::SETCC,   MVT::v4i64,   1 },
1478     { ISD::SETCC,   MVT::v8i32,   1 },
1479     { ISD::SETCC,   MVT::v16i16,  1 },
1480     { ISD::SETCC,   MVT::v32i8,   1 },
1481   };
1482 
1483   static const CostTblEntry AVX512CostTbl[] = {
1484     { ISD::SETCC,   MVT::v8i64,   1 },
1485     { ISD::SETCC,   MVT::v16i32,  1 },
1486     { ISD::SETCC,   MVT::v8f64,   1 },
1487     { ISD::SETCC,   MVT::v16f32,  1 },
1488   };
1489 
1490   if (ST->hasAVX512())
1491     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
1492       return LT.first * Entry->Cost;
1493 
1494   if (ST->hasAVX2())
1495     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
1496       return LT.first * Entry->Cost;
1497 
1498   if (ST->hasAVX())
1499     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
1500       return LT.first * Entry->Cost;
1501 
1502   if (ST->hasSSE42())
1503     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
1504       return LT.first * Entry->Cost;
1505 
1506   if (ST->hasSSE2())
1507     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
1508       return LT.first * Entry->Cost;
1509 
1510   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, I);
1511 }
1512 
1513 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; }
1514 
1515 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1516                                       ArrayRef<Type *> Tys, FastMathFlags FMF,
1517                                       unsigned ScalarizationCostPassed) {
1518   // Costs should match the codegen from:
1519   // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll
1520   // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll
1521   // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll
1522   // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll
1523   // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll
1524   static const CostTblEntry AVX512CDCostTbl[] = {
1525     { ISD::CTLZ,       MVT::v8i64,   1 },
1526     { ISD::CTLZ,       MVT::v16i32,  1 },
1527     { ISD::CTLZ,       MVT::v32i16,  8 },
1528     { ISD::CTLZ,       MVT::v64i8,  20 },
1529     { ISD::CTLZ,       MVT::v4i64,   1 },
1530     { ISD::CTLZ,       MVT::v8i32,   1 },
1531     { ISD::CTLZ,       MVT::v16i16,  4 },
1532     { ISD::CTLZ,       MVT::v32i8,  10 },
1533     { ISD::CTLZ,       MVT::v2i64,   1 },
1534     { ISD::CTLZ,       MVT::v4i32,   1 },
1535     { ISD::CTLZ,       MVT::v8i16,   4 },
1536     { ISD::CTLZ,       MVT::v16i8,   4 },
1537   };
1538   static const CostTblEntry AVX512BWCostTbl[] = {
1539     { ISD::BITREVERSE, MVT::v8i64,   5 },
1540     { ISD::BITREVERSE, MVT::v16i32,  5 },
1541     { ISD::BITREVERSE, MVT::v32i16,  5 },
1542     { ISD::BITREVERSE, MVT::v64i8,   5 },
1543     { ISD::CTLZ,       MVT::v8i64,  23 },
1544     { ISD::CTLZ,       MVT::v16i32, 22 },
1545     { ISD::CTLZ,       MVT::v32i16, 18 },
1546     { ISD::CTLZ,       MVT::v64i8,  17 },
1547     { ISD::CTPOP,      MVT::v8i64,   7 },
1548     { ISD::CTPOP,      MVT::v16i32, 11 },
1549     { ISD::CTPOP,      MVT::v32i16,  9 },
1550     { ISD::CTPOP,      MVT::v64i8,   6 },
1551     { ISD::CTTZ,       MVT::v8i64,  10 },
1552     { ISD::CTTZ,       MVT::v16i32, 14 },
1553     { ISD::CTTZ,       MVT::v32i16, 12 },
1554     { ISD::CTTZ,       MVT::v64i8,   9 },
1555   };
1556   static const CostTblEntry AVX512CostTbl[] = {
1557     { ISD::BITREVERSE, MVT::v8i64,  36 },
1558     { ISD::BITREVERSE, MVT::v16i32, 24 },
1559     { ISD::CTLZ,       MVT::v8i64,  29 },
1560     { ISD::CTLZ,       MVT::v16i32, 35 },
1561     { ISD::CTPOP,      MVT::v8i64,  16 },
1562     { ISD::CTPOP,      MVT::v16i32, 24 },
1563     { ISD::CTTZ,       MVT::v8i64,  20 },
1564     { ISD::CTTZ,       MVT::v16i32, 28 },
1565   };
1566   static const CostTblEntry XOPCostTbl[] = {
1567     { ISD::BITREVERSE, MVT::v4i64,   4 },
1568     { ISD::BITREVERSE, MVT::v8i32,   4 },
1569     { ISD::BITREVERSE, MVT::v16i16,  4 },
1570     { ISD::BITREVERSE, MVT::v32i8,   4 },
1571     { ISD::BITREVERSE, MVT::v2i64,   1 },
1572     { ISD::BITREVERSE, MVT::v4i32,   1 },
1573     { ISD::BITREVERSE, MVT::v8i16,   1 },
1574     { ISD::BITREVERSE, MVT::v16i8,   1 },
1575     { ISD::BITREVERSE, MVT::i64,     3 },
1576     { ISD::BITREVERSE, MVT::i32,     3 },
1577     { ISD::BITREVERSE, MVT::i16,     3 },
1578     { ISD::BITREVERSE, MVT::i8,      3 }
1579   };
1580   static const CostTblEntry AVX2CostTbl[] = {
1581     { ISD::BITREVERSE, MVT::v4i64,   5 },
1582     { ISD::BITREVERSE, MVT::v8i32,   5 },
1583     { ISD::BITREVERSE, MVT::v16i16,  5 },
1584     { ISD::BITREVERSE, MVT::v32i8,   5 },
1585     { ISD::BSWAP,      MVT::v4i64,   1 },
1586     { ISD::BSWAP,      MVT::v8i32,   1 },
1587     { ISD::BSWAP,      MVT::v16i16,  1 },
1588     { ISD::CTLZ,       MVT::v4i64,  23 },
1589     { ISD::CTLZ,       MVT::v8i32,  18 },
1590     { ISD::CTLZ,       MVT::v16i16, 14 },
1591     { ISD::CTLZ,       MVT::v32i8,   9 },
1592     { ISD::CTPOP,      MVT::v4i64,   7 },
1593     { ISD::CTPOP,      MVT::v8i32,  11 },
1594     { ISD::CTPOP,      MVT::v16i16,  9 },
1595     { ISD::CTPOP,      MVT::v32i8,   6 },
1596     { ISD::CTTZ,       MVT::v4i64,  10 },
1597     { ISD::CTTZ,       MVT::v8i32,  14 },
1598     { ISD::CTTZ,       MVT::v16i16, 12 },
1599     { ISD::CTTZ,       MVT::v32i8,   9 },
1600     { ISD::FSQRT,      MVT::f32,     7 }, // Haswell from http://www.agner.org/
1601     { ISD::FSQRT,      MVT::v4f32,   7 }, // Haswell from http://www.agner.org/
1602     { ISD::FSQRT,      MVT::v8f32,  14 }, // Haswell from http://www.agner.org/
1603     { ISD::FSQRT,      MVT::f64,    14 }, // Haswell from http://www.agner.org/
1604     { ISD::FSQRT,      MVT::v2f64,  14 }, // Haswell from http://www.agner.org/
1605     { ISD::FSQRT,      MVT::v4f64,  28 }, // Haswell from http://www.agner.org/
1606   };
1607   static const CostTblEntry AVX1CostTbl[] = {
1608     { ISD::BITREVERSE, MVT::v4i64,  12 }, // 2 x 128-bit Op + extract/insert
1609     { ISD::BITREVERSE, MVT::v8i32,  12 }, // 2 x 128-bit Op + extract/insert
1610     { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert
1611     { ISD::BITREVERSE, MVT::v32i8,  12 }, // 2 x 128-bit Op + extract/insert
1612     { ISD::BSWAP,      MVT::v4i64,   4 },
1613     { ISD::BSWAP,      MVT::v8i32,   4 },
1614     { ISD::BSWAP,      MVT::v16i16,  4 },
1615     { ISD::CTLZ,       MVT::v4i64,  48 }, // 2 x 128-bit Op + extract/insert
1616     { ISD::CTLZ,       MVT::v8i32,  38 }, // 2 x 128-bit Op + extract/insert
1617     { ISD::CTLZ,       MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert
1618     { ISD::CTLZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
1619     { ISD::CTPOP,      MVT::v4i64,  16 }, // 2 x 128-bit Op + extract/insert
1620     { ISD::CTPOP,      MVT::v8i32,  24 }, // 2 x 128-bit Op + extract/insert
1621     { ISD::CTPOP,      MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert
1622     { ISD::CTPOP,      MVT::v32i8,  14 }, // 2 x 128-bit Op + extract/insert
1623     { ISD::CTTZ,       MVT::v4i64,  22 }, // 2 x 128-bit Op + extract/insert
1624     { ISD::CTTZ,       MVT::v8i32,  30 }, // 2 x 128-bit Op + extract/insert
1625     { ISD::CTTZ,       MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert
1626     { ISD::CTTZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
1627     { ISD::FSQRT,      MVT::f32,    14 }, // SNB from http://www.agner.org/
1628     { ISD::FSQRT,      MVT::v4f32,  14 }, // SNB from http://www.agner.org/
1629     { ISD::FSQRT,      MVT::v8f32,  28 }, // SNB from http://www.agner.org/
1630     { ISD::FSQRT,      MVT::f64,    21 }, // SNB from http://www.agner.org/
1631     { ISD::FSQRT,      MVT::v2f64,  21 }, // SNB from http://www.agner.org/
1632     { ISD::FSQRT,      MVT::v4f64,  43 }, // SNB from http://www.agner.org/
1633   };
1634   static const CostTblEntry SSE42CostTbl[] = {
1635     { ISD::FSQRT,      MVT::f32,    18 }, // Nehalem from http://www.agner.org/
1636     { ISD::FSQRT,      MVT::v4f32,  18 }, // Nehalem from http://www.agner.org/
1637   };
1638   static const CostTblEntry SSSE3CostTbl[] = {
1639     { ISD::BITREVERSE, MVT::v2i64,   5 },
1640     { ISD::BITREVERSE, MVT::v4i32,   5 },
1641     { ISD::BITREVERSE, MVT::v8i16,   5 },
1642     { ISD::BITREVERSE, MVT::v16i8,   5 },
1643     { ISD::BSWAP,      MVT::v2i64,   1 },
1644     { ISD::BSWAP,      MVT::v4i32,   1 },
1645     { ISD::BSWAP,      MVT::v8i16,   1 },
1646     { ISD::CTLZ,       MVT::v2i64,  23 },
1647     { ISD::CTLZ,       MVT::v4i32,  18 },
1648     { ISD::CTLZ,       MVT::v8i16,  14 },
1649     { ISD::CTLZ,       MVT::v16i8,   9 },
1650     { ISD::CTPOP,      MVT::v2i64,   7 },
1651     { ISD::CTPOP,      MVT::v4i32,  11 },
1652     { ISD::CTPOP,      MVT::v8i16,   9 },
1653     { ISD::CTPOP,      MVT::v16i8,   6 },
1654     { ISD::CTTZ,       MVT::v2i64,  10 },
1655     { ISD::CTTZ,       MVT::v4i32,  14 },
1656     { ISD::CTTZ,       MVT::v8i16,  12 },
1657     { ISD::CTTZ,       MVT::v16i8,   9 }
1658   };
1659   static const CostTblEntry SSE2CostTbl[] = {
1660     { ISD::BITREVERSE, MVT::v2i64,  29 },
1661     { ISD::BITREVERSE, MVT::v4i32,  27 },
1662     { ISD::BITREVERSE, MVT::v8i16,  27 },
1663     { ISD::BITREVERSE, MVT::v16i8,  20 },
1664     { ISD::BSWAP,      MVT::v2i64,   7 },
1665     { ISD::BSWAP,      MVT::v4i32,   7 },
1666     { ISD::BSWAP,      MVT::v8i16,   7 },
1667     { ISD::CTLZ,       MVT::v2i64,  25 },
1668     { ISD::CTLZ,       MVT::v4i32,  26 },
1669     { ISD::CTLZ,       MVT::v8i16,  20 },
1670     { ISD::CTLZ,       MVT::v16i8,  17 },
1671     { ISD::CTPOP,      MVT::v2i64,  12 },
1672     { ISD::CTPOP,      MVT::v4i32,  15 },
1673     { ISD::CTPOP,      MVT::v8i16,  13 },
1674     { ISD::CTPOP,      MVT::v16i8,  10 },
1675     { ISD::CTTZ,       MVT::v2i64,  14 },
1676     { ISD::CTTZ,       MVT::v4i32,  18 },
1677     { ISD::CTTZ,       MVT::v8i16,  16 },
1678     { ISD::CTTZ,       MVT::v16i8,  13 },
1679     { ISD::FSQRT,      MVT::f64,    32 }, // Nehalem from http://www.agner.org/
1680     { ISD::FSQRT,      MVT::v2f64,  32 }, // Nehalem from http://www.agner.org/
1681   };
1682   static const CostTblEntry SSE1CostTbl[] = {
1683     { ISD::FSQRT,      MVT::f32,    28 }, // Pentium III from http://www.agner.org/
1684     { ISD::FSQRT,      MVT::v4f32,  56 }, // Pentium III from http://www.agner.org/
1685   };
1686   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
1687     { ISD::BITREVERSE, MVT::i64,    14 }
1688   };
1689   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
1690     { ISD::BITREVERSE, MVT::i32,    14 },
1691     { ISD::BITREVERSE, MVT::i16,    14 },
1692     { ISD::BITREVERSE, MVT::i8,     11 }
1693   };
1694 
1695   unsigned ISD = ISD::DELETED_NODE;
1696   switch (IID) {
1697   default:
1698     break;
1699   case Intrinsic::bitreverse:
1700     ISD = ISD::BITREVERSE;
1701     break;
1702   case Intrinsic::bswap:
1703     ISD = ISD::BSWAP;
1704     break;
1705   case Intrinsic::ctlz:
1706     ISD = ISD::CTLZ;
1707     break;
1708   case Intrinsic::ctpop:
1709     ISD = ISD::CTPOP;
1710     break;
1711   case Intrinsic::cttz:
1712     ISD = ISD::CTTZ;
1713     break;
1714   case Intrinsic::sqrt:
1715     ISD = ISD::FSQRT;
1716     break;
1717   }
1718 
1719   // Legalize the type.
1720   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
1721   MVT MTy = LT.second;
1722 
1723   // Attempt to lookup cost.
1724   if (ST->hasCDI())
1725     if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy))
1726       return LT.first * Entry->Cost;
1727 
1728   if (ST->hasBWI())
1729     if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
1730       return LT.first * Entry->Cost;
1731 
1732   if (ST->hasAVX512())
1733     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
1734       return LT.first * Entry->Cost;
1735 
1736   if (ST->hasXOP())
1737     if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
1738       return LT.first * Entry->Cost;
1739 
1740   if (ST->hasAVX2())
1741     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
1742       return LT.first * Entry->Cost;
1743 
1744   if (ST->hasAVX())
1745     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
1746       return LT.first * Entry->Cost;
1747 
1748   if (ST->hasSSE42())
1749     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
1750       return LT.first * Entry->Cost;
1751 
1752   if (ST->hasSSSE3())
1753     if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy))
1754       return LT.first * Entry->Cost;
1755 
1756   if (ST->hasSSE2())
1757     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
1758       return LT.first * Entry->Cost;
1759 
1760   if (ST->hasSSE1())
1761     if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
1762       return LT.first * Entry->Cost;
1763 
1764   if (ST->is64Bit())
1765     if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
1766       return LT.first * Entry->Cost;
1767 
1768   if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
1769     return LT.first * Entry->Cost;
1770 
1771   return BaseT::getIntrinsicInstrCost(IID, RetTy, Tys, FMF, ScalarizationCostPassed);
1772 }
1773 
1774 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1775                      ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) {
1776   return BaseT::getIntrinsicInstrCost(IID, RetTy, Args, FMF, VF);
1777 }
1778 
1779 int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
1780   assert(Val->isVectorTy() && "This must be a vector type");
1781 
1782   Type *ScalarType = Val->getScalarType();
1783 
1784   if (Index != -1U) {
1785     // Legalize the type.
1786     std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Val);
1787 
1788     // This type is legalized to a scalar type.
1789     if (!LT.second.isVector())
1790       return 0;
1791 
1792     // The type may be split. Normalize the index to the new type.
1793     unsigned Width = LT.second.getVectorNumElements();
1794     Index = Index % Width;
1795 
1796     // Floating point scalars are already located in index #0.
1797     if (ScalarType->isFloatingPointTy() && Index == 0)
1798       return 0;
1799   }
1800 
1801   // Add to the base cost if we know that the extracted element of a vector is
1802   // destined to be moved to and used in the integer register file.
1803   int RegisterFileMoveCost = 0;
1804   if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy())
1805     RegisterFileMoveCost = 1;
1806 
1807   return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost;
1808 }
1809 
1810 int X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
1811                                 unsigned AddressSpace, const Instruction *I) {
1812   // Handle non-power-of-two vectors such as <3 x float>
1813   if (VectorType *VTy = dyn_cast<VectorType>(Src)) {
1814     unsigned NumElem = VTy->getVectorNumElements();
1815 
1816     // Handle a few common cases:
1817     // <3 x float>
1818     if (NumElem == 3 && VTy->getScalarSizeInBits() == 32)
1819       // Cost = 64 bit store + extract + 32 bit store.
1820       return 3;
1821 
1822     // <3 x double>
1823     if (NumElem == 3 && VTy->getScalarSizeInBits() == 64)
1824       // Cost = 128 bit store + unpack + 64 bit store.
1825       return 3;
1826 
1827     // Assume that all other non-power-of-two numbers are scalarized.
1828     if (!isPowerOf2_32(NumElem)) {
1829       int Cost = BaseT::getMemoryOpCost(Opcode, VTy->getScalarType(), Alignment,
1830                                         AddressSpace);
1831       int SplitCost = getScalarizationOverhead(Src, Opcode == Instruction::Load,
1832                                                Opcode == Instruction::Store);
1833       return NumElem * Cost + SplitCost;
1834     }
1835   }
1836 
1837   // Legalize the type.
1838   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
1839   assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
1840          "Invalid Opcode");
1841 
1842   // Each load/store unit costs 1.
1843   int Cost = LT.first * 1;
1844 
1845   // This isn't exactly right. We're using slow unaligned 32-byte accesses as a
1846   // proxy for a double-pumped AVX memory interface such as on Sandybridge.
1847   if (LT.second.getStoreSize() == 32 && ST->isUnalignedMem32Slow())
1848     Cost *= 2;
1849 
1850   return Cost;
1851 }
1852 
1853 int X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy,
1854                                       unsigned Alignment,
1855                                       unsigned AddressSpace) {
1856   VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy);
1857   if (!SrcVTy)
1858     // To calculate scalar take the regular cost, without mask
1859     return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace);
1860 
1861   unsigned NumElem = SrcVTy->getVectorNumElements();
1862   VectorType *MaskTy =
1863     VectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem);
1864   if ((Opcode == Instruction::Load && !isLegalMaskedLoad(SrcVTy)) ||
1865       (Opcode == Instruction::Store && !isLegalMaskedStore(SrcVTy)) ||
1866       !isPowerOf2_32(NumElem)) {
1867     // Scalarization
1868     int MaskSplitCost = getScalarizationOverhead(MaskTy, false, true);
1869     int ScalarCompareCost = getCmpSelInstrCost(
1870         Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr);
1871     int BranchCost = getCFInstrCost(Instruction::Br);
1872     int MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost);
1873 
1874     int ValueSplitCost = getScalarizationOverhead(
1875         SrcVTy, Opcode == Instruction::Load, Opcode == Instruction::Store);
1876     int MemopCost =
1877         NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
1878                                          Alignment, AddressSpace);
1879     return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost;
1880   }
1881 
1882   // Legalize the type.
1883   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy);
1884   auto VT = TLI->getValueType(DL, SrcVTy);
1885   int Cost = 0;
1886   if (VT.isSimple() && LT.second != VT.getSimpleVT() &&
1887       LT.second.getVectorNumElements() == NumElem)
1888     // Promotion requires expand/truncate for data and a shuffle for mask.
1889     Cost += getShuffleCost(TTI::SK_Alternate, SrcVTy, 0, nullptr) +
1890             getShuffleCost(TTI::SK_Alternate, MaskTy, 0, nullptr);
1891 
1892   else if (LT.second.getVectorNumElements() > NumElem) {
1893     VectorType *NewMaskTy = VectorType::get(MaskTy->getVectorElementType(),
1894                                             LT.second.getVectorNumElements());
1895     // Expanding requires fill mask with zeroes
1896     Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, 0, MaskTy);
1897   }
1898   if (!ST->hasAVX512())
1899     return Cost + LT.first*4; // Each maskmov costs 4
1900 
1901   // AVX-512 masked load/store is cheapper
1902   return Cost+LT.first;
1903 }
1904 
1905 int X86TTIImpl::getAddressComputationCost(Type *Ty, ScalarEvolution *SE,
1906                                           const SCEV *Ptr) {
1907   // Address computations in vectorized code with non-consecutive addresses will
1908   // likely result in more instructions compared to scalar code where the
1909   // computation can more often be merged into the index mode. The resulting
1910   // extra micro-ops can significantly decrease throughput.
1911   unsigned NumVectorInstToHideOverhead = 10;
1912 
1913   // Cost modeling of Strided Access Computation is hidden by the indexing
1914   // modes of X86 regardless of the stride value. We dont believe that there
1915   // is a difference between constant strided access in gerenal and constant
1916   // strided value which is less than or equal to 64.
1917   // Even in the case of (loop invariant) stride whose value is not known at
1918   // compile time, the address computation will not incur more than one extra
1919   // ADD instruction.
1920   if (Ty->isVectorTy() && SE) {
1921     if (!BaseT::isStridedAccess(Ptr))
1922       return NumVectorInstToHideOverhead;
1923     if (!BaseT::getConstantStrideStep(SE, Ptr))
1924       return 1;
1925   }
1926 
1927   return BaseT::getAddressComputationCost(Ty, SE, Ptr);
1928 }
1929 
1930 int X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, Type *ValTy,
1931                                            bool IsPairwise) {
1932 
1933   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
1934 
1935   MVT MTy = LT.second;
1936 
1937   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1938   assert(ISD && "Invalid opcode");
1939 
1940   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
1941   // and make it as the cost.
1942 
1943   static const CostTblEntry SSE42CostTblPairWise[] = {
1944     { ISD::FADD,  MVT::v2f64,   2 },
1945     { ISD::FADD,  MVT::v4f32,   4 },
1946     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
1947     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.5".
1948     { ISD::ADD,   MVT::v8i16,   5 },
1949   };
1950 
1951   static const CostTblEntry AVX1CostTblPairWise[] = {
1952     { ISD::FADD,  MVT::v4f32,   4 },
1953     { ISD::FADD,  MVT::v4f64,   5 },
1954     { ISD::FADD,  MVT::v8f32,   7 },
1955     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
1956     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.5".
1957     { ISD::ADD,   MVT::v4i64,   5 },      // The data reported by the IACA tool is "4.8".
1958     { ISD::ADD,   MVT::v8i16,   5 },
1959     { ISD::ADD,   MVT::v8i32,   5 },
1960   };
1961 
1962   static const CostTblEntry SSE42CostTblNoPairWise[] = {
1963     { ISD::FADD,  MVT::v2f64,   2 },
1964     { ISD::FADD,  MVT::v4f32,   4 },
1965     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
1966     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.3".
1967     { ISD::ADD,   MVT::v8i16,   4 },      // The data reported by the IACA tool is "4.3".
1968   };
1969 
1970   static const CostTblEntry AVX1CostTblNoPairWise[] = {
1971     { ISD::FADD,  MVT::v4f32,   3 },
1972     { ISD::FADD,  MVT::v4f64,   3 },
1973     { ISD::FADD,  MVT::v8f32,   4 },
1974     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
1975     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "2.8".
1976     { ISD::ADD,   MVT::v4i64,   3 },
1977     { ISD::ADD,   MVT::v8i16,   4 },
1978     { ISD::ADD,   MVT::v8i32,   5 },
1979   };
1980 
1981   if (IsPairwise) {
1982     if (ST->hasAVX())
1983       if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy))
1984         return LT.first * Entry->Cost;
1985 
1986     if (ST->hasSSE42())
1987       if (const auto *Entry = CostTableLookup(SSE42CostTblPairWise, ISD, MTy))
1988         return LT.first * Entry->Cost;
1989   } else {
1990     if (ST->hasAVX())
1991       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
1992         return LT.first * Entry->Cost;
1993 
1994     if (ST->hasSSE42())
1995       if (const auto *Entry = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy))
1996         return LT.first * Entry->Cost;
1997   }
1998 
1999   return BaseT::getArithmeticReductionCost(Opcode, ValTy, IsPairwise);
2000 }
2001 
2002 int X86TTIImpl::getMinMaxReductionCost(Type *ValTy, Type *CondTy,
2003                                        bool IsPairwise, bool IsUnsigned) {
2004   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
2005 
2006   MVT MTy = LT.second;
2007 
2008   int ISD;
2009   if (ValTy->isIntOrIntVectorTy()) {
2010     ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN;
2011   } else {
2012     assert(ValTy->isFPOrFPVectorTy() &&
2013            "Expected float point or integer vector type.");
2014     ISD = ISD::FMINNUM;
2015   }
2016 
2017   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
2018   // and make it as the cost.
2019 
2020   static const CostTblEntry SSE42CostTblPairWise[] = {
2021       {ISD::FMINNUM, MVT::v2f64, 3},
2022       {ISD::FMINNUM, MVT::v4f32, 2},
2023       {ISD::SMIN, MVT::v2i64, 7}, // The data reported by the IACA is "6.8"
2024       {ISD::UMIN, MVT::v2i64, 8}, // The data reported by the IACA is "8.6"
2025       {ISD::SMIN, MVT::v4i32, 1}, // The data reported by the IACA is "1.5"
2026       {ISD::UMIN, MVT::v4i32, 2}, // The data reported by the IACA is "1.8"
2027       {ISD::SMIN, MVT::v8i16, 2},
2028       {ISD::UMIN, MVT::v8i16, 2},
2029   };
2030 
2031   static const CostTblEntry AVX1CostTblPairWise[] = {
2032       {ISD::FMINNUM, MVT::v4f32, 1},
2033       {ISD::FMINNUM, MVT::v4f64, 1},
2034       {ISD::FMINNUM, MVT::v8f32, 2},
2035       {ISD::SMIN, MVT::v2i64, 3},
2036       {ISD::UMIN, MVT::v2i64, 3},
2037       {ISD::SMIN, MVT::v4i32, 1},
2038       {ISD::UMIN, MVT::v4i32, 1},
2039       {ISD::SMIN, MVT::v8i16, 1},
2040       {ISD::UMIN, MVT::v8i16, 1},
2041       {ISD::SMIN, MVT::v8i32, 3},
2042       {ISD::UMIN, MVT::v8i32, 3},
2043   };
2044 
2045   static const CostTblEntry AVX2CostTblPairWise[] = {
2046       {ISD::SMIN, MVT::v4i64, 2},
2047       {ISD::UMIN, MVT::v4i64, 2},
2048       {ISD::SMIN, MVT::v8i32, 1},
2049       {ISD::UMIN, MVT::v8i32, 1},
2050       {ISD::SMIN, MVT::v16i16, 1},
2051       {ISD::UMIN, MVT::v16i16, 1},
2052       {ISD::SMIN, MVT::v32i8, 2},
2053       {ISD::UMIN, MVT::v32i8, 2},
2054   };
2055 
2056   static const CostTblEntry AVX512CostTblPairWise[] = {
2057       {ISD::FMINNUM, MVT::v8f64, 1},
2058       {ISD::FMINNUM, MVT::v16f32, 2},
2059       {ISD::SMIN, MVT::v8i64, 2},
2060       {ISD::UMIN, MVT::v8i64, 2},
2061       {ISD::SMIN, MVT::v16i32, 1},
2062       {ISD::UMIN, MVT::v16i32, 1},
2063   };
2064 
2065   static const CostTblEntry SSE42CostTblNoPairWise[] = {
2066       {ISD::FMINNUM, MVT::v2f64, 3},
2067       {ISD::FMINNUM, MVT::v4f32, 3},
2068       {ISD::SMIN, MVT::v2i64, 7}, // The data reported by the IACA is "6.8"
2069       {ISD::UMIN, MVT::v2i64, 9}, // The data reported by the IACA is "8.6"
2070       {ISD::SMIN, MVT::v4i32, 1}, // The data reported by the IACA is "1.5"
2071       {ISD::UMIN, MVT::v4i32, 2}, // The data reported by the IACA is "1.8"
2072       {ISD::SMIN, MVT::v8i16, 1}, // The data reported by the IACA is "1.5"
2073       {ISD::UMIN, MVT::v8i16, 2}, // The data reported by the IACA is "1.8"
2074   };
2075 
2076   static const CostTblEntry AVX1CostTblNoPairWise[] = {
2077       {ISD::FMINNUM, MVT::v4f32, 1},
2078       {ISD::FMINNUM, MVT::v4f64, 1},
2079       {ISD::FMINNUM, MVT::v8f32, 1},
2080       {ISD::SMIN, MVT::v2i64, 3},
2081       {ISD::UMIN, MVT::v2i64, 3},
2082       {ISD::SMIN, MVT::v4i32, 1},
2083       {ISD::UMIN, MVT::v4i32, 1},
2084       {ISD::SMIN, MVT::v8i16, 1},
2085       {ISD::UMIN, MVT::v8i16, 1},
2086       {ISD::SMIN, MVT::v8i32, 2},
2087       {ISD::UMIN, MVT::v8i32, 2},
2088   };
2089 
2090   static const CostTblEntry AVX2CostTblNoPairWise[] = {
2091       {ISD::SMIN, MVT::v4i64, 1},
2092       {ISD::UMIN, MVT::v4i64, 1},
2093       {ISD::SMIN, MVT::v8i32, 1},
2094       {ISD::UMIN, MVT::v8i32, 1},
2095       {ISD::SMIN, MVT::v16i16, 1},
2096       {ISD::UMIN, MVT::v16i16, 1},
2097       {ISD::SMIN, MVT::v32i8, 1},
2098       {ISD::UMIN, MVT::v32i8, 1},
2099   };
2100 
2101   static const CostTblEntry AVX512CostTblNoPairWise[] = {
2102       {ISD::FMINNUM, MVT::v8f64, 1},
2103       {ISD::FMINNUM, MVT::v16f32, 2},
2104       {ISD::SMIN, MVT::v8i64, 1},
2105       {ISD::UMIN, MVT::v8i64, 1},
2106       {ISD::SMIN, MVT::v16i32, 1},
2107       {ISD::UMIN, MVT::v16i32, 1},
2108   };
2109 
2110   if (IsPairwise) {
2111     if (ST->hasAVX512())
2112       if (const auto *Entry = CostTableLookup(AVX512CostTblPairWise, ISD, MTy))
2113         return LT.first * Entry->Cost;
2114 
2115     if (ST->hasAVX2())
2116       if (const auto *Entry = CostTableLookup(AVX2CostTblPairWise, ISD, MTy))
2117         return LT.first * Entry->Cost;
2118 
2119     if (ST->hasAVX())
2120       if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy))
2121         return LT.first * Entry->Cost;
2122 
2123     if (ST->hasSSE42())
2124       if (const auto *Entry = CostTableLookup(SSE42CostTblPairWise, ISD, MTy))
2125         return LT.first * Entry->Cost;
2126   } else {
2127     if (ST->hasAVX512())
2128       if (const auto *Entry =
2129               CostTableLookup(AVX512CostTblNoPairWise, ISD, MTy))
2130         return LT.first * Entry->Cost;
2131 
2132     if (ST->hasAVX2())
2133       if (const auto *Entry = CostTableLookup(AVX2CostTblNoPairWise, ISD, MTy))
2134         return LT.first * Entry->Cost;
2135 
2136     if (ST->hasAVX())
2137       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
2138         return LT.first * Entry->Cost;
2139 
2140     if (ST->hasSSE42())
2141       if (const auto *Entry = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy))
2142         return LT.first * Entry->Cost;
2143   }
2144 
2145   return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsPairwise, IsUnsigned);
2146 }
2147 
2148 /// \brief Calculate the cost of materializing a 64-bit value. This helper
2149 /// method might only calculate a fraction of a larger immediate. Therefore it
2150 /// is valid to return a cost of ZERO.
2151 int X86TTIImpl::getIntImmCost(int64_t Val) {
2152   if (Val == 0)
2153     return TTI::TCC_Free;
2154 
2155   if (isInt<32>(Val))
2156     return TTI::TCC_Basic;
2157 
2158   return 2 * TTI::TCC_Basic;
2159 }
2160 
2161 int X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) {
2162   assert(Ty->isIntegerTy());
2163 
2164   unsigned BitSize = Ty->getPrimitiveSizeInBits();
2165   if (BitSize == 0)
2166     return ~0U;
2167 
2168   // Never hoist constants larger than 128bit, because this might lead to
2169   // incorrect code generation or assertions in codegen.
2170   // Fixme: Create a cost model for types larger than i128 once the codegen
2171   // issues have been fixed.
2172   if (BitSize > 128)
2173     return TTI::TCC_Free;
2174 
2175   if (Imm == 0)
2176     return TTI::TCC_Free;
2177 
2178   // Sign-extend all constants to a multiple of 64-bit.
2179   APInt ImmVal = Imm;
2180   if (BitSize & 0x3f)
2181     ImmVal = Imm.sext((BitSize + 63) & ~0x3fU);
2182 
2183   // Split the constant into 64-bit chunks and calculate the cost for each
2184   // chunk.
2185   int Cost = 0;
2186   for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
2187     APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
2188     int64_t Val = Tmp.getSExtValue();
2189     Cost += getIntImmCost(Val);
2190   }
2191   // We need at least one instruction to materialize the constant.
2192   return std::max(1, Cost);
2193 }
2194 
2195 int X86TTIImpl::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
2196                               Type *Ty) {
2197   assert(Ty->isIntegerTy());
2198 
2199   unsigned BitSize = Ty->getPrimitiveSizeInBits();
2200   // There is no cost model for constants with a bit size of 0. Return TCC_Free
2201   // here, so that constant hoisting will ignore this constant.
2202   if (BitSize == 0)
2203     return TTI::TCC_Free;
2204 
2205   unsigned ImmIdx = ~0U;
2206   switch (Opcode) {
2207   default:
2208     return TTI::TCC_Free;
2209   case Instruction::GetElementPtr:
2210     // Always hoist the base address of a GetElementPtr. This prevents the
2211     // creation of new constants for every base constant that gets constant
2212     // folded with the offset.
2213     if (Idx == 0)
2214       return 2 * TTI::TCC_Basic;
2215     return TTI::TCC_Free;
2216   case Instruction::Store:
2217     ImmIdx = 0;
2218     break;
2219   case Instruction::ICmp:
2220     // This is an imperfect hack to prevent constant hoisting of
2221     // compares that might be trying to check if a 64-bit value fits in
2222     // 32-bits. The backend can optimize these cases using a right shift by 32.
2223     // Ideally we would check the compare predicate here. There also other
2224     // similar immediates the backend can use shifts for.
2225     if (Idx == 1 && Imm.getBitWidth() == 64) {
2226       uint64_t ImmVal = Imm.getZExtValue();
2227       if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff)
2228         return TTI::TCC_Free;
2229     }
2230     ImmIdx = 1;
2231     break;
2232   case Instruction::And:
2233     // We support 64-bit ANDs with immediates with 32-bits of leading zeroes
2234     // by using a 32-bit operation with implicit zero extension. Detect such
2235     // immediates here as the normal path expects bit 31 to be sign extended.
2236     if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue()))
2237       return TTI::TCC_Free;
2238     LLVM_FALLTHROUGH;
2239   case Instruction::Add:
2240   case Instruction::Sub:
2241   case Instruction::Mul:
2242   case Instruction::UDiv:
2243   case Instruction::SDiv:
2244   case Instruction::URem:
2245   case Instruction::SRem:
2246   case Instruction::Or:
2247   case Instruction::Xor:
2248     ImmIdx = 1;
2249     break;
2250   // Always return TCC_Free for the shift value of a shift instruction.
2251   case Instruction::Shl:
2252   case Instruction::LShr:
2253   case Instruction::AShr:
2254     if (Idx == 1)
2255       return TTI::TCC_Free;
2256     break;
2257   case Instruction::Trunc:
2258   case Instruction::ZExt:
2259   case Instruction::SExt:
2260   case Instruction::IntToPtr:
2261   case Instruction::PtrToInt:
2262   case Instruction::BitCast:
2263   case Instruction::PHI:
2264   case Instruction::Call:
2265   case Instruction::Select:
2266   case Instruction::Ret:
2267   case Instruction::Load:
2268     break;
2269   }
2270 
2271   if (Idx == ImmIdx) {
2272     int NumConstants = (BitSize + 63) / 64;
2273     int Cost = X86TTIImpl::getIntImmCost(Imm, Ty);
2274     return (Cost <= NumConstants * TTI::TCC_Basic)
2275                ? static_cast<int>(TTI::TCC_Free)
2276                : Cost;
2277   }
2278 
2279   return X86TTIImpl::getIntImmCost(Imm, Ty);
2280 }
2281 
2282 int X86TTIImpl::getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
2283                               Type *Ty) {
2284   assert(Ty->isIntegerTy());
2285 
2286   unsigned BitSize = Ty->getPrimitiveSizeInBits();
2287   // There is no cost model for constants with a bit size of 0. Return TCC_Free
2288   // here, so that constant hoisting will ignore this constant.
2289   if (BitSize == 0)
2290     return TTI::TCC_Free;
2291 
2292   switch (IID) {
2293   default:
2294     return TTI::TCC_Free;
2295   case Intrinsic::sadd_with_overflow:
2296   case Intrinsic::uadd_with_overflow:
2297   case Intrinsic::ssub_with_overflow:
2298   case Intrinsic::usub_with_overflow:
2299   case Intrinsic::smul_with_overflow:
2300   case Intrinsic::umul_with_overflow:
2301     if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
2302       return TTI::TCC_Free;
2303     break;
2304   case Intrinsic::experimental_stackmap:
2305     if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
2306       return TTI::TCC_Free;
2307     break;
2308   case Intrinsic::experimental_patchpoint_void:
2309   case Intrinsic::experimental_patchpoint_i64:
2310     if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
2311       return TTI::TCC_Free;
2312     break;
2313   }
2314   return X86TTIImpl::getIntImmCost(Imm, Ty);
2315 }
2316 
2317 unsigned X86TTIImpl::getUserCost(const User *U,
2318                                  ArrayRef<const Value *> Operands) {
2319   if (isa<StoreInst>(U)) {
2320     Value *Ptr = U->getOperand(1);
2321     // Store instruction with index and scale costs 2 Uops.
2322     // Check the preceding GEP to identify non-const indices.
2323     if (auto GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
2324       if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); }))
2325         return TTI::TCC_Basic * 2;
2326     }
2327     return TTI::TCC_Basic;
2328   }
2329   return BaseT::getUserCost(U, Operands);
2330 }
2331 
2332 // Return an average cost of Gather / Scatter instruction, maybe improved later
2333 int X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy, Value *Ptr,
2334                                 unsigned Alignment, unsigned AddressSpace) {
2335 
2336   assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost");
2337   unsigned VF = SrcVTy->getVectorNumElements();
2338 
2339   // Try to reduce index size from 64 bit (default for GEP)
2340   // to 32. It is essential for VF 16. If the index can't be reduced to 32, the
2341   // operation will use 16 x 64 indices which do not fit in a zmm and needs
2342   // to split. Also check that the base pointer is the same for all lanes,
2343   // and that there's at most one variable index.
2344   auto getIndexSizeInBits = [](Value *Ptr, const DataLayout& DL) {
2345     unsigned IndexSize = DL.getPointerSizeInBits();
2346     GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
2347     if (IndexSize < 64 || !GEP)
2348       return IndexSize;
2349 
2350     unsigned NumOfVarIndices = 0;
2351     Value *Ptrs = GEP->getPointerOperand();
2352     if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs))
2353       return IndexSize;
2354     for (unsigned i = 1; i < GEP->getNumOperands(); ++i) {
2355       if (isa<Constant>(GEP->getOperand(i)))
2356         continue;
2357       Type *IndxTy = GEP->getOperand(i)->getType();
2358       if (IndxTy->isVectorTy())
2359         IndxTy = IndxTy->getVectorElementType();
2360       if ((IndxTy->getPrimitiveSizeInBits() == 64 &&
2361           !isa<SExtInst>(GEP->getOperand(i))) ||
2362          ++NumOfVarIndices > 1)
2363         return IndexSize; // 64
2364     }
2365     return (unsigned)32;
2366   };
2367 
2368 
2369   // Trying to reduce IndexSize to 32 bits for vector 16.
2370   // By default the IndexSize is equal to pointer size.
2371   unsigned IndexSize = (ST->hasAVX512() && VF >= 16)
2372                            ? getIndexSizeInBits(Ptr, DL)
2373                            : DL.getPointerSizeInBits();
2374 
2375   Type *IndexVTy = VectorType::get(IntegerType::get(SrcVTy->getContext(),
2376                                                     IndexSize), VF);
2377   std::pair<int, MVT> IdxsLT = TLI->getTypeLegalizationCost(DL, IndexVTy);
2378   std::pair<int, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, SrcVTy);
2379   int SplitFactor = std::max(IdxsLT.first, SrcLT.first);
2380   if (SplitFactor > 1) {
2381     // Handle splitting of vector of pointers
2382     Type *SplitSrcTy = VectorType::get(SrcVTy->getScalarType(), VF / SplitFactor);
2383     return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment,
2384                                          AddressSpace);
2385   }
2386 
2387   // The gather / scatter cost is given by Intel architects. It is a rough
2388   // number since we are looking at one instruction in a time.
2389   const int GSOverhead = (Opcode == Instruction::Load)
2390                              ? ST->getGatherOverhead()
2391                              : ST->getScatterOverhead();
2392   return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
2393                                            Alignment, AddressSpace);
2394 }
2395 
2396 /// Return the cost of full scalarization of gather / scatter operation.
2397 ///
2398 /// Opcode - Load or Store instruction.
2399 /// SrcVTy - The type of the data vector that should be gathered or scattered.
2400 /// VariableMask - The mask is non-constant at compile time.
2401 /// Alignment - Alignment for one element.
2402 /// AddressSpace - pointer[s] address space.
2403 ///
2404 int X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy,
2405                                 bool VariableMask, unsigned Alignment,
2406                                 unsigned AddressSpace) {
2407   unsigned VF = SrcVTy->getVectorNumElements();
2408 
2409   int MaskUnpackCost = 0;
2410   if (VariableMask) {
2411     VectorType *MaskTy =
2412       VectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF);
2413     MaskUnpackCost = getScalarizationOverhead(MaskTy, false, true);
2414     int ScalarCompareCost =
2415       getCmpSelInstrCost(Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()),
2416                          nullptr);
2417     int BranchCost = getCFInstrCost(Instruction::Br);
2418     MaskUnpackCost += VF * (BranchCost + ScalarCompareCost);
2419   }
2420 
2421   // The cost of the scalar loads/stores.
2422   int MemoryOpCost = VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
2423                                           Alignment, AddressSpace);
2424 
2425   int InsertExtractCost = 0;
2426   if (Opcode == Instruction::Load)
2427     for (unsigned i = 0; i < VF; ++i)
2428       // Add the cost of inserting each scalar load into the vector
2429       InsertExtractCost +=
2430         getVectorInstrCost(Instruction::InsertElement, SrcVTy, i);
2431   else
2432     for (unsigned i = 0; i < VF; ++i)
2433       // Add the cost of extracting each element out of the data vector
2434       InsertExtractCost +=
2435         getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i);
2436 
2437   return MemoryOpCost + MaskUnpackCost + InsertExtractCost;
2438 }
2439 
2440 /// Calculate the cost of Gather / Scatter operation
2441 int X86TTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *SrcVTy,
2442                                        Value *Ptr, bool VariableMask,
2443                                        unsigned Alignment) {
2444   assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter");
2445   unsigned VF = SrcVTy->getVectorNumElements();
2446   PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
2447   if (!PtrTy && Ptr->getType()->isVectorTy())
2448     PtrTy = dyn_cast<PointerType>(Ptr->getType()->getVectorElementType());
2449   assert(PtrTy && "Unexpected type for Ptr argument");
2450   unsigned AddressSpace = PtrTy->getAddressSpace();
2451 
2452   bool Scalarize = false;
2453   if ((Opcode == Instruction::Load && !isLegalMaskedGather(SrcVTy)) ||
2454       (Opcode == Instruction::Store && !isLegalMaskedScatter(SrcVTy)))
2455     Scalarize = true;
2456   // Gather / Scatter for vector 2 is not profitable on KNL / SKX
2457   // Vector-4 of gather/scatter instruction does not exist on KNL.
2458   // We can extend it to 8 elements, but zeroing upper bits of
2459   // the mask vector will add more instructions. Right now we give the scalar
2460   // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter instruction
2461   // is better in the VariableMask case.
2462   if (ST->hasAVX512() && (VF == 2 || (VF == 4 && !ST->hasVLX())))
2463     Scalarize = true;
2464 
2465   if (Scalarize)
2466     return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment,
2467                            AddressSpace);
2468 
2469   return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace);
2470 }
2471 
2472 bool X86TTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1,
2473                                TargetTransformInfo::LSRCost &C2) {
2474     // X86 specific here are "instruction number 1st priority".
2475     return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost,
2476                     C1.NumIVMuls, C1.NumBaseAdds,
2477                     C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
2478            std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost,
2479                     C2.NumIVMuls, C2.NumBaseAdds,
2480                     C2.ScaleCost, C2.ImmCost, C2.SetupCost);
2481 }
2482 
2483 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy) {
2484   // The backend can't handle a single element vector.
2485   if (isa<VectorType>(DataTy) && DataTy->getVectorNumElements() == 1)
2486     return false;
2487   Type *ScalarTy = DataTy->getScalarType();
2488   int DataWidth = isa<PointerType>(ScalarTy) ?
2489     DL.getPointerSizeInBits() : ScalarTy->getPrimitiveSizeInBits();
2490 
2491   return ((DataWidth == 32 || DataWidth == 64) && ST->hasAVX()) ||
2492          ((DataWidth == 8 || DataWidth == 16) && ST->hasBWI());
2493 }
2494 
2495 bool X86TTIImpl::isLegalMaskedStore(Type *DataType) {
2496   return isLegalMaskedLoad(DataType);
2497 }
2498 
2499 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy) {
2500   // This function is called now in two cases: from the Loop Vectorizer
2501   // and from the Scalarizer.
2502   // When the Loop Vectorizer asks about legality of the feature,
2503   // the vectorization factor is not calculated yet. The Loop Vectorizer
2504   // sends a scalar type and the decision is based on the width of the
2505   // scalar element.
2506   // Later on, the cost model will estimate usage this intrinsic based on
2507   // the vector type.
2508   // The Scalarizer asks again about legality. It sends a vector type.
2509   // In this case we can reject non-power-of-2 vectors.
2510   // We also reject single element vectors as the type legalizer can't
2511   // scalarize it.
2512   if (isa<VectorType>(DataTy)) {
2513     unsigned NumElts = DataTy->getVectorNumElements();
2514     if (NumElts == 1 || !isPowerOf2_32(NumElts))
2515       return false;
2516   }
2517   Type *ScalarTy = DataTy->getScalarType();
2518   int DataWidth = isa<PointerType>(ScalarTy) ?
2519     DL.getPointerSizeInBits() : ScalarTy->getPrimitiveSizeInBits();
2520 
2521   // Some CPUs have better gather performance than others.
2522   // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only
2523   // enable gather with a -march.
2524   return (DataWidth == 32 || DataWidth == 64) &&
2525     (ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2()));
2526 }
2527 
2528 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType) {
2529   // AVX2 doesn't support scatter
2530   if (!ST->hasAVX512())
2531     return false;
2532   return isLegalMaskedGather(DataType);
2533 }
2534 
2535 bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) {
2536   EVT VT = TLI->getValueType(DL, DataType);
2537   return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT);
2538 }
2539 
2540 bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
2541   return false;
2542 }
2543 
2544 bool X86TTIImpl::areInlineCompatible(const Function *Caller,
2545                                      const Function *Callee) const {
2546   const TargetMachine &TM = getTLI()->getTargetMachine();
2547 
2548   // Work this as a subsetting of subtarget features.
2549   const FeatureBitset &CallerBits =
2550       TM.getSubtargetImpl(*Caller)->getFeatureBits();
2551   const FeatureBitset &CalleeBits =
2552       TM.getSubtargetImpl(*Callee)->getFeatureBits();
2553 
2554   // FIXME: This is likely too limiting as it will include subtarget features
2555   // that we might not care about for inlining, but it is conservatively
2556   // correct.
2557   return (CallerBits & CalleeBits) == CalleeBits;
2558 }
2559 
2560 const X86TTIImpl::TTI::MemCmpExpansionOptions *
2561 X86TTIImpl::enableMemCmpExpansion(bool IsZeroCmp) const {
2562   // Only enable vector loads for equality comparison.
2563   // Right now the vector version is not as fast, see #33329.
2564   static const auto ThreeWayOptions = [this]() {
2565     TTI::MemCmpExpansionOptions Options;
2566     if (ST->is64Bit()) {
2567       Options.LoadSizes.push_back(8);
2568     }
2569     Options.LoadSizes.push_back(4);
2570     Options.LoadSizes.push_back(2);
2571     Options.LoadSizes.push_back(1);
2572     return Options;
2573   }();
2574   static const auto EqZeroOptions = [this]() {
2575     TTI::MemCmpExpansionOptions Options;
2576     // TODO: enable AVX512 when the DAG is ready.
2577     // if (ST->hasAVX512()) Options.LoadSizes.push_back(64);
2578     if (ST->hasAVX2()) Options.LoadSizes.push_back(32);
2579     if (ST->hasSSE2()) Options.LoadSizes.push_back(16);
2580     if (ST->is64Bit()) {
2581       Options.LoadSizes.push_back(8);
2582     }
2583     Options.LoadSizes.push_back(4);
2584     Options.LoadSizes.push_back(2);
2585     Options.LoadSizes.push_back(1);
2586     return Options;
2587   }();
2588   return IsZeroCmp ? &EqZeroOptions : &ThreeWayOptions;
2589 }
2590 
2591 bool X86TTIImpl::enableInterleavedAccessVectorization() {
2592   // TODO: We expect this to be beneficial regardless of arch,
2593   // but there are currently some unexplained performance artifacts on Atom.
2594   // As a temporary solution, disable on Atom.
2595   return !(ST->isAtom());
2596 }
2597 
2598 // Get estimation for interleaved load/store operations for AVX2.
2599 // \p Factor is the interleaved-access factor (stride) - number of
2600 // (interleaved) elements in the group.
2601 // \p Indices contains the indices for a strided load: when the
2602 // interleaved load has gaps they indicate which elements are used.
2603 // If Indices is empty (or if the number of indices is equal to the size
2604 // of the interleaved-access as given in \p Factor) the access has no gaps.
2605 //
2606 // As opposed to AVX-512, AVX2 does not have generic shuffles that allow
2607 // computing the cost using a generic formula as a function of generic
2608 // shuffles. We therefore use a lookup table instead, filled according to
2609 // the instruction sequences that codegen currently generates.
2610 int X86TTIImpl::getInterleavedMemoryOpCostAVX2(unsigned Opcode, Type *VecTy,
2611                                                unsigned Factor,
2612                                                ArrayRef<unsigned> Indices,
2613                                                unsigned Alignment,
2614                                                unsigned AddressSpace) {
2615 
2616   // We currently Support only fully-interleaved groups, with no gaps.
2617   // TODO: Support also strided loads (interleaved-groups with gaps).
2618   if (Indices.size() && Indices.size() != Factor)
2619     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2620                                              Alignment, AddressSpace);
2621 
2622   // VecTy for interleave memop is <VF*Factor x Elt>.
2623   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
2624   // VecTy = <12 x i32>.
2625   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
2626 
2627   // This function can be called with VecTy=<6xi128>, Factor=3, in which case
2628   // the VF=2, while v2i128 is an unsupported MVT vector type
2629   // (see MachineValueType.h::getVectorVT()).
2630   if (!LegalVT.isVector())
2631     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2632                                              Alignment, AddressSpace);
2633 
2634   unsigned VF = VecTy->getVectorNumElements() / Factor;
2635   Type *ScalarTy = VecTy->getVectorElementType();
2636 
2637   // Calculate the number of memory operations (NumOfMemOps), required
2638   // for load/store the VecTy.
2639   unsigned VecTySize = DL.getTypeStoreSize(VecTy);
2640   unsigned LegalVTSize = LegalVT.getStoreSize();
2641   unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize;
2642 
2643   // Get the cost of one memory operation.
2644   Type *SingleMemOpTy = VectorType::get(VecTy->getVectorElementType(),
2645                                         LegalVT.getVectorNumElements());
2646   unsigned MemOpCost =
2647       getMemoryOpCost(Opcode, SingleMemOpTy, Alignment, AddressSpace);
2648 
2649   VectorType *VT = VectorType::get(ScalarTy, VF);
2650   EVT ETy = TLI->getValueType(DL, VT);
2651   if (!ETy.isSimple())
2652     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2653                                              Alignment, AddressSpace);
2654 
2655   // TODO: Complete for other data-types and strides.
2656   // Each combination of Stride, ElementTy and VF results in a different
2657   // sequence; The cost tables are therefore accessed with:
2658   // Factor (stride) and VectorType=VFxElemType.
2659   // The Cost accounts only for the shuffle sequence;
2660   // The cost of the loads/stores is accounted for separately.
2661   //
2662   static const CostTblEntry AVX2InterleavedLoadTbl[] = {
2663     { 2, MVT::v4i64, 6 }, //(load 8i64 and) deinterleave into 2 x 4i64
2664     { 2, MVT::v4f64, 6 }, //(load 8f64 and) deinterleave into 2 x 4f64
2665 
2666     { 3, MVT::v2i8,  10 }, //(load 6i8 and)  deinterleave into 3 x 2i8
2667     { 3, MVT::v4i8,  4 },  //(load 12i8 and) deinterleave into 3 x 4i8
2668     { 3, MVT::v8i8,  9 },  //(load 24i8 and) deinterleave into 3 x 8i8
2669     { 3, MVT::v16i8, 11},  //(load 48i8 and) deinterleave into 3 x 16i8
2670     { 3, MVT::v32i8, 13},  //(load 96i8 and) deinterleave into 3 x 32i8
2671     { 3, MVT::v8f32, 17 }, //(load 24f32 and)deinterleave into 3 x 8f32
2672 
2673     { 4, MVT::v2i8,  12 }, //(load 8i8 and)   deinterleave into 4 x 2i8
2674     { 4, MVT::v4i8,  4 },  //(load 16i8 and)  deinterleave into 4 x 4i8
2675     { 4, MVT::v8i8,  20 }, //(load 32i8 and)  deinterleave into 4 x 8i8
2676     { 4, MVT::v16i8, 39 }, //(load 64i8 and)  deinterleave into 4 x 16i8
2677     { 4, MVT::v32i8, 80 }, //(load 128i8 and) deinterleave into 4 x 32i8
2678 
2679     { 8, MVT::v8f32, 40 }  //(load 64f32 and)deinterleave into 8 x 8f32
2680   };
2681 
2682   static const CostTblEntry AVX2InterleavedStoreTbl[] = {
2683     { 2, MVT::v4i64, 6 }, //interleave into 2 x 4i64 into 8i64 (and store)
2684     { 2, MVT::v4f64, 6 }, //interleave into 2 x 4f64 into 8f64 (and store)
2685 
2686     { 3, MVT::v2i8,  7 },  //interleave 3 x 2i8  into 6i8 (and store)
2687     { 3, MVT::v4i8,  8 },  //interleave 3 x 4i8  into 12i8 (and store)
2688     { 3, MVT::v8i8,  11 }, //interleave 3 x 8i8  into 24i8 (and store)
2689     { 3, MVT::v16i8, 11 }, //interleave 3 x 16i8 into 48i8 (and store)
2690     { 3, MVT::v32i8, 13 }, //interleave 3 x 32i8 into 96i8 (and store)
2691 
2692     { 4, MVT::v2i8,  12 }, //interleave 4 x 2i8  into 8i8 (and store)
2693     { 4, MVT::v4i8,  9 },  //interleave 4 x 4i8  into 16i8 (and store)
2694     { 4, MVT::v8i8,  10 }, //interleave 4 x 8i8  into 32i8 (and store)
2695     { 4, MVT::v16i8, 10 }, //interleave 4 x 16i8 into 64i8 (and store)
2696     { 4, MVT::v32i8, 12 }  //interleave 4 x 32i8 into 128i8 (and store)
2697   };
2698 
2699   if (Opcode == Instruction::Load) {
2700     if (const auto *Entry =
2701             CostTableLookup(AVX2InterleavedLoadTbl, Factor, ETy.getSimpleVT()))
2702       return NumOfMemOps * MemOpCost + Entry->Cost;
2703   } else {
2704     assert(Opcode == Instruction::Store &&
2705            "Expected Store Instruction at this  point");
2706     if (const auto *Entry =
2707             CostTableLookup(AVX2InterleavedStoreTbl, Factor, ETy.getSimpleVT()))
2708       return NumOfMemOps * MemOpCost + Entry->Cost;
2709   }
2710 
2711   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2712                                            Alignment, AddressSpace);
2713 }
2714 
2715 // Get estimation for interleaved load/store operations and strided load.
2716 // \p Indices contains indices for strided load.
2717 // \p Factor - the factor of interleaving.
2718 // AVX-512 provides 3-src shuffles that significantly reduces the cost.
2719 int X86TTIImpl::getInterleavedMemoryOpCostAVX512(unsigned Opcode, Type *VecTy,
2720                                                  unsigned Factor,
2721                                                  ArrayRef<unsigned> Indices,
2722                                                  unsigned Alignment,
2723                                                  unsigned AddressSpace) {
2724 
2725   // VecTy for interleave memop is <VF*Factor x Elt>.
2726   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
2727   // VecTy = <12 x i32>.
2728 
2729   // Calculate the number of memory operations (NumOfMemOps), required
2730   // for load/store the VecTy.
2731   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
2732   unsigned VecTySize = DL.getTypeStoreSize(VecTy);
2733   unsigned LegalVTSize = LegalVT.getStoreSize();
2734   unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize;
2735 
2736   // Get the cost of one memory operation.
2737   Type *SingleMemOpTy = VectorType::get(VecTy->getVectorElementType(),
2738                                         LegalVT.getVectorNumElements());
2739   unsigned MemOpCost =
2740       getMemoryOpCost(Opcode, SingleMemOpTy, Alignment, AddressSpace);
2741 
2742   unsigned VF = VecTy->getVectorNumElements() / Factor;
2743   MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF);
2744 
2745   if (Opcode == Instruction::Load) {
2746     // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl)
2747     // contain the cost of the optimized shuffle sequence that the
2748     // X86InterleavedAccess pass will generate.
2749     // The cost of loads and stores are computed separately from the table.
2750 
2751     // X86InterleavedAccess support only the following interleaved-access group.
2752     static const CostTblEntry AVX512InterleavedLoadTbl[] = {
2753         {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8
2754         {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8
2755         {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8
2756     };
2757 
2758     if (const auto *Entry =
2759             CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT))
2760       return NumOfMemOps * MemOpCost + Entry->Cost;
2761     //If an entry does not exist, fallback to the default implementation.
2762 
2763     // Kind of shuffle depends on number of loaded values.
2764     // If we load the entire data in one register, we can use a 1-src shuffle.
2765     // Otherwise, we'll merge 2 sources in each operation.
2766     TTI::ShuffleKind ShuffleKind =
2767         (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc;
2768 
2769     unsigned ShuffleCost =
2770         getShuffleCost(ShuffleKind, SingleMemOpTy, 0, nullptr);
2771 
2772     unsigned NumOfLoadsInInterleaveGrp =
2773         Indices.size() ? Indices.size() : Factor;
2774     Type *ResultTy = VectorType::get(VecTy->getVectorElementType(),
2775                                      VecTy->getVectorNumElements() / Factor);
2776     unsigned NumOfResults =
2777         getTLI()->getTypeLegalizationCost(DL, ResultTy).first *
2778         NumOfLoadsInInterleaveGrp;
2779 
2780     // About a half of the loads may be folded in shuffles when we have only
2781     // one result. If we have more than one result, we do not fold loads at all.
2782     unsigned NumOfUnfoldedLoads =
2783         NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2;
2784 
2785     // Get a number of shuffle operations per result.
2786     unsigned NumOfShufflesPerResult =
2787         std::max((unsigned)1, (unsigned)(NumOfMemOps - 1));
2788 
2789     // The SK_MergeTwoSrc shuffle clobbers one of src operands.
2790     // When we have more than one destination, we need additional instructions
2791     // to keep sources.
2792     unsigned NumOfMoves = 0;
2793     if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc)
2794       NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2;
2795 
2796     int Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost +
2797                NumOfUnfoldedLoads * MemOpCost + NumOfMoves;
2798 
2799     return Cost;
2800   }
2801 
2802   // Store.
2803   assert(Opcode == Instruction::Store &&
2804          "Expected Store Instruction at this  point");
2805   // X86InterleavedAccess support only the following interleaved-access group.
2806   static const CostTblEntry AVX512InterleavedStoreTbl[] = {
2807       {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store)
2808       {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store)
2809       {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store)
2810 
2811       {4, MVT::v8i8, 10},  // interleave 4 x 8i8  into 32i8  (and store)
2812       {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8  (and store)
2813       {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store)
2814       {4, MVT::v64i8, 24}  // interleave 4 x 32i8 into 256i8 (and store)
2815   };
2816 
2817   if (const auto *Entry =
2818           CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT))
2819     return NumOfMemOps * MemOpCost + Entry->Cost;
2820   //If an entry does not exist, fallback to the default implementation.
2821 
2822   // There is no strided stores meanwhile. And store can't be folded in
2823   // shuffle.
2824   unsigned NumOfSources = Factor; // The number of values to be merged.
2825   unsigned ShuffleCost =
2826       getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, 0, nullptr);
2827   unsigned NumOfShufflesPerStore = NumOfSources - 1;
2828 
2829   // The SK_MergeTwoSrc shuffle clobbers one of src operands.
2830   // We need additional instructions to keep sources.
2831   unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2;
2832   int Cost = NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) +
2833              NumOfMoves;
2834   return Cost;
2835 }
2836 
2837 int X86TTIImpl::getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
2838                                            unsigned Factor,
2839                                            ArrayRef<unsigned> Indices,
2840                                            unsigned Alignment,
2841                                            unsigned AddressSpace) {
2842   auto isSupportedOnAVX512 = [](Type *VecTy, bool HasBW) {
2843     Type *EltTy = VecTy->getVectorElementType();
2844     if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) ||
2845         EltTy->isIntegerTy(32) || EltTy->isPointerTy())
2846       return true;
2847     if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8))
2848       return HasBW;
2849     return false;
2850   };
2851   if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI()))
2852     return getInterleavedMemoryOpCostAVX512(Opcode, VecTy, Factor, Indices,
2853                                             Alignment, AddressSpace);
2854   if (ST->hasAVX2())
2855     return getInterleavedMemoryOpCostAVX2(Opcode, VecTy, Factor, Indices,
2856                                           Alignment, AddressSpace);
2857 
2858   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2859                                            Alignment, AddressSpace);
2860 }
2861