1 //===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 /// \file
10 /// This file implements a TargetTransformInfo analysis pass specific to the
11 /// X86 target machine. It uses the target's detailed information to provide
12 /// more precise answers to certain TTI queries, while letting the target
13 /// independent and default TTI implementations handle the rest.
14 ///
15 //===----------------------------------------------------------------------===//
16 /// About Cost Model numbers used below it's necessary to say the following:
17 /// the numbers correspond to some "generic" X86 CPU instead of usage of
18 /// concrete CPU model. Usually the numbers correspond to CPU where the feature
19 /// apeared at the first time. For example, if we do Subtarget.hasSSE42() in
20 /// the lookups below the cost is based on Nehalem as that was the first CPU
21 /// to support that feature level and thus has most likely the worst case cost.
22 /// Some examples of other technologies/CPUs:
23 ///   SSE 3   - Pentium4 / Athlon64
24 ///   SSE 4.1 - Penryn
25 ///   SSE 4.2 - Nehalem
26 ///   AVX     - Sandy Bridge
27 ///   AVX2    - Haswell
28 ///   AVX-512 - Xeon Phi / Skylake
29 /// And some examples of instruction target dependent costs (latency)
30 ///                   divss     sqrtss          rsqrtss
31 ///   AMD K7            11-16     19              3
32 ///   Piledriver        9-24      13-15           5
33 ///   Jaguar            14        16              2
34 ///   Pentium II,III    18        30              2
35 ///   Nehalem           7-14      7-18            3
36 ///   Haswell           10-13     11              5
37 /// TODO: Develop and implement  the target dependent cost model and
38 /// specialize cost numbers for different Cost Model Targets such as throughput,
39 /// code size, latency and uop count.
40 //===----------------------------------------------------------------------===//
41 
42 #include "X86TargetTransformInfo.h"
43 #include "llvm/Analysis/TargetTransformInfo.h"
44 #include "llvm/CodeGen/BasicTTIImpl.h"
45 #include "llvm/CodeGen/CostTable.h"
46 #include "llvm/CodeGen/TargetLowering.h"
47 #include "llvm/IR/IntrinsicInst.h"
48 #include "llvm/Support/Debug.h"
49 
50 using namespace llvm;
51 
52 #define DEBUG_TYPE "x86tti"
53 
54 //===----------------------------------------------------------------------===//
55 //
56 // X86 cost model.
57 //
58 //===----------------------------------------------------------------------===//
59 
60 TargetTransformInfo::PopcntSupportKind
61 X86TTIImpl::getPopcntSupport(unsigned TyWidth) {
62   assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
63   // TODO: Currently the __builtin_popcount() implementation using SSE3
64   //   instructions is inefficient. Once the problem is fixed, we should
65   //   call ST->hasSSE3() instead of ST->hasPOPCNT().
66   return ST->hasPOPCNT() ? TTI::PSK_FastHardware : TTI::PSK_Software;
67 }
68 
69 llvm::Optional<unsigned> X86TTIImpl::getCacheSize(
70   TargetTransformInfo::CacheLevel Level) const {
71   switch (Level) {
72   case TargetTransformInfo::CacheLevel::L1D:
73     //   - Penryn
74     //   - Nehalem
75     //   - Westmere
76     //   - Sandy Bridge
77     //   - Ivy Bridge
78     //   - Haswell
79     //   - Broadwell
80     //   - Skylake
81     //   - Kabylake
82     return 32 * 1024;  //  32 KByte
83   case TargetTransformInfo::CacheLevel::L2D:
84     //   - Penryn
85     //   - Nehalem
86     //   - Westmere
87     //   - Sandy Bridge
88     //   - Ivy Bridge
89     //   - Haswell
90     //   - Broadwell
91     //   - Skylake
92     //   - Kabylake
93     return 256 * 1024; // 256 KByte
94   }
95 
96   llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
97 }
98 
99 llvm::Optional<unsigned> X86TTIImpl::getCacheAssociativity(
100   TargetTransformInfo::CacheLevel Level) const {
101   //   - Penryn
102   //   - Nehalem
103   //   - Westmere
104   //   - Sandy Bridge
105   //   - Ivy Bridge
106   //   - Haswell
107   //   - Broadwell
108   //   - Skylake
109   //   - Kabylake
110   switch (Level) {
111   case TargetTransformInfo::CacheLevel::L1D:
112     LLVM_FALLTHROUGH;
113   case TargetTransformInfo::CacheLevel::L2D:
114     return 8;
115   }
116 
117   llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
118 }
119 
120 unsigned X86TTIImpl::getNumberOfRegisters(bool Vector) {
121   if (Vector && !ST->hasSSE1())
122     return 0;
123 
124   if (ST->is64Bit()) {
125     if (Vector && ST->hasAVX512())
126       return 32;
127     return 16;
128   }
129   return 8;
130 }
131 
132 unsigned X86TTIImpl::getRegisterBitWidth(bool Vector) const {
133   unsigned PreferVectorWidth = ST->getPreferVectorWidth();
134   if (Vector) {
135     if (ST->hasAVX512() && PreferVectorWidth >= 512)
136       return 512;
137     if (ST->hasAVX() && PreferVectorWidth >= 256)
138       return 256;
139     if (ST->hasSSE1() && PreferVectorWidth >= 128)
140       return 128;
141     return 0;
142   }
143 
144   if (ST->is64Bit())
145     return 64;
146 
147   return 32;
148 }
149 
150 unsigned X86TTIImpl::getLoadStoreVecRegBitWidth(unsigned) const {
151   return getRegisterBitWidth(true);
152 }
153 
154 unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) {
155   // If the loop will not be vectorized, don't interleave the loop.
156   // Let regular unroll to unroll the loop, which saves the overflow
157   // check and memory check cost.
158   if (VF == 1)
159     return 1;
160 
161   if (ST->isAtom())
162     return 1;
163 
164   // Sandybridge and Haswell have multiple execution ports and pipelined
165   // vector units.
166   if (ST->hasAVX())
167     return 4;
168 
169   return 2;
170 }
171 
172 int X86TTIImpl::getArithmeticInstrCost(
173     unsigned Opcode, Type *Ty,
174     TTI::OperandValueKind Op1Info, TTI::OperandValueKind Op2Info,
175     TTI::OperandValueProperties Opd1PropInfo,
176     TTI::OperandValueProperties Opd2PropInfo,
177     ArrayRef<const Value *> Args) {
178   // Legalize the type.
179   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
180 
181   int ISD = TLI->InstructionOpcodeToISD(Opcode);
182   assert(ISD && "Invalid opcode");
183 
184   static const CostTblEntry SLMCostTable[] = {
185     { ISD::MUL,  MVT::v4i32, 11 }, // pmulld
186     { ISD::MUL,  MVT::v8i16, 2  }, // pmullw
187     { ISD::MUL,  MVT::v16i8, 14 }, // extend/pmullw/trunc sequence.
188     { ISD::FMUL, MVT::f64,   2  }, // mulsd
189     { ISD::FMUL, MVT::v2f64, 4  }, // mulpd
190     { ISD::FMUL, MVT::v4f32, 2  }, // mulps
191     { ISD::FDIV, MVT::f32,   17 }, // divss
192     { ISD::FDIV, MVT::v4f32, 39 }, // divps
193     { ISD::FDIV, MVT::f64,   32 }, // divsd
194     { ISD::FDIV, MVT::v2f64, 69 }, // divpd
195     { ISD::FADD, MVT::v2f64, 2  }, // addpd
196     { ISD::FSUB, MVT::v2f64, 2  }, // subpd
197     // v2i64/v4i64 mul is custom lowered as a series of long:
198     // multiplies(3), shifts(3) and adds(2)
199     // slm muldq version throughput is 2 and addq throughput 4
200     // thus: 3X2 (muldq throughput) + 3X1 (shift throughput) +
201     //       3X4 (addq throughput) = 17
202     { ISD::MUL,  MVT::v2i64, 17 },
203     // slm addq\subq throughput is 4
204     { ISD::ADD,  MVT::v2i64, 4  },
205     { ISD::SUB,  MVT::v2i64, 4  },
206   };
207 
208   if (ST->isSLM()) {
209     if (Args.size() == 2 && ISD == ISD::MUL && LT.second == MVT::v4i32) {
210       // Check if the operands can be shrinked into a smaller datatype.
211       bool Op1Signed = false;
212       unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed);
213       bool Op2Signed = false;
214       unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed);
215 
216       bool signedMode = Op1Signed | Op2Signed;
217       unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize);
218 
219       if (OpMinSize <= 7)
220         return LT.first * 3; // pmullw/sext
221       if (!signedMode && OpMinSize <= 8)
222         return LT.first * 3; // pmullw/zext
223       if (OpMinSize <= 15)
224         return LT.first * 5; // pmullw/pmulhw/pshuf
225       if (!signedMode && OpMinSize <= 16)
226         return LT.first * 5; // pmullw/pmulhw/pshuf
227     }
228     if (const auto *Entry = CostTableLookup(SLMCostTable, ISD,
229                                             LT.second)) {
230       return LT.first * Entry->Cost;
231     }
232   }
233 
234   if (ISD == ISD::SDIV &&
235       Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
236       Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
237     // On X86, vector signed division by constants power-of-two are
238     // normally expanded to the sequence SRA + SRL + ADD + SRA.
239     // The OperandValue properties many not be same as that of previous
240     // operation;conservatively assume OP_None.
241     int Cost = 2 * getArithmeticInstrCost(Instruction::AShr, Ty, Op1Info,
242                                           Op2Info, TargetTransformInfo::OP_None,
243                                           TargetTransformInfo::OP_None);
244     Cost += getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info,
245                                    TargetTransformInfo::OP_None,
246                                    TargetTransformInfo::OP_None);
247     Cost += getArithmeticInstrCost(Instruction::Add, Ty, Op1Info, Op2Info,
248                                    TargetTransformInfo::OP_None,
249                                    TargetTransformInfo::OP_None);
250 
251     return Cost;
252   }
253 
254   static const CostTblEntry AVX512BWUniformConstCostTable[] = {
255     { ISD::SHL,  MVT::v64i8,   2 }, // psllw + pand.
256     { ISD::SRL,  MVT::v64i8,   2 }, // psrlw + pand.
257     { ISD::SRA,  MVT::v64i8,   4 }, // psrlw, pand, pxor, psubb.
258 
259     { ISD::SDIV, MVT::v32i16,  6 }, // vpmulhw sequence
260     { ISD::UDIV, MVT::v32i16,  6 }, // vpmulhuw sequence
261   };
262 
263   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
264       ST->hasBWI()) {
265     if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD,
266                                             LT.second))
267       return LT.first * Entry->Cost;
268   }
269 
270   static const CostTblEntry AVX512UniformConstCostTable[] = {
271     { ISD::SRA,  MVT::v2i64,   1 },
272     { ISD::SRA,  MVT::v4i64,   1 },
273     { ISD::SRA,  MVT::v8i64,   1 },
274 
275     { ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence
276     { ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence
277   };
278 
279   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
280       ST->hasAVX512()) {
281     if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD,
282                                             LT.second))
283       return LT.first * Entry->Cost;
284   }
285 
286   static const CostTblEntry AVX2UniformConstCostTable[] = {
287     { ISD::SHL,  MVT::v32i8,   2 }, // psllw + pand.
288     { ISD::SRL,  MVT::v32i8,   2 }, // psrlw + pand.
289     { ISD::SRA,  MVT::v32i8,   4 }, // psrlw, pand, pxor, psubb.
290 
291     { ISD::SRA,  MVT::v4i64,   4 }, // 2 x psrad + shuffle.
292 
293     { ISD::SDIV, MVT::v16i16,  6 }, // vpmulhw sequence
294     { ISD::UDIV, MVT::v16i16,  6 }, // vpmulhuw sequence
295     { ISD::SDIV, MVT::v8i32,  15 }, // vpmuldq sequence
296     { ISD::UDIV, MVT::v8i32,  15 }, // vpmuludq sequence
297   };
298 
299   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
300       ST->hasAVX2()) {
301     if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD,
302                                             LT.second))
303       return LT.first * Entry->Cost;
304   }
305 
306   static const CostTblEntry SSE2UniformConstCostTable[] = {
307     { ISD::SHL,  MVT::v16i8,     2 }, // psllw + pand.
308     { ISD::SRL,  MVT::v16i8,     2 }, // psrlw + pand.
309     { ISD::SRA,  MVT::v16i8,     4 }, // psrlw, pand, pxor, psubb.
310 
311     { ISD::SHL,  MVT::v32i8,   4+2 }, // 2*(psllw + pand) + split.
312     { ISD::SRL,  MVT::v32i8,   4+2 }, // 2*(psrlw + pand) + split.
313     { ISD::SRA,  MVT::v32i8,   8+2 }, // 2*(psrlw, pand, pxor, psubb) + split.
314 
315     { ISD::SDIV, MVT::v16i16, 12+2 }, // 2*pmulhw sequence + split.
316     { ISD::SDIV, MVT::v8i16,     6 }, // pmulhw sequence
317     { ISD::UDIV, MVT::v16i16, 12+2 }, // 2*pmulhuw sequence + split.
318     { ISD::UDIV, MVT::v8i16,     6 }, // pmulhuw sequence
319     { ISD::SDIV, MVT::v8i32,  38+2 }, // 2*pmuludq sequence + split.
320     { ISD::SDIV, MVT::v4i32,    19 }, // pmuludq sequence
321     { ISD::UDIV, MVT::v8i32,  30+2 }, // 2*pmuludq sequence + split.
322     { ISD::UDIV, MVT::v4i32,    15 }, // pmuludq sequence
323   };
324 
325   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
326       ST->hasSSE2()) {
327     // pmuldq sequence.
328     if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX())
329       return LT.first * 32;
330     if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
331       return LT.first * 15;
332 
333     // XOP has faster vXi8 shifts.
334     if ((ISD != ISD::SHL && ISD != ISD::SRL && ISD != ISD::SRA) ||
335         !ST->hasXOP())
336       if (const auto *Entry =
337               CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second))
338         return LT.first * Entry->Cost;
339   }
340 
341   static const CostTblEntry AVX2UniformCostTable[] = {
342     // Uniform splats are cheaper for the following instructions.
343     { ISD::SHL,  MVT::v16i16, 1 }, // psllw.
344     { ISD::SRL,  MVT::v16i16, 1 }, // psrlw.
345     { ISD::SRA,  MVT::v16i16, 1 }, // psraw.
346   };
347 
348   if (ST->hasAVX2() &&
349       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
350        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
351     if (const auto *Entry =
352             CostTableLookup(AVX2UniformCostTable, ISD, LT.second))
353       return LT.first * Entry->Cost;
354   }
355 
356   static const CostTblEntry SSE2UniformCostTable[] = {
357     // Uniform splats are cheaper for the following instructions.
358     { ISD::SHL,  MVT::v8i16,  1 }, // psllw.
359     { ISD::SHL,  MVT::v4i32,  1 }, // pslld
360     { ISD::SHL,  MVT::v2i64,  1 }, // psllq.
361 
362     { ISD::SRL,  MVT::v8i16,  1 }, // psrlw.
363     { ISD::SRL,  MVT::v4i32,  1 }, // psrld.
364     { ISD::SRL,  MVT::v2i64,  1 }, // psrlq.
365 
366     { ISD::SRA,  MVT::v8i16,  1 }, // psraw.
367     { ISD::SRA,  MVT::v4i32,  1 }, // psrad.
368   };
369 
370   if (ST->hasSSE2() &&
371       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
372        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
373     if (const auto *Entry =
374             CostTableLookup(SSE2UniformCostTable, ISD, LT.second))
375       return LT.first * Entry->Cost;
376   }
377 
378   static const CostTblEntry AVX512DQCostTable[] = {
379     { ISD::MUL,  MVT::v2i64, 1 },
380     { ISD::MUL,  MVT::v4i64, 1 },
381     { ISD::MUL,  MVT::v8i64, 1 }
382   };
383 
384   // Look for AVX512DQ lowering tricks for custom cases.
385   if (ST->hasDQI())
386     if (const auto *Entry = CostTableLookup(AVX512DQCostTable, ISD, LT.second))
387       return LT.first * Entry->Cost;
388 
389   static const CostTblEntry AVX512BWCostTable[] = {
390     { ISD::SHL,   MVT::v8i16,      1 }, // vpsllvw
391     { ISD::SRL,   MVT::v8i16,      1 }, // vpsrlvw
392     { ISD::SRA,   MVT::v8i16,      1 }, // vpsravw
393 
394     { ISD::SHL,   MVT::v16i16,     1 }, // vpsllvw
395     { ISD::SRL,   MVT::v16i16,     1 }, // vpsrlvw
396     { ISD::SRA,   MVT::v16i16,     1 }, // vpsravw
397 
398     { ISD::SHL,   MVT::v32i16,     1 }, // vpsllvw
399     { ISD::SRL,   MVT::v32i16,     1 }, // vpsrlvw
400     { ISD::SRA,   MVT::v32i16,     1 }, // vpsravw
401 
402     { ISD::SHL,   MVT::v64i8,     11 }, // vpblendvb sequence.
403     { ISD::SRL,   MVT::v64i8,     11 }, // vpblendvb sequence.
404     { ISD::SRA,   MVT::v64i8,     24 }, // vpblendvb sequence.
405 
406     { ISD::MUL,   MVT::v64i8,     11 }, // extend/pmullw/trunc sequence.
407     { ISD::MUL,   MVT::v32i8,      4 }, // extend/pmullw/trunc sequence.
408     { ISD::MUL,   MVT::v16i8,      4 }, // extend/pmullw/trunc sequence.
409 
410     // Vectorizing division is a bad idea. See the SSE2 table for more comments.
411     { ISD::SDIV,  MVT::v64i8,  64*20 },
412     { ISD::SDIV,  MVT::v32i16, 32*20 },
413     { ISD::UDIV,  MVT::v64i8,  64*20 },
414     { ISD::UDIV,  MVT::v32i16, 32*20 }
415   };
416 
417   // Look for AVX512BW lowering tricks for custom cases.
418   if (ST->hasBWI())
419     if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD, LT.second))
420       return LT.first * Entry->Cost;
421 
422   static const CostTblEntry AVX512CostTable[] = {
423     { ISD::SHL,     MVT::v16i32,     1 },
424     { ISD::SRL,     MVT::v16i32,     1 },
425     { ISD::SRA,     MVT::v16i32,     1 },
426 
427     { ISD::SHL,     MVT::v8i64,      1 },
428     { ISD::SRL,     MVT::v8i64,      1 },
429 
430     { ISD::SRA,     MVT::v2i64,      1 },
431     { ISD::SRA,     MVT::v4i64,      1 },
432     { ISD::SRA,     MVT::v8i64,      1 },
433 
434     { ISD::MUL,     MVT::v32i8,     13 }, // extend/pmullw/trunc sequence.
435     { ISD::MUL,     MVT::v16i8,      5 }, // extend/pmullw/trunc sequence.
436     { ISD::MUL,     MVT::v16i32,     1 }, // pmulld (Skylake from agner.org)
437     { ISD::MUL,     MVT::v8i32,      1 }, // pmulld (Skylake from agner.org)
438     { ISD::MUL,     MVT::v4i32,      1 }, // pmulld (Skylake from agner.org)
439     { ISD::MUL,     MVT::v8i64,      8 }, // 3*pmuludq/3*shift/2*add
440 
441     // Vectorizing division is a bad idea. See the SSE2 table for more comments.
442     { ISD::SDIV,    MVT::v16i32, 16*20 },
443     { ISD::SDIV,    MVT::v8i64,   8*20 },
444     { ISD::UDIV,    MVT::v16i32, 16*20 },
445     { ISD::UDIV,    MVT::v8i64,   8*20 }
446   };
447 
448   if (ST->hasAVX512())
449     if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second))
450       return LT.first * Entry->Cost;
451 
452   static const CostTblEntry AVX2ShiftCostTable[] = {
453     // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
454     // customize them to detect the cases where shift amount is a scalar one.
455     { ISD::SHL,     MVT::v4i32,    1 },
456     { ISD::SRL,     MVT::v4i32,    1 },
457     { ISD::SRA,     MVT::v4i32,    1 },
458     { ISD::SHL,     MVT::v8i32,    1 },
459     { ISD::SRL,     MVT::v8i32,    1 },
460     { ISD::SRA,     MVT::v8i32,    1 },
461     { ISD::SHL,     MVT::v2i64,    1 },
462     { ISD::SRL,     MVT::v2i64,    1 },
463     { ISD::SHL,     MVT::v4i64,    1 },
464     { ISD::SRL,     MVT::v4i64,    1 },
465   };
466 
467   // Look for AVX2 lowering tricks.
468   if (ST->hasAVX2()) {
469     if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
470         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
471          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
472       // On AVX2, a packed v16i16 shift left by a constant build_vector
473       // is lowered into a vector multiply (vpmullw).
474       return LT.first;
475 
476     if (const auto *Entry = CostTableLookup(AVX2ShiftCostTable, ISD, LT.second))
477       return LT.first * Entry->Cost;
478   }
479 
480   static const CostTblEntry XOPShiftCostTable[] = {
481     // 128bit shifts take 1cy, but right shifts require negation beforehand.
482     { ISD::SHL,     MVT::v16i8,    1 },
483     { ISD::SRL,     MVT::v16i8,    2 },
484     { ISD::SRA,     MVT::v16i8,    2 },
485     { ISD::SHL,     MVT::v8i16,    1 },
486     { ISD::SRL,     MVT::v8i16,    2 },
487     { ISD::SRA,     MVT::v8i16,    2 },
488     { ISD::SHL,     MVT::v4i32,    1 },
489     { ISD::SRL,     MVT::v4i32,    2 },
490     { ISD::SRA,     MVT::v4i32,    2 },
491     { ISD::SHL,     MVT::v2i64,    1 },
492     { ISD::SRL,     MVT::v2i64,    2 },
493     { ISD::SRA,     MVT::v2i64,    2 },
494     // 256bit shifts require splitting if AVX2 didn't catch them above.
495     { ISD::SHL,     MVT::v32i8,  2+2 },
496     { ISD::SRL,     MVT::v32i8,  4+2 },
497     { ISD::SRA,     MVT::v32i8,  4+2 },
498     { ISD::SHL,     MVT::v16i16, 2+2 },
499     { ISD::SRL,     MVT::v16i16, 4+2 },
500     { ISD::SRA,     MVT::v16i16, 4+2 },
501     { ISD::SHL,     MVT::v8i32,  2+2 },
502     { ISD::SRL,     MVT::v8i32,  4+2 },
503     { ISD::SRA,     MVT::v8i32,  4+2 },
504     { ISD::SHL,     MVT::v4i64,  2+2 },
505     { ISD::SRL,     MVT::v4i64,  4+2 },
506     { ISD::SRA,     MVT::v4i64,  4+2 },
507   };
508 
509   // Look for XOP lowering tricks.
510   if (ST->hasXOP())
511     if (const auto *Entry = CostTableLookup(XOPShiftCostTable, ISD, LT.second))
512       return LT.first * Entry->Cost;
513 
514   static const CostTblEntry SSE2UniformShiftCostTable[] = {
515     // Uniform splats are cheaper for the following instructions.
516     { ISD::SHL,  MVT::v16i16, 2+2 }, // 2*psllw + split.
517     { ISD::SHL,  MVT::v8i32,  2+2 }, // 2*pslld + split.
518     { ISD::SHL,  MVT::v4i64,  2+2 }, // 2*psllq + split.
519 
520     { ISD::SRL,  MVT::v16i16, 2+2 }, // 2*psrlw + split.
521     { ISD::SRL,  MVT::v8i32,  2+2 }, // 2*psrld + split.
522     { ISD::SRL,  MVT::v4i64,  2+2 }, // 2*psrlq + split.
523 
524     { ISD::SRA,  MVT::v16i16, 2+2 }, // 2*psraw + split.
525     { ISD::SRA,  MVT::v8i32,  2+2 }, // 2*psrad + split.
526     { ISD::SRA,  MVT::v2i64,    4 }, // 2*psrad + shuffle.
527     { ISD::SRA,  MVT::v4i64,  8+2 }, // 2*(2*psrad + shuffle) + split.
528   };
529 
530   if (ST->hasSSE2() &&
531       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
532        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
533 
534     // Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table.
535     if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2())
536       return LT.first * 4; // 2*psrad + shuffle.
537 
538     if (const auto *Entry =
539             CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second))
540       return LT.first * Entry->Cost;
541   }
542 
543   if (ISD == ISD::SHL &&
544       Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
545     MVT VT = LT.second;
546     // Vector shift left by non uniform constant can be lowered
547     // into vector multiply.
548     if (((VT == MVT::v8i16 || VT == MVT::v4i32) && ST->hasSSE2()) ||
549         ((VT == MVT::v16i16 || VT == MVT::v8i32) && ST->hasAVX()))
550       ISD = ISD::MUL;
551   }
552 
553   static const CostTblEntry AVX2CostTable[] = {
554     { ISD::SHL,  MVT::v32i8,     11 }, // vpblendvb sequence.
555     { ISD::SHL,  MVT::v16i16,    10 }, // extend/vpsrlvd/pack sequence.
556 
557     { ISD::SRL,  MVT::v32i8,     11 }, // vpblendvb sequence.
558     { ISD::SRL,  MVT::v16i16,    10 }, // extend/vpsrlvd/pack sequence.
559 
560     { ISD::SRA,  MVT::v32i8,     24 }, // vpblendvb sequence.
561     { ISD::SRA,  MVT::v16i16,    10 }, // extend/vpsravd/pack sequence.
562     { ISD::SRA,  MVT::v2i64,      4 }, // srl/xor/sub sequence.
563     { ISD::SRA,  MVT::v4i64,      4 }, // srl/xor/sub sequence.
564 
565     { ISD::SUB,  MVT::v32i8,      1 }, // psubb
566     { ISD::ADD,  MVT::v32i8,      1 }, // paddb
567     { ISD::SUB,  MVT::v16i16,     1 }, // psubw
568     { ISD::ADD,  MVT::v16i16,     1 }, // paddw
569     { ISD::SUB,  MVT::v8i32,      1 }, // psubd
570     { ISD::ADD,  MVT::v8i32,      1 }, // paddd
571     { ISD::SUB,  MVT::v4i64,      1 }, // psubq
572     { ISD::ADD,  MVT::v4i64,      1 }, // paddq
573 
574     { ISD::MUL,  MVT::v32i8,     17 }, // extend/pmullw/trunc sequence.
575     { ISD::MUL,  MVT::v16i8,      7 }, // extend/pmullw/trunc sequence.
576     { ISD::MUL,  MVT::v16i16,     1 }, // pmullw
577     { ISD::MUL,  MVT::v8i32,      2 }, // pmulld (Haswell from agner.org)
578     { ISD::MUL,  MVT::v4i64,      8 }, // 3*pmuludq/3*shift/2*add
579 
580     { ISD::FDIV, MVT::f32,        7 }, // Haswell from http://www.agner.org/
581     { ISD::FDIV, MVT::v4f32,      7 }, // Haswell from http://www.agner.org/
582     { ISD::FDIV, MVT::v8f32,     14 }, // Haswell from http://www.agner.org/
583     { ISD::FDIV, MVT::f64,       14 }, // Haswell from http://www.agner.org/
584     { ISD::FDIV, MVT::v2f64,     14 }, // Haswell from http://www.agner.org/
585     { ISD::FDIV, MVT::v4f64,     28 }, // Haswell from http://www.agner.org/
586   };
587 
588   // Look for AVX2 lowering tricks for custom cases.
589   if (ST->hasAVX2())
590     if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second))
591       return LT.first * Entry->Cost;
592 
593   static const CostTblEntry AVX1CostTable[] = {
594     // We don't have to scalarize unsupported ops. We can issue two half-sized
595     // operations and we only need to extract the upper YMM half.
596     // Two ops + 1 extract + 1 insert = 4.
597     { ISD::MUL,     MVT::v16i16,     4 },
598     { ISD::MUL,     MVT::v8i32,      4 },
599     { ISD::SUB,     MVT::v32i8,      4 },
600     { ISD::ADD,     MVT::v32i8,      4 },
601     { ISD::SUB,     MVT::v16i16,     4 },
602     { ISD::ADD,     MVT::v16i16,     4 },
603     { ISD::SUB,     MVT::v8i32,      4 },
604     { ISD::ADD,     MVT::v8i32,      4 },
605     { ISD::SUB,     MVT::v4i64,      4 },
606     { ISD::ADD,     MVT::v4i64,      4 },
607 
608     // A v4i64 multiply is custom lowered as two split v2i64 vectors that then
609     // are lowered as a series of long multiplies(3), shifts(3) and adds(2)
610     // Because we believe v4i64 to be a legal type, we must also include the
611     // extract+insert in the cost table. Therefore, the cost here is 18
612     // instead of 8.
613     { ISD::MUL,     MVT::v4i64,     18 },
614 
615     { ISD::MUL,     MVT::v32i8,     26 }, // extend/pmullw/trunc sequence.
616 
617     { ISD::FDIV,    MVT::f32,       14 }, // SNB from http://www.agner.org/
618     { ISD::FDIV,    MVT::v4f32,     14 }, // SNB from http://www.agner.org/
619     { ISD::FDIV,    MVT::v8f32,     28 }, // SNB from http://www.agner.org/
620     { ISD::FDIV,    MVT::f64,       22 }, // SNB from http://www.agner.org/
621     { ISD::FDIV,    MVT::v2f64,     22 }, // SNB from http://www.agner.org/
622     { ISD::FDIV,    MVT::v4f64,     44 }, // SNB from http://www.agner.org/
623 
624     // Vectorizing division is a bad idea. See the SSE2 table for more comments.
625     { ISD::SDIV,    MVT::v32i8,  32*20 },
626     { ISD::SDIV,    MVT::v16i16, 16*20 },
627     { ISD::SDIV,    MVT::v8i32,   8*20 },
628     { ISD::SDIV,    MVT::v4i64,   4*20 },
629     { ISD::UDIV,    MVT::v32i8,  32*20 },
630     { ISD::UDIV,    MVT::v16i16, 16*20 },
631     { ISD::UDIV,    MVT::v8i32,   8*20 },
632     { ISD::UDIV,    MVT::v4i64,   4*20 },
633   };
634 
635   if (ST->hasAVX())
636     if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, LT.second))
637       return LT.first * Entry->Cost;
638 
639   static const CostTblEntry SSE42CostTable[] = {
640     { ISD::FDIV,  MVT::f32,   14 }, // Nehalem from http://www.agner.org/
641     { ISD::FDIV,  MVT::v4f32, 14 }, // Nehalem from http://www.agner.org/
642     { ISD::FDIV,  MVT::f64,   22 }, // Nehalem from http://www.agner.org/
643     { ISD::FDIV,  MVT::v2f64, 22 }, // Nehalem from http://www.agner.org/
644   };
645 
646   if (ST->hasSSE42())
647     if (const auto *Entry = CostTableLookup(SSE42CostTable, ISD, LT.second))
648       return LT.first * Entry->Cost;
649 
650   static const CostTblEntry SSE41CostTable[] = {
651     { ISD::SHL,  MVT::v16i8,      11 }, // pblendvb sequence.
652     { ISD::SHL,  MVT::v32i8,  2*11+2 }, // pblendvb sequence + split.
653     { ISD::SHL,  MVT::v8i16,      14 }, // pblendvb sequence.
654     { ISD::SHL,  MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
655     { ISD::SHL,  MVT::v4i32,       4 }, // pslld/paddd/cvttps2dq/pmulld
656     { ISD::SHL,  MVT::v8i32,   2*4+2 }, // pslld/paddd/cvttps2dq/pmulld + split
657 
658     { ISD::SRL,  MVT::v16i8,      12 }, // pblendvb sequence.
659     { ISD::SRL,  MVT::v32i8,  2*12+2 }, // pblendvb sequence + split.
660     { ISD::SRL,  MVT::v8i16,      14 }, // pblendvb sequence.
661     { ISD::SRL,  MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
662     { ISD::SRL,  MVT::v4i32,      11 }, // Shift each lane + blend.
663     { ISD::SRL,  MVT::v8i32,  2*11+2 }, // Shift each lane + blend + split.
664 
665     { ISD::SRA,  MVT::v16i8,      24 }, // pblendvb sequence.
666     { ISD::SRA,  MVT::v32i8,  2*24+2 }, // pblendvb sequence + split.
667     { ISD::SRA,  MVT::v8i16,      14 }, // pblendvb sequence.
668     { ISD::SRA,  MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
669     { ISD::SRA,  MVT::v4i32,      12 }, // Shift each lane + blend.
670     { ISD::SRA,  MVT::v8i32,  2*12+2 }, // Shift each lane + blend + split.
671 
672     { ISD::MUL,  MVT::v4i32,       2 }  // pmulld (Nehalem from agner.org)
673   };
674 
675   if (ST->hasSSE41())
676     if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second))
677       return LT.first * Entry->Cost;
678 
679   static const CostTblEntry SSE2CostTable[] = {
680     // We don't correctly identify costs of casts because they are marked as
681     // custom.
682     { ISD::SHL,  MVT::v16i8,      26 }, // cmpgtb sequence.
683     { ISD::SHL,  MVT::v8i16,      32 }, // cmpgtb sequence.
684     { ISD::SHL,  MVT::v4i32,     2*5 }, // We optimized this using mul.
685     { ISD::SHL,  MVT::v2i64,       4 }, // splat+shuffle sequence.
686     { ISD::SHL,  MVT::v4i64,   2*4+2 }, // splat+shuffle sequence + split.
687 
688     { ISD::SRL,  MVT::v16i8,      26 }, // cmpgtb sequence.
689     { ISD::SRL,  MVT::v8i16,      32 }, // cmpgtb sequence.
690     { ISD::SRL,  MVT::v4i32,      16 }, // Shift each lane + blend.
691     { ISD::SRL,  MVT::v2i64,       4 }, // splat+shuffle sequence.
692     { ISD::SRL,  MVT::v4i64,   2*4+2 }, // splat+shuffle sequence + split.
693 
694     { ISD::SRA,  MVT::v16i8,      54 }, // unpacked cmpgtb sequence.
695     { ISD::SRA,  MVT::v8i16,      32 }, // cmpgtb sequence.
696     { ISD::SRA,  MVT::v4i32,      16 }, // Shift each lane + blend.
697     { ISD::SRA,  MVT::v2i64,      12 }, // srl/xor/sub sequence.
698     { ISD::SRA,  MVT::v4i64,  2*12+2 }, // srl/xor/sub sequence+split.
699 
700     { ISD::MUL,  MVT::v16i8,      12 }, // extend/pmullw/trunc sequence.
701     { ISD::MUL,  MVT::v8i16,       1 }, // pmullw
702     { ISD::MUL,  MVT::v4i32,       6 }, // 3*pmuludq/4*shuffle
703     { ISD::MUL,  MVT::v2i64,       8 }, // 3*pmuludq/3*shift/2*add
704 
705     { ISD::FDIV, MVT::f32,        23 }, // Pentium IV from http://www.agner.org/
706     { ISD::FDIV, MVT::v4f32,      39 }, // Pentium IV from http://www.agner.org/
707     { ISD::FDIV, MVT::f64,        38 }, // Pentium IV from http://www.agner.org/
708     { ISD::FDIV, MVT::v2f64,      69 }, // Pentium IV from http://www.agner.org/
709 
710     // It is not a good idea to vectorize division. We have to scalarize it and
711     // in the process we will often end up having to spilling regular
712     // registers. The overhead of division is going to dominate most kernels
713     // anyways so try hard to prevent vectorization of division - it is
714     // generally a bad idea. Assume somewhat arbitrarily that we have to be able
715     // to hide "20 cycles" for each lane.
716     { ISD::SDIV,  MVT::v16i8,  16*20 },
717     { ISD::SDIV,  MVT::v8i16,   8*20 },
718     { ISD::SDIV,  MVT::v4i32,   4*20 },
719     { ISD::SDIV,  MVT::v2i64,   2*20 },
720     { ISD::UDIV,  MVT::v16i8,  16*20 },
721     { ISD::UDIV,  MVT::v8i16,   8*20 },
722     { ISD::UDIV,  MVT::v4i32,   4*20 },
723     { ISD::UDIV,  MVT::v2i64,   2*20 },
724   };
725 
726   if (ST->hasSSE2())
727     if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second))
728       return LT.first * Entry->Cost;
729 
730   static const CostTblEntry SSE1CostTable[] = {
731     { ISD::FDIV, MVT::f32,   17 }, // Pentium III from http://www.agner.org/
732     { ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/
733   };
734 
735   if (ST->hasSSE1())
736     if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second))
737       return LT.first * Entry->Cost;
738 
739   // Fallback to the default implementation.
740   return BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
741 }
742 
743 int X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
744                                Type *SubTp) {
745   // 64-bit packed float vectors (v2f32) are widened to type v4f32.
746   // 64-bit packed integer vectors (v2i32) are promoted to type v2i64.
747   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
748 
749   // For Broadcasts we are splatting the first element from the first input
750   // register, so only need to reference that input and all the output
751   // registers are the same.
752   if (Kind == TTI::SK_Broadcast)
753     LT.first = 1;
754 
755   // We are going to permute multiple sources and the result will be in multiple
756   // destinations. Providing an accurate cost only for splits where the element
757   // type remains the same.
758   if (Kind == TTI::SK_PermuteSingleSrc && LT.first != 1) {
759     MVT LegalVT = LT.second;
760     if (LegalVT.isVector() &&
761         LegalVT.getVectorElementType().getSizeInBits() ==
762             Tp->getVectorElementType()->getPrimitiveSizeInBits() &&
763         LegalVT.getVectorNumElements() < Tp->getVectorNumElements()) {
764 
765       unsigned VecTySize = DL.getTypeStoreSize(Tp);
766       unsigned LegalVTSize = LegalVT.getStoreSize();
767       // Number of source vectors after legalization:
768       unsigned NumOfSrcs = (VecTySize + LegalVTSize - 1) / LegalVTSize;
769       // Number of destination vectors after legalization:
770       unsigned NumOfDests = LT.first;
771 
772       Type *SingleOpTy = VectorType::get(Tp->getVectorElementType(),
773                                          LegalVT.getVectorNumElements());
774 
775       unsigned NumOfShuffles = (NumOfSrcs - 1) * NumOfDests;
776       return NumOfShuffles *
777              getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy, 0, nullptr);
778     }
779 
780     return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
781   }
782 
783   // For 2-input shuffles, we must account for splitting the 2 inputs into many.
784   if (Kind == TTI::SK_PermuteTwoSrc && LT.first != 1) {
785     // We assume that source and destination have the same vector type.
786     int NumOfDests = LT.first;
787     int NumOfShufflesPerDest = LT.first * 2 - 1;
788     LT.first = NumOfDests * NumOfShufflesPerDest;
789   }
790 
791   static const CostTblEntry AVX512VBMIShuffleTbl[] = {
792     { TTI::SK_Reverse,          MVT::v64i8,  1 }, // vpermb
793     { TTI::SK_Reverse,          MVT::v32i8,  1 }, // vpermb
794 
795     { TTI::SK_PermuteSingleSrc, MVT::v64i8,  1 }, // vpermb
796     { TTI::SK_PermuteSingleSrc, MVT::v32i8,  1 }, // vpermb
797 
798     { TTI::SK_PermuteTwoSrc,    MVT::v64i8,  1 }, // vpermt2b
799     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,  1 }, // vpermt2b
800     { TTI::SK_PermuteTwoSrc,    MVT::v16i8,  1 }  // vpermt2b
801   };
802 
803   if (ST->hasVBMI())
804     if (const auto *Entry =
805             CostTableLookup(AVX512VBMIShuffleTbl, Kind, LT.second))
806       return LT.first * Entry->Cost;
807 
808   static const CostTblEntry AVX512BWShuffleTbl[] = {
809     { TTI::SK_Broadcast,        MVT::v32i16, 1 }, // vpbroadcastw
810     { TTI::SK_Broadcast,        MVT::v64i8,  1 }, // vpbroadcastb
811 
812     { TTI::SK_Reverse,          MVT::v32i16, 1 }, // vpermw
813     { TTI::SK_Reverse,          MVT::v16i16, 1 }, // vpermw
814     { TTI::SK_Reverse,          MVT::v64i8,  2 }, // pshufb + vshufi64x2
815 
816     { TTI::SK_PermuteSingleSrc, MVT::v32i16, 1 }, // vpermw
817     { TTI::SK_PermuteSingleSrc, MVT::v16i16, 1 }, // vpermw
818     { TTI::SK_PermuteSingleSrc, MVT::v8i16,  1 }, // vpermw
819     { TTI::SK_PermuteSingleSrc, MVT::v64i8,  8 }, // extend to v32i16
820     { TTI::SK_PermuteSingleSrc, MVT::v32i8,  3 }, // vpermw + zext/trunc
821 
822     { TTI::SK_PermuteTwoSrc,    MVT::v32i16, 1 }, // vpermt2w
823     { TTI::SK_PermuteTwoSrc,    MVT::v16i16, 1 }, // vpermt2w
824     { TTI::SK_PermuteTwoSrc,    MVT::v8i16,  1 }, // vpermt2w
825     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,  3 }, // zext + vpermt2w + trunc
826     { TTI::SK_PermuteTwoSrc,    MVT::v64i8, 19 }, // 6 * v32i8 + 1
827     { TTI::SK_PermuteTwoSrc,    MVT::v16i8,  3 }  // zext + vpermt2w + trunc
828   };
829 
830   if (ST->hasBWI())
831     if (const auto *Entry =
832             CostTableLookup(AVX512BWShuffleTbl, Kind, LT.second))
833       return LT.first * Entry->Cost;
834 
835   static const CostTblEntry AVX512ShuffleTbl[] = {
836     { TTI::SK_Broadcast,        MVT::v8f64,  1 }, // vbroadcastpd
837     { TTI::SK_Broadcast,        MVT::v16f32, 1 }, // vbroadcastps
838     { TTI::SK_Broadcast,        MVT::v8i64,  1 }, // vpbroadcastq
839     { TTI::SK_Broadcast,        MVT::v16i32, 1 }, // vpbroadcastd
840 
841     { TTI::SK_Reverse,          MVT::v8f64,  1 }, // vpermpd
842     { TTI::SK_Reverse,          MVT::v16f32, 1 }, // vpermps
843     { TTI::SK_Reverse,          MVT::v8i64,  1 }, // vpermq
844     { TTI::SK_Reverse,          MVT::v16i32, 1 }, // vpermd
845 
846     { TTI::SK_PermuteSingleSrc, MVT::v8f64,  1 }, // vpermpd
847     { TTI::SK_PermuteSingleSrc, MVT::v4f64,  1 }, // vpermpd
848     { TTI::SK_PermuteSingleSrc, MVT::v2f64,  1 }, // vpermpd
849     { TTI::SK_PermuteSingleSrc, MVT::v16f32, 1 }, // vpermps
850     { TTI::SK_PermuteSingleSrc, MVT::v8f32,  1 }, // vpermps
851     { TTI::SK_PermuteSingleSrc, MVT::v4f32,  1 }, // vpermps
852     { TTI::SK_PermuteSingleSrc, MVT::v8i64,  1 }, // vpermq
853     { TTI::SK_PermuteSingleSrc, MVT::v4i64,  1 }, // vpermq
854     { TTI::SK_PermuteSingleSrc, MVT::v2i64,  1 }, // vpermq
855     { TTI::SK_PermuteSingleSrc, MVT::v16i32, 1 }, // vpermd
856     { TTI::SK_PermuteSingleSrc, MVT::v8i32,  1 }, // vpermd
857     { TTI::SK_PermuteSingleSrc, MVT::v4i32,  1 }, // vpermd
858     { TTI::SK_PermuteSingleSrc, MVT::v16i8,  1 }, // pshufb
859 
860     { TTI::SK_PermuteTwoSrc,    MVT::v8f64,  1 }, // vpermt2pd
861     { TTI::SK_PermuteTwoSrc,    MVT::v16f32, 1 }, // vpermt2ps
862     { TTI::SK_PermuteTwoSrc,    MVT::v8i64,  1 }, // vpermt2q
863     { TTI::SK_PermuteTwoSrc,    MVT::v16i32, 1 }, // vpermt2d
864     { TTI::SK_PermuteTwoSrc,    MVT::v4f64,  1 }, // vpermt2pd
865     { TTI::SK_PermuteTwoSrc,    MVT::v8f32,  1 }, // vpermt2ps
866     { TTI::SK_PermuteTwoSrc,    MVT::v4i64,  1 }, // vpermt2q
867     { TTI::SK_PermuteTwoSrc,    MVT::v8i32,  1 }, // vpermt2d
868     { TTI::SK_PermuteTwoSrc,    MVT::v2f64,  1 }, // vpermt2pd
869     { TTI::SK_PermuteTwoSrc,    MVT::v4f32,  1 }, // vpermt2ps
870     { TTI::SK_PermuteTwoSrc,    MVT::v2i64,  1 }, // vpermt2q
871     { TTI::SK_PermuteTwoSrc,    MVT::v4i32,  1 }  // vpermt2d
872   };
873 
874   if (ST->hasAVX512())
875     if (const auto *Entry = CostTableLookup(AVX512ShuffleTbl, Kind, LT.second))
876       return LT.first * Entry->Cost;
877 
878   static const CostTblEntry AVX2ShuffleTbl[] = {
879     { TTI::SK_Broadcast, MVT::v4f64,  1 }, // vbroadcastpd
880     { TTI::SK_Broadcast, MVT::v8f32,  1 }, // vbroadcastps
881     { TTI::SK_Broadcast, MVT::v4i64,  1 }, // vpbroadcastq
882     { TTI::SK_Broadcast, MVT::v8i32,  1 }, // vpbroadcastd
883     { TTI::SK_Broadcast, MVT::v16i16, 1 }, // vpbroadcastw
884     { TTI::SK_Broadcast, MVT::v32i8,  1 }, // vpbroadcastb
885 
886     { TTI::SK_Reverse,   MVT::v4f64,  1 }, // vpermpd
887     { TTI::SK_Reverse,   MVT::v8f32,  1 }, // vpermps
888     { TTI::SK_Reverse,   MVT::v4i64,  1 }, // vpermq
889     { TTI::SK_Reverse,   MVT::v8i32,  1 }, // vpermd
890     { TTI::SK_Reverse,   MVT::v16i16, 2 }, // vperm2i128 + pshufb
891     { TTI::SK_Reverse,   MVT::v32i8,  2 }, // vperm2i128 + pshufb
892 
893     { TTI::SK_Alternate, MVT::v16i16, 1 }, // vpblendw
894     { TTI::SK_Alternate, MVT::v32i8,  1 }, // vpblendvb
895 
896     { TTI::SK_PermuteSingleSrc, MVT::v4f64,  1 }, // vpermpd
897     { TTI::SK_PermuteSingleSrc, MVT::v8f32,  1 }, // vpermps
898     { TTI::SK_PermuteSingleSrc, MVT::v4i64,  1 }, // vpermq
899     { TTI::SK_PermuteSingleSrc, MVT::v8i32,  1 }, // vpermd
900     { TTI::SK_PermuteSingleSrc, MVT::v16i16, 4 }, // vperm2i128 + 2*vpshufb
901                                                   // + vpblendvb
902     { TTI::SK_PermuteSingleSrc, MVT::v32i8,  4 }, // vperm2i128 + 2*vpshufb
903                                                   // + vpblendvb
904 
905     { TTI::SK_PermuteTwoSrc,    MVT::v4f64,  3 }, // 2*vpermpd + vblendpd
906     { TTI::SK_PermuteTwoSrc,    MVT::v8f32,  3 }, // 2*vpermps + vblendps
907     { TTI::SK_PermuteTwoSrc,    MVT::v4i64,  3 }, // 2*vpermq + vpblendd
908     { TTI::SK_PermuteTwoSrc,    MVT::v8i32,  3 }, // 2*vpermd + vpblendd
909     { TTI::SK_PermuteTwoSrc,    MVT::v16i16, 7 }, // 2*vperm2i128 + 4*vpshufb
910                                                   // + vpblendvb
911     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,  7 }, // 2*vperm2i128 + 4*vpshufb
912                                                   // + vpblendvb
913   };
914 
915   if (ST->hasAVX2())
916     if (const auto *Entry = CostTableLookup(AVX2ShuffleTbl, Kind, LT.second))
917       return LT.first * Entry->Cost;
918 
919   static const CostTblEntry XOPShuffleTbl[] = {
920     { TTI::SK_PermuteSingleSrc, MVT::v4f64,   2 }, // vperm2f128 + vpermil2pd
921     { TTI::SK_PermuteSingleSrc, MVT::v8f32,   2 }, // vperm2f128 + vpermil2ps
922     { TTI::SK_PermuteSingleSrc, MVT::v4i64,   2 }, // vperm2f128 + vpermil2pd
923     { TTI::SK_PermuteSingleSrc, MVT::v8i32,   2 }, // vperm2f128 + vpermil2ps
924     { TTI::SK_PermuteSingleSrc, MVT::v16i16,  4 }, // vextractf128 + 2*vpperm
925                                                    // + vinsertf128
926     { TTI::SK_PermuteSingleSrc, MVT::v32i8,   4 }, // vextractf128 + 2*vpperm
927                                                    // + vinsertf128
928 
929     { TTI::SK_PermuteTwoSrc,    MVT::v16i16,  9 }, // 2*vextractf128 + 6*vpperm
930                                                    // + vinsertf128
931     { TTI::SK_PermuteTwoSrc,    MVT::v8i16,   1 }, // vpperm
932     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,   9 }, // 2*vextractf128 + 6*vpperm
933                                                    // + vinsertf128
934     { TTI::SK_PermuteTwoSrc,    MVT::v16i8,   1 }, // vpperm
935   };
936 
937   if (ST->hasXOP())
938     if (const auto *Entry = CostTableLookup(XOPShuffleTbl, Kind, LT.second))
939       return LT.first * Entry->Cost;
940 
941   static const CostTblEntry AVX1ShuffleTbl[] = {
942     { TTI::SK_Broadcast, MVT::v4f64,  2 }, // vperm2f128 + vpermilpd
943     { TTI::SK_Broadcast, MVT::v8f32,  2 }, // vperm2f128 + vpermilps
944     { TTI::SK_Broadcast, MVT::v4i64,  2 }, // vperm2f128 + vpermilpd
945     { TTI::SK_Broadcast, MVT::v8i32,  2 }, // vperm2f128 + vpermilps
946     { TTI::SK_Broadcast, MVT::v16i16, 3 }, // vpshuflw + vpshufd + vinsertf128
947     { TTI::SK_Broadcast, MVT::v32i8,  2 }, // vpshufb + vinsertf128
948 
949     { TTI::SK_Reverse,   MVT::v4f64,  2 }, // vperm2f128 + vpermilpd
950     { TTI::SK_Reverse,   MVT::v8f32,  2 }, // vperm2f128 + vpermilps
951     { TTI::SK_Reverse,   MVT::v4i64,  2 }, // vperm2f128 + vpermilpd
952     { TTI::SK_Reverse,   MVT::v8i32,  2 }, // vperm2f128 + vpermilps
953     { TTI::SK_Reverse,   MVT::v16i16, 4 }, // vextractf128 + 2*pshufb
954                                            // + vinsertf128
955     { TTI::SK_Reverse,   MVT::v32i8,  4 }, // vextractf128 + 2*pshufb
956                                            // + vinsertf128
957 
958     { TTI::SK_Alternate, MVT::v4i64,  1 }, // vblendpd
959     { TTI::SK_Alternate, MVT::v4f64,  1 }, // vblendpd
960     { TTI::SK_Alternate, MVT::v8i32,  1 }, // vblendps
961     { TTI::SK_Alternate, MVT::v8f32,  1 }, // vblendps
962     { TTI::SK_Alternate, MVT::v16i16, 3 }, // vpand + vpandn + vpor
963     { TTI::SK_Alternate, MVT::v32i8,  3 }, // vpand + vpandn + vpor
964 
965     { TTI::SK_PermuteSingleSrc, MVT::v4f64,  3 }, // 2*vperm2f128 + vshufpd
966     { TTI::SK_PermuteSingleSrc, MVT::v4i64,  3 }, // 2*vperm2f128 + vshufpd
967     { TTI::SK_PermuteSingleSrc, MVT::v8f32,  4 }, // 2*vperm2f128 + 2*vshufps
968     { TTI::SK_PermuteSingleSrc, MVT::v8i32,  4 }, // 2*vperm2f128 + 2*vshufps
969     { TTI::SK_PermuteSingleSrc, MVT::v16i16, 8 }, // vextractf128 + 4*pshufb
970                                                   // + 2*por + vinsertf128
971     { TTI::SK_PermuteSingleSrc, MVT::v32i8,  8 }, // vextractf128 + 4*pshufb
972                                                   // + 2*por + vinsertf128
973 
974     { TTI::SK_PermuteTwoSrc,    MVT::v4f64,   4 }, // 2*vperm2f128 + 2*vshufpd
975     { TTI::SK_PermuteTwoSrc,    MVT::v8f32,   4 }, // 2*vperm2f128 + 2*vshufps
976     { TTI::SK_PermuteTwoSrc,    MVT::v4i64,   4 }, // 2*vperm2f128 + 2*vshufpd
977     { TTI::SK_PermuteTwoSrc,    MVT::v8i32,   4 }, // 2*vperm2f128 + 2*vshufps
978     { TTI::SK_PermuteTwoSrc,    MVT::v16i16, 15 }, // 2*vextractf128 + 8*pshufb
979                                                    // + 4*por + vinsertf128
980     { TTI::SK_PermuteTwoSrc,    MVT::v32i8,  15 }, // 2*vextractf128 + 8*pshufb
981                                                    // + 4*por + vinsertf128
982   };
983 
984   if (ST->hasAVX())
985     if (const auto *Entry = CostTableLookup(AVX1ShuffleTbl, Kind, LT.second))
986       return LT.first * Entry->Cost;
987 
988   static const CostTblEntry SSE41ShuffleTbl[] = {
989     { TTI::SK_Alternate, MVT::v2i64,  1 }, // pblendw
990     { TTI::SK_Alternate, MVT::v2f64,  1 }, // movsd
991     { TTI::SK_Alternate, MVT::v4i32,  1 }, // pblendw
992     { TTI::SK_Alternate, MVT::v4f32,  1 }, // blendps
993     { TTI::SK_Alternate, MVT::v8i16,  1 }, // pblendw
994     { TTI::SK_Alternate, MVT::v16i8,  1 }  // pblendvb
995   };
996 
997   if (ST->hasSSE41())
998     if (const auto *Entry = CostTableLookup(SSE41ShuffleTbl, Kind, LT.second))
999       return LT.first * Entry->Cost;
1000 
1001   static const CostTblEntry SSSE3ShuffleTbl[] = {
1002     { TTI::SK_Broadcast, MVT::v8i16,  1 }, // pshufb
1003     { TTI::SK_Broadcast, MVT::v16i8,  1 }, // pshufb
1004 
1005     { TTI::SK_Reverse,   MVT::v8i16,  1 }, // pshufb
1006     { TTI::SK_Reverse,   MVT::v16i8,  1 }, // pshufb
1007 
1008     { TTI::SK_Alternate, MVT::v8i16,  3 }, // 2*pshufb + por
1009     { TTI::SK_Alternate, MVT::v16i8,  3 }, // 2*pshufb + por
1010 
1011     { TTI::SK_PermuteSingleSrc, MVT::v8i16, 1 }, // pshufb
1012     { TTI::SK_PermuteSingleSrc, MVT::v16i8, 1 }, // pshufb
1013 
1014     { TTI::SK_PermuteTwoSrc,    MVT::v8i16, 3 }, // 2*pshufb + por
1015     { TTI::SK_PermuteTwoSrc,    MVT::v16i8, 3 }, // 2*pshufb + por
1016   };
1017 
1018   if (ST->hasSSSE3())
1019     if (const auto *Entry = CostTableLookup(SSSE3ShuffleTbl, Kind, LT.second))
1020       return LT.first * Entry->Cost;
1021 
1022   static const CostTblEntry SSE2ShuffleTbl[] = {
1023     { TTI::SK_Broadcast, MVT::v2f64,  1 }, // shufpd
1024     { TTI::SK_Broadcast, MVT::v2i64,  1 }, // pshufd
1025     { TTI::SK_Broadcast, MVT::v4i32,  1 }, // pshufd
1026     { TTI::SK_Broadcast, MVT::v8i16,  2 }, // pshuflw + pshufd
1027     { TTI::SK_Broadcast, MVT::v16i8,  3 }, // unpck + pshuflw + pshufd
1028 
1029     { TTI::SK_Reverse,   MVT::v2f64,  1 }, // shufpd
1030     { TTI::SK_Reverse,   MVT::v2i64,  1 }, // pshufd
1031     { TTI::SK_Reverse,   MVT::v4i32,  1 }, // pshufd
1032     { TTI::SK_Reverse,   MVT::v8i16,  3 }, // pshuflw + pshufhw + pshufd
1033     { TTI::SK_Reverse,   MVT::v16i8,  9 }, // 2*pshuflw + 2*pshufhw
1034                                            // + 2*pshufd + 2*unpck + packus
1035 
1036     { TTI::SK_Alternate, MVT::v2i64,  1 }, // movsd
1037     { TTI::SK_Alternate, MVT::v2f64,  1 }, // movsd
1038     { TTI::SK_Alternate, MVT::v4i32,  2 }, // 2*shufps
1039     { TTI::SK_Alternate, MVT::v8i16,  3 }, // pand + pandn + por
1040     { TTI::SK_Alternate, MVT::v16i8,  3 }, // pand + pandn + por
1041 
1042     { TTI::SK_PermuteSingleSrc, MVT::v2f64,  1 }, // shufpd
1043     { TTI::SK_PermuteSingleSrc, MVT::v2i64,  1 }, // pshufd
1044     { TTI::SK_PermuteSingleSrc, MVT::v4i32,  1 }, // pshufd
1045     { TTI::SK_PermuteSingleSrc, MVT::v8i16,  5 }, // 2*pshuflw + 2*pshufhw
1046                                                   // + pshufd/unpck
1047     { TTI::SK_PermuteSingleSrc, MVT::v16i8, 10 }, // 2*pshuflw + 2*pshufhw
1048                                                   // + 2*pshufd + 2*unpck + 2*packus
1049 
1050     { TTI::SK_PermuteTwoSrc,    MVT::v2f64,  1 }, // shufpd
1051     { TTI::SK_PermuteTwoSrc,    MVT::v2i64,  1 }, // shufpd
1052     { TTI::SK_PermuteTwoSrc,    MVT::v4i32,  2 }, // 2*{unpck,movsd,pshufd}
1053     { TTI::SK_PermuteTwoSrc,    MVT::v8i16,  8 }, // blend+permute
1054     { TTI::SK_PermuteTwoSrc,    MVT::v16i8, 13 }, // blend+permute
1055   };
1056 
1057   if (ST->hasSSE2())
1058     if (const auto *Entry = CostTableLookup(SSE2ShuffleTbl, Kind, LT.second))
1059       return LT.first * Entry->Cost;
1060 
1061   static const CostTblEntry SSE1ShuffleTbl[] = {
1062     { TTI::SK_Broadcast,        MVT::v4f32, 1 }, // shufps
1063     { TTI::SK_Reverse,          MVT::v4f32, 1 }, // shufps
1064     { TTI::SK_Alternate,        MVT::v4f32, 2 }, // 2*shufps
1065     { TTI::SK_PermuteSingleSrc, MVT::v4f32, 1 }, // shufps
1066     { TTI::SK_PermuteTwoSrc,    MVT::v4f32, 2 }, // 2*shufps
1067   };
1068 
1069   if (ST->hasSSE1())
1070     if (const auto *Entry = CostTableLookup(SSE1ShuffleTbl, Kind, LT.second))
1071       return LT.first * Entry->Cost;
1072 
1073   return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
1074 }
1075 
1076 int X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
1077                                  const Instruction *I) {
1078   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1079   assert(ISD && "Invalid opcode");
1080 
1081   // FIXME: Need a better design of the cost table to handle non-simple types of
1082   // potential massive combinations (elem_num x src_type x dst_type).
1083 
1084   static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = {
1085     { ISD::SINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
1086     { ISD::SINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
1087     { ISD::SINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
1088     { ISD::SINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
1089     { ISD::SINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
1090     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
1091 
1092     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
1093     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
1094     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
1095     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
1096     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
1097     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
1098 
1099     { ISD::FP_TO_SINT,  MVT::v2i64,  MVT::v2f32,  1 },
1100     { ISD::FP_TO_SINT,  MVT::v4i64,  MVT::v4f32,  1 },
1101     { ISD::FP_TO_SINT,  MVT::v8i64,  MVT::v8f32,  1 },
1102     { ISD::FP_TO_SINT,  MVT::v2i64,  MVT::v2f64,  1 },
1103     { ISD::FP_TO_SINT,  MVT::v4i64,  MVT::v4f64,  1 },
1104     { ISD::FP_TO_SINT,  MVT::v8i64,  MVT::v8f64,  1 },
1105 
1106     { ISD::FP_TO_UINT,  MVT::v2i64,  MVT::v2f32,  1 },
1107     { ISD::FP_TO_UINT,  MVT::v4i64,  MVT::v4f32,  1 },
1108     { ISD::FP_TO_UINT,  MVT::v8i64,  MVT::v8f32,  1 },
1109     { ISD::FP_TO_UINT,  MVT::v2i64,  MVT::v2f64,  1 },
1110     { ISD::FP_TO_UINT,  MVT::v4i64,  MVT::v4f64,  1 },
1111     { ISD::FP_TO_UINT,  MVT::v8i64,  MVT::v8f64,  1 },
1112   };
1113 
1114   // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and
1115   // 256-bit wide vectors.
1116 
1117   static const TypeConversionCostTblEntry AVX512FConversionTbl[] = {
1118     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v8f32,  1 },
1119     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v16f32, 3 },
1120     { ISD::FP_ROUND,  MVT::v8f32,   MVT::v8f64,  1 },
1121 
1122     { ISD::TRUNCATE,  MVT::v16i8,   MVT::v16i32, 1 },
1123     { ISD::TRUNCATE,  MVT::v16i16,  MVT::v16i32, 1 },
1124     { ISD::TRUNCATE,  MVT::v8i16,   MVT::v8i64,  1 },
1125     { ISD::TRUNCATE,  MVT::v8i32,   MVT::v8i64,  1 },
1126 
1127     // v16i1 -> v16i32 - load + broadcast
1128     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
1129     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
1130     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
1131     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
1132     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1133     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1134     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
1135     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
1136     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
1137     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
1138 
1139     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
1140     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
1141     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i8,   2 },
1142     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i8,  2 },
1143     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
1144     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i16, 2 },
1145     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
1146     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
1147     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64, 26 },
1148     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64, 26 },
1149 
1150     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
1151     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
1152     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i8,   2 },
1153     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i8,   2 },
1154     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i8,   2 },
1155     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i8,   2 },
1156     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i8,  2 },
1157     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i16,  5 },
1158     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i16,  2 },
1159     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  2 },
1160     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
1161     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i16, 2 },
1162     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  2 },
1163     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i32,  1 },
1164     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  1 },
1165     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  1 },
1166     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  1 },
1167     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
1168     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
1169     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  5 },
1170     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  5 },
1171     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64, 12 },
1172     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64, 26 },
1173 
1174     { ISD::FP_TO_UINT,  MVT::v2i32,  MVT::v2f32,  1 },
1175     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  1 },
1176     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f32,  1 },
1177     { ISD::FP_TO_UINT,  MVT::v8i16,  MVT::v8f64,  2 },
1178     { ISD::FP_TO_UINT,  MVT::v8i8,   MVT::v8f64,  2 },
1179     { ISD::FP_TO_UINT,  MVT::v16i32, MVT::v16f32, 1 },
1180     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v16f32, 2 },
1181     { ISD::FP_TO_UINT,  MVT::v16i8,  MVT::v16f32, 2 },
1182   };
1183 
1184   static const TypeConversionCostTblEntry AVX2ConversionTbl[] = {
1185     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
1186     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
1187     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
1188     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
1189     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,   3 },
1190     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,   3 },
1191     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   3 },
1192     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   3 },
1193     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
1194     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
1195     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
1196     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
1197     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
1198     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
1199     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
1200     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
1201 
1202     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i64,  2 },
1203     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i64,  2 },
1204     { ISD::TRUNCATE,    MVT::v4i32,  MVT::v4i64,  2 },
1205     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  2 },
1206     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  2 },
1207     { ISD::TRUNCATE,    MVT::v8i32,  MVT::v8i64,  4 },
1208 
1209     { ISD::FP_EXTEND,   MVT::v8f64,  MVT::v8f32,  3 },
1210     { ISD::FP_ROUND,    MVT::v8f32,  MVT::v8f64,  3 },
1211 
1212     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  8 },
1213   };
1214 
1215   static const TypeConversionCostTblEntry AVXConversionTbl[] = {
1216     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,  6 },
1217     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,  4 },
1218     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,  7 },
1219     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,  4 },
1220     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,  6 },
1221     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,  4 },
1222     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,  7 },
1223     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,  4 },
1224     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
1225     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
1226     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16, 6 },
1227     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16, 3 },
1228     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16, 4 },
1229     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16, 4 },
1230     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32, 4 },
1231     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32, 4 },
1232 
1233     { ISD::TRUNCATE,    MVT::v16i8, MVT::v16i16, 4 },
1234     { ISD::TRUNCATE,    MVT::v8i8,  MVT::v8i32,  4 },
1235     { ISD::TRUNCATE,    MVT::v8i16, MVT::v8i32,  5 },
1236     { ISD::TRUNCATE,    MVT::v4i8,  MVT::v4i64,  4 },
1237     { ISD::TRUNCATE,    MVT::v4i16, MVT::v4i64,  4 },
1238     { ISD::TRUNCATE,    MVT::v4i32, MVT::v4i64,  4 },
1239     { ISD::TRUNCATE,    MVT::v8i32, MVT::v8i64,  9 },
1240 
1241     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i1,  3 },
1242     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i1,  3 },
1243     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i1,  8 },
1244     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i8,  3 },
1245     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i8,  3 },
1246     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i8,  8 },
1247     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i16, 3 },
1248     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i16, 3 },
1249     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
1250     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },
1251     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i32, 1 },
1252     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i32, 1 },
1253 
1254     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i1,  7 },
1255     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i1,  7 },
1256     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i1,  6 },
1257     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i8,  2 },
1258     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i8,  2 },
1259     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i8,  5 },
1260     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
1261     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i16, 2 },
1262     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
1263     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i32, 6 },
1264     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i32, 6 },
1265     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i32, 6 },
1266     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i32, 9 },
1267     // The generic code to compute the scalar overhead is currently broken.
1268     // Workaround this limitation by estimating the scalarization overhead
1269     // here. We have roughly 10 instructions per scalar element.
1270     // Multiply that by the vector width.
1271     // FIXME: remove that when PR19268 is fixed.
1272     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i64, 10 },
1273     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i64, 20 },
1274     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
1275     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
1276 
1277     { ISD::FP_TO_SINT,  MVT::v4i8,  MVT::v4f32, 1 },
1278     { ISD::FP_TO_SINT,  MVT::v8i8,  MVT::v8f32, 7 },
1279     // This node is expanded into scalarized operations but BasicTTI is overly
1280     // optimistic estimating its cost.  It computes 3 per element (one
1281     // vector-extract, one scalar conversion and one vector-insert).  The
1282     // problem is that the inserts form a read-modify-write chain so latency
1283     // should be factored in too.  Inflating the cost per element by 1.
1284     { ISD::FP_TO_UINT,  MVT::v8i32, MVT::v8f32, 8*4 },
1285     { ISD::FP_TO_UINT,  MVT::v4i32, MVT::v4f64, 4*4 },
1286 
1287     { ISD::FP_EXTEND,   MVT::v4f64,  MVT::v4f32,  1 },
1288     { ISD::FP_ROUND,    MVT::v4f32,  MVT::v4f64,  1 },
1289   };
1290 
1291   static const TypeConversionCostTblEntry SSE41ConversionTbl[] = {
1292     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8,    2 },
1293     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8,    2 },
1294     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16,   2 },
1295     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16,   2 },
1296     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32,   2 },
1297     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32,   2 },
1298 
1299     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i8,   1 },
1300     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i8,   2 },
1301     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i8,   1 },
1302     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i8,   1 },
1303     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
1304     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
1305     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   2 },
1306     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   2 },
1307     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
1308     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
1309     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  4 },
1310     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  4 },
1311     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
1312     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
1313     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
1314     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
1315     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 4 },
1316     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 4 },
1317 
1318     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  2 },
1319     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  1 },
1320     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i32,  1 },
1321     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i32,  1 },
1322     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  3 },
1323     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  3 },
1324     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 6 },
1325 
1326   };
1327 
1328   static const TypeConversionCostTblEntry SSE2ConversionTbl[] = {
1329     // These are somewhat magic numbers justified by looking at the output of
1330     // Intel's IACA, running some kernels and making sure when we take
1331     // legalization into account the throughput will be overestimated.
1332     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
1333     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
1334     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
1335     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
1336     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 },
1337     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
1338     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
1339     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
1340 
1341     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
1342     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
1343     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
1344     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
1345     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
1346     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 8 },
1347     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
1348     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
1349 
1350     { ISD::FP_TO_SINT,  MVT::v2i32,  MVT::v2f64,  3 },
1351 
1352     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i8,   1 },
1353     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i8,   6 },
1354     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i8,   2 },
1355     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i8,   3 },
1356     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,   4 },
1357     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,   8 },
1358     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
1359     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i8,   2 },
1360     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   6 },
1361     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   6 },
1362     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  3 },
1363     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  4 },
1364     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  9 },
1365     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  12 },
1366     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
1367     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i16,  2 },
1368     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
1369     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16,  10 },
1370     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  3 },
1371     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  4 },
1372     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 6 },
1373     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 8 },
1374     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  3 },
1375     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  5 },
1376 
1377     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  4 },
1378     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  2 },
1379     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 3 },
1380     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i32,  3 },
1381     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i32,  3 },
1382     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  4 },
1383     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i32, 7 },
1384     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  5 },
1385     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 10 },
1386   };
1387 
1388   std::pair<int, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src);
1389   std::pair<int, MVT> LTDest = TLI->getTypeLegalizationCost(DL, Dst);
1390 
1391   if (ST->hasSSE2() && !ST->hasAVX()) {
1392     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
1393                                                    LTDest.second, LTSrc.second))
1394       return LTSrc.first * Entry->Cost;
1395   }
1396 
1397   EVT SrcTy = TLI->getValueType(DL, Src);
1398   EVT DstTy = TLI->getValueType(DL, Dst);
1399 
1400   // The function getSimpleVT only handles simple value types.
1401   if (!SrcTy.isSimple() || !DstTy.isSimple())
1402     return BaseT::getCastInstrCost(Opcode, Dst, Src);
1403 
1404   if (ST->hasDQI())
1405     if (const auto *Entry = ConvertCostTableLookup(AVX512DQConversionTbl, ISD,
1406                                                    DstTy.getSimpleVT(),
1407                                                    SrcTy.getSimpleVT()))
1408       return Entry->Cost;
1409 
1410   if (ST->hasAVX512())
1411     if (const auto *Entry = ConvertCostTableLookup(AVX512FConversionTbl, ISD,
1412                                                    DstTy.getSimpleVT(),
1413                                                    SrcTy.getSimpleVT()))
1414       return Entry->Cost;
1415 
1416   if (ST->hasAVX2()) {
1417     if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
1418                                                    DstTy.getSimpleVT(),
1419                                                    SrcTy.getSimpleVT()))
1420       return Entry->Cost;
1421   }
1422 
1423   if (ST->hasAVX()) {
1424     if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
1425                                                    DstTy.getSimpleVT(),
1426                                                    SrcTy.getSimpleVT()))
1427       return Entry->Cost;
1428   }
1429 
1430   if (ST->hasSSE41()) {
1431     if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
1432                                                    DstTy.getSimpleVT(),
1433                                                    SrcTy.getSimpleVT()))
1434       return Entry->Cost;
1435   }
1436 
1437   if (ST->hasSSE2()) {
1438     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
1439                                                    DstTy.getSimpleVT(),
1440                                                    SrcTy.getSimpleVT()))
1441       return Entry->Cost;
1442   }
1443 
1444   return BaseT::getCastInstrCost(Opcode, Dst, Src, I);
1445 }
1446 
1447 int X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
1448                                    const Instruction *I) {
1449   // Legalize the type.
1450   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
1451 
1452   MVT MTy = LT.second;
1453 
1454   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1455   assert(ISD && "Invalid opcode");
1456 
1457   static const CostTblEntry SSE2CostTbl[] = {
1458     { ISD::SETCC,   MVT::v2i64,   8 },
1459     { ISD::SETCC,   MVT::v4i32,   1 },
1460     { ISD::SETCC,   MVT::v8i16,   1 },
1461     { ISD::SETCC,   MVT::v16i8,   1 },
1462   };
1463 
1464   static const CostTblEntry SSE42CostTbl[] = {
1465     { ISD::SETCC,   MVT::v2f64,   1 },
1466     { ISD::SETCC,   MVT::v4f32,   1 },
1467     { ISD::SETCC,   MVT::v2i64,   1 },
1468   };
1469 
1470   static const CostTblEntry AVX1CostTbl[] = {
1471     { ISD::SETCC,   MVT::v4f64,   1 },
1472     { ISD::SETCC,   MVT::v8f32,   1 },
1473     // AVX1 does not support 8-wide integer compare.
1474     { ISD::SETCC,   MVT::v4i64,   4 },
1475     { ISD::SETCC,   MVT::v8i32,   4 },
1476     { ISD::SETCC,   MVT::v16i16,  4 },
1477     { ISD::SETCC,   MVT::v32i8,   4 },
1478   };
1479 
1480   static const CostTblEntry AVX2CostTbl[] = {
1481     { ISD::SETCC,   MVT::v4i64,   1 },
1482     { ISD::SETCC,   MVT::v8i32,   1 },
1483     { ISD::SETCC,   MVT::v16i16,  1 },
1484     { ISD::SETCC,   MVT::v32i8,   1 },
1485   };
1486 
1487   static const CostTblEntry AVX512CostTbl[] = {
1488     { ISD::SETCC,   MVT::v8i64,   1 },
1489     { ISD::SETCC,   MVT::v16i32,  1 },
1490     { ISD::SETCC,   MVT::v8f64,   1 },
1491     { ISD::SETCC,   MVT::v16f32,  1 },
1492   };
1493 
1494   if (ST->hasAVX512())
1495     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
1496       return LT.first * Entry->Cost;
1497 
1498   if (ST->hasAVX2())
1499     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
1500       return LT.first * Entry->Cost;
1501 
1502   if (ST->hasAVX())
1503     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
1504       return LT.first * Entry->Cost;
1505 
1506   if (ST->hasSSE42())
1507     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
1508       return LT.first * Entry->Cost;
1509 
1510   if (ST->hasSSE2())
1511     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
1512       return LT.first * Entry->Cost;
1513 
1514   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, I);
1515 }
1516 
1517 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; }
1518 
1519 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1520                                       ArrayRef<Type *> Tys, FastMathFlags FMF,
1521                                       unsigned ScalarizationCostPassed) {
1522   // Costs should match the codegen from:
1523   // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll
1524   // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll
1525   // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll
1526   // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll
1527   // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll
1528   static const CostTblEntry AVX512CDCostTbl[] = {
1529     { ISD::CTLZ,       MVT::v8i64,   1 },
1530     { ISD::CTLZ,       MVT::v16i32,  1 },
1531     { ISD::CTLZ,       MVT::v32i16,  8 },
1532     { ISD::CTLZ,       MVT::v64i8,  20 },
1533     { ISD::CTLZ,       MVT::v4i64,   1 },
1534     { ISD::CTLZ,       MVT::v8i32,   1 },
1535     { ISD::CTLZ,       MVT::v16i16,  4 },
1536     { ISD::CTLZ,       MVT::v32i8,  10 },
1537     { ISD::CTLZ,       MVT::v2i64,   1 },
1538     { ISD::CTLZ,       MVT::v4i32,   1 },
1539     { ISD::CTLZ,       MVT::v8i16,   4 },
1540     { ISD::CTLZ,       MVT::v16i8,   4 },
1541   };
1542   static const CostTblEntry AVX512BWCostTbl[] = {
1543     { ISD::BITREVERSE, MVT::v8i64,   5 },
1544     { ISD::BITREVERSE, MVT::v16i32,  5 },
1545     { ISD::BITREVERSE, MVT::v32i16,  5 },
1546     { ISD::BITREVERSE, MVT::v64i8,   5 },
1547     { ISD::CTLZ,       MVT::v8i64,  23 },
1548     { ISD::CTLZ,       MVT::v16i32, 22 },
1549     { ISD::CTLZ,       MVT::v32i16, 18 },
1550     { ISD::CTLZ,       MVT::v64i8,  17 },
1551     { ISD::CTPOP,      MVT::v8i64,   7 },
1552     { ISD::CTPOP,      MVT::v16i32, 11 },
1553     { ISD::CTPOP,      MVT::v32i16,  9 },
1554     { ISD::CTPOP,      MVT::v64i8,   6 },
1555     { ISD::CTTZ,       MVT::v8i64,  10 },
1556     { ISD::CTTZ,       MVT::v16i32, 14 },
1557     { ISD::CTTZ,       MVT::v32i16, 12 },
1558     { ISD::CTTZ,       MVT::v64i8,   9 },
1559   };
1560   static const CostTblEntry AVX512CostTbl[] = {
1561     { ISD::BITREVERSE, MVT::v8i64,  36 },
1562     { ISD::BITREVERSE, MVT::v16i32, 24 },
1563     { ISD::CTLZ,       MVT::v8i64,  29 },
1564     { ISD::CTLZ,       MVT::v16i32, 35 },
1565     { ISD::CTPOP,      MVT::v8i64,  16 },
1566     { ISD::CTPOP,      MVT::v16i32, 24 },
1567     { ISD::CTTZ,       MVT::v8i64,  20 },
1568     { ISD::CTTZ,       MVT::v16i32, 28 },
1569   };
1570   static const CostTblEntry XOPCostTbl[] = {
1571     { ISD::BITREVERSE, MVT::v4i64,   4 },
1572     { ISD::BITREVERSE, MVT::v8i32,   4 },
1573     { ISD::BITREVERSE, MVT::v16i16,  4 },
1574     { ISD::BITREVERSE, MVT::v32i8,   4 },
1575     { ISD::BITREVERSE, MVT::v2i64,   1 },
1576     { ISD::BITREVERSE, MVT::v4i32,   1 },
1577     { ISD::BITREVERSE, MVT::v8i16,   1 },
1578     { ISD::BITREVERSE, MVT::v16i8,   1 },
1579     { ISD::BITREVERSE, MVT::i64,     3 },
1580     { ISD::BITREVERSE, MVT::i32,     3 },
1581     { ISD::BITREVERSE, MVT::i16,     3 },
1582     { ISD::BITREVERSE, MVT::i8,      3 }
1583   };
1584   static const CostTblEntry AVX2CostTbl[] = {
1585     { ISD::BITREVERSE, MVT::v4i64,   5 },
1586     { ISD::BITREVERSE, MVT::v8i32,   5 },
1587     { ISD::BITREVERSE, MVT::v16i16,  5 },
1588     { ISD::BITREVERSE, MVT::v32i8,   5 },
1589     { ISD::BSWAP,      MVT::v4i64,   1 },
1590     { ISD::BSWAP,      MVT::v8i32,   1 },
1591     { ISD::BSWAP,      MVT::v16i16,  1 },
1592     { ISD::CTLZ,       MVT::v4i64,  23 },
1593     { ISD::CTLZ,       MVT::v8i32,  18 },
1594     { ISD::CTLZ,       MVT::v16i16, 14 },
1595     { ISD::CTLZ,       MVT::v32i8,   9 },
1596     { ISD::CTPOP,      MVT::v4i64,   7 },
1597     { ISD::CTPOP,      MVT::v8i32,  11 },
1598     { ISD::CTPOP,      MVT::v16i16,  9 },
1599     { ISD::CTPOP,      MVT::v32i8,   6 },
1600     { ISD::CTTZ,       MVT::v4i64,  10 },
1601     { ISD::CTTZ,       MVT::v8i32,  14 },
1602     { ISD::CTTZ,       MVT::v16i16, 12 },
1603     { ISD::CTTZ,       MVT::v32i8,   9 },
1604     { ISD::FSQRT,      MVT::f32,     7 }, // Haswell from http://www.agner.org/
1605     { ISD::FSQRT,      MVT::v4f32,   7 }, // Haswell from http://www.agner.org/
1606     { ISD::FSQRT,      MVT::v8f32,  14 }, // Haswell from http://www.agner.org/
1607     { ISD::FSQRT,      MVT::f64,    14 }, // Haswell from http://www.agner.org/
1608     { ISD::FSQRT,      MVT::v2f64,  14 }, // Haswell from http://www.agner.org/
1609     { ISD::FSQRT,      MVT::v4f64,  28 }, // Haswell from http://www.agner.org/
1610   };
1611   static const CostTblEntry AVX1CostTbl[] = {
1612     { ISD::BITREVERSE, MVT::v4i64,  12 }, // 2 x 128-bit Op + extract/insert
1613     { ISD::BITREVERSE, MVT::v8i32,  12 }, // 2 x 128-bit Op + extract/insert
1614     { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert
1615     { ISD::BITREVERSE, MVT::v32i8,  12 }, // 2 x 128-bit Op + extract/insert
1616     { ISD::BSWAP,      MVT::v4i64,   4 },
1617     { ISD::BSWAP,      MVT::v8i32,   4 },
1618     { ISD::BSWAP,      MVT::v16i16,  4 },
1619     { ISD::CTLZ,       MVT::v4i64,  48 }, // 2 x 128-bit Op + extract/insert
1620     { ISD::CTLZ,       MVT::v8i32,  38 }, // 2 x 128-bit Op + extract/insert
1621     { ISD::CTLZ,       MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert
1622     { ISD::CTLZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
1623     { ISD::CTPOP,      MVT::v4i64,  16 }, // 2 x 128-bit Op + extract/insert
1624     { ISD::CTPOP,      MVT::v8i32,  24 }, // 2 x 128-bit Op + extract/insert
1625     { ISD::CTPOP,      MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert
1626     { ISD::CTPOP,      MVT::v32i8,  14 }, // 2 x 128-bit Op + extract/insert
1627     { ISD::CTTZ,       MVT::v4i64,  22 }, // 2 x 128-bit Op + extract/insert
1628     { ISD::CTTZ,       MVT::v8i32,  30 }, // 2 x 128-bit Op + extract/insert
1629     { ISD::CTTZ,       MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert
1630     { ISD::CTTZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
1631     { ISD::FSQRT,      MVT::f32,    14 }, // SNB from http://www.agner.org/
1632     { ISD::FSQRT,      MVT::v4f32,  14 }, // SNB from http://www.agner.org/
1633     { ISD::FSQRT,      MVT::v8f32,  28 }, // SNB from http://www.agner.org/
1634     { ISD::FSQRT,      MVT::f64,    21 }, // SNB from http://www.agner.org/
1635     { ISD::FSQRT,      MVT::v2f64,  21 }, // SNB from http://www.agner.org/
1636     { ISD::FSQRT,      MVT::v4f64,  43 }, // SNB from http://www.agner.org/
1637   };
1638   static const CostTblEntry SSE42CostTbl[] = {
1639     { ISD::FSQRT,      MVT::f32,    18 }, // Nehalem from http://www.agner.org/
1640     { ISD::FSQRT,      MVT::v4f32,  18 }, // Nehalem from http://www.agner.org/
1641   };
1642   static const CostTblEntry SSSE3CostTbl[] = {
1643     { ISD::BITREVERSE, MVT::v2i64,   5 },
1644     { ISD::BITREVERSE, MVT::v4i32,   5 },
1645     { ISD::BITREVERSE, MVT::v8i16,   5 },
1646     { ISD::BITREVERSE, MVT::v16i8,   5 },
1647     { ISD::BSWAP,      MVT::v2i64,   1 },
1648     { ISD::BSWAP,      MVT::v4i32,   1 },
1649     { ISD::BSWAP,      MVT::v8i16,   1 },
1650     { ISD::CTLZ,       MVT::v2i64,  23 },
1651     { ISD::CTLZ,       MVT::v4i32,  18 },
1652     { ISD::CTLZ,       MVT::v8i16,  14 },
1653     { ISD::CTLZ,       MVT::v16i8,   9 },
1654     { ISD::CTPOP,      MVT::v2i64,   7 },
1655     { ISD::CTPOP,      MVT::v4i32,  11 },
1656     { ISD::CTPOP,      MVT::v8i16,   9 },
1657     { ISD::CTPOP,      MVT::v16i8,   6 },
1658     { ISD::CTTZ,       MVT::v2i64,  10 },
1659     { ISD::CTTZ,       MVT::v4i32,  14 },
1660     { ISD::CTTZ,       MVT::v8i16,  12 },
1661     { ISD::CTTZ,       MVT::v16i8,   9 }
1662   };
1663   static const CostTblEntry SSE2CostTbl[] = {
1664     { ISD::BITREVERSE, MVT::v2i64,  29 },
1665     { ISD::BITREVERSE, MVT::v4i32,  27 },
1666     { ISD::BITREVERSE, MVT::v8i16,  27 },
1667     { ISD::BITREVERSE, MVT::v16i8,  20 },
1668     { ISD::BSWAP,      MVT::v2i64,   7 },
1669     { ISD::BSWAP,      MVT::v4i32,   7 },
1670     { ISD::BSWAP,      MVT::v8i16,   7 },
1671     { ISD::CTLZ,       MVT::v2i64,  25 },
1672     { ISD::CTLZ,       MVT::v4i32,  26 },
1673     { ISD::CTLZ,       MVT::v8i16,  20 },
1674     { ISD::CTLZ,       MVT::v16i8,  17 },
1675     { ISD::CTPOP,      MVT::v2i64,  12 },
1676     { ISD::CTPOP,      MVT::v4i32,  15 },
1677     { ISD::CTPOP,      MVT::v8i16,  13 },
1678     { ISD::CTPOP,      MVT::v16i8,  10 },
1679     { ISD::CTTZ,       MVT::v2i64,  14 },
1680     { ISD::CTTZ,       MVT::v4i32,  18 },
1681     { ISD::CTTZ,       MVT::v8i16,  16 },
1682     { ISD::CTTZ,       MVT::v16i8,  13 },
1683     { ISD::FSQRT,      MVT::f64,    32 }, // Nehalem from http://www.agner.org/
1684     { ISD::FSQRT,      MVT::v2f64,  32 }, // Nehalem from http://www.agner.org/
1685   };
1686   static const CostTblEntry SSE1CostTbl[] = {
1687     { ISD::FSQRT,      MVT::f32,    28 }, // Pentium III from http://www.agner.org/
1688     { ISD::FSQRT,      MVT::v4f32,  56 }, // Pentium III from http://www.agner.org/
1689   };
1690   static const CostTblEntry X64CostTbl[] = { // 64-bit targets
1691     { ISD::BITREVERSE, MVT::i64,    14 }
1692   };
1693   static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
1694     { ISD::BITREVERSE, MVT::i32,    14 },
1695     { ISD::BITREVERSE, MVT::i16,    14 },
1696     { ISD::BITREVERSE, MVT::i8,     11 }
1697   };
1698 
1699   unsigned ISD = ISD::DELETED_NODE;
1700   switch (IID) {
1701   default:
1702     break;
1703   case Intrinsic::bitreverse:
1704     ISD = ISD::BITREVERSE;
1705     break;
1706   case Intrinsic::bswap:
1707     ISD = ISD::BSWAP;
1708     break;
1709   case Intrinsic::ctlz:
1710     ISD = ISD::CTLZ;
1711     break;
1712   case Intrinsic::ctpop:
1713     ISD = ISD::CTPOP;
1714     break;
1715   case Intrinsic::cttz:
1716     ISD = ISD::CTTZ;
1717     break;
1718   case Intrinsic::sqrt:
1719     ISD = ISD::FSQRT;
1720     break;
1721   }
1722 
1723   // Legalize the type.
1724   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
1725   MVT MTy = LT.second;
1726 
1727   // Attempt to lookup cost.
1728   if (ST->hasCDI())
1729     if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy))
1730       return LT.first * Entry->Cost;
1731 
1732   if (ST->hasBWI())
1733     if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
1734       return LT.first * Entry->Cost;
1735 
1736   if (ST->hasAVX512())
1737     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
1738       return LT.first * Entry->Cost;
1739 
1740   if (ST->hasXOP())
1741     if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
1742       return LT.first * Entry->Cost;
1743 
1744   if (ST->hasAVX2())
1745     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
1746       return LT.first * Entry->Cost;
1747 
1748   if (ST->hasAVX())
1749     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
1750       return LT.first * Entry->Cost;
1751 
1752   if (ST->hasSSE42())
1753     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
1754       return LT.first * Entry->Cost;
1755 
1756   if (ST->hasSSSE3())
1757     if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy))
1758       return LT.first * Entry->Cost;
1759 
1760   if (ST->hasSSE2())
1761     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
1762       return LT.first * Entry->Cost;
1763 
1764   if (ST->hasSSE1())
1765     if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
1766       return LT.first * Entry->Cost;
1767 
1768   if (ST->is64Bit())
1769     if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
1770       return LT.first * Entry->Cost;
1771 
1772   if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
1773     return LT.first * Entry->Cost;
1774 
1775   return BaseT::getIntrinsicInstrCost(IID, RetTy, Tys, FMF, ScalarizationCostPassed);
1776 }
1777 
1778 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1779                      ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) {
1780   return BaseT::getIntrinsicInstrCost(IID, RetTy, Args, FMF, VF);
1781 }
1782 
1783 int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
1784   assert(Val->isVectorTy() && "This must be a vector type");
1785 
1786   Type *ScalarType = Val->getScalarType();
1787 
1788   if (Index != -1U) {
1789     // Legalize the type.
1790     std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Val);
1791 
1792     // This type is legalized to a scalar type.
1793     if (!LT.second.isVector())
1794       return 0;
1795 
1796     // The type may be split. Normalize the index to the new type.
1797     unsigned Width = LT.second.getVectorNumElements();
1798     Index = Index % Width;
1799 
1800     // Floating point scalars are already located in index #0.
1801     if (ScalarType->isFloatingPointTy() && Index == 0)
1802       return 0;
1803   }
1804 
1805   // Add to the base cost if we know that the extracted element of a vector is
1806   // destined to be moved to and used in the integer register file.
1807   int RegisterFileMoveCost = 0;
1808   if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy())
1809     RegisterFileMoveCost = 1;
1810 
1811   return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost;
1812 }
1813 
1814 int X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
1815                                 unsigned AddressSpace, const Instruction *I) {
1816   // Handle non-power-of-two vectors such as <3 x float>
1817   if (VectorType *VTy = dyn_cast<VectorType>(Src)) {
1818     unsigned NumElem = VTy->getVectorNumElements();
1819 
1820     // Handle a few common cases:
1821     // <3 x float>
1822     if (NumElem == 3 && VTy->getScalarSizeInBits() == 32)
1823       // Cost = 64 bit store + extract + 32 bit store.
1824       return 3;
1825 
1826     // <3 x double>
1827     if (NumElem == 3 && VTy->getScalarSizeInBits() == 64)
1828       // Cost = 128 bit store + unpack + 64 bit store.
1829       return 3;
1830 
1831     // Assume that all other non-power-of-two numbers are scalarized.
1832     if (!isPowerOf2_32(NumElem)) {
1833       int Cost = BaseT::getMemoryOpCost(Opcode, VTy->getScalarType(), Alignment,
1834                                         AddressSpace);
1835       int SplitCost = getScalarizationOverhead(Src, Opcode == Instruction::Load,
1836                                                Opcode == Instruction::Store);
1837       return NumElem * Cost + SplitCost;
1838     }
1839   }
1840 
1841   // Legalize the type.
1842   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
1843   assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
1844          "Invalid Opcode");
1845 
1846   // Each load/store unit costs 1.
1847   int Cost = LT.first * 1;
1848 
1849   // This isn't exactly right. We're using slow unaligned 32-byte accesses as a
1850   // proxy for a double-pumped AVX memory interface such as on Sandybridge.
1851   if (LT.second.getStoreSize() == 32 && ST->isUnalignedMem32Slow())
1852     Cost *= 2;
1853 
1854   return Cost;
1855 }
1856 
1857 int X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy,
1858                                       unsigned Alignment,
1859                                       unsigned AddressSpace) {
1860   VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy);
1861   if (!SrcVTy)
1862     // To calculate scalar take the regular cost, without mask
1863     return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace);
1864 
1865   unsigned NumElem = SrcVTy->getVectorNumElements();
1866   VectorType *MaskTy =
1867     VectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem);
1868   if ((Opcode == Instruction::Load && !isLegalMaskedLoad(SrcVTy)) ||
1869       (Opcode == Instruction::Store && !isLegalMaskedStore(SrcVTy)) ||
1870       !isPowerOf2_32(NumElem)) {
1871     // Scalarization
1872     int MaskSplitCost = getScalarizationOverhead(MaskTy, false, true);
1873     int ScalarCompareCost = getCmpSelInstrCost(
1874         Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr);
1875     int BranchCost = getCFInstrCost(Instruction::Br);
1876     int MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost);
1877 
1878     int ValueSplitCost = getScalarizationOverhead(
1879         SrcVTy, Opcode == Instruction::Load, Opcode == Instruction::Store);
1880     int MemopCost =
1881         NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
1882                                          Alignment, AddressSpace);
1883     return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost;
1884   }
1885 
1886   // Legalize the type.
1887   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy);
1888   auto VT = TLI->getValueType(DL, SrcVTy);
1889   int Cost = 0;
1890   if (VT.isSimple() && LT.second != VT.getSimpleVT() &&
1891       LT.second.getVectorNumElements() == NumElem)
1892     // Promotion requires expand/truncate for data and a shuffle for mask.
1893     Cost += getShuffleCost(TTI::SK_Alternate, SrcVTy, 0, nullptr) +
1894             getShuffleCost(TTI::SK_Alternate, MaskTy, 0, nullptr);
1895 
1896   else if (LT.second.getVectorNumElements() > NumElem) {
1897     VectorType *NewMaskTy = VectorType::get(MaskTy->getVectorElementType(),
1898                                             LT.second.getVectorNumElements());
1899     // Expanding requires fill mask with zeroes
1900     Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, 0, MaskTy);
1901   }
1902   if (!ST->hasAVX512())
1903     return Cost + LT.first*4; // Each maskmov costs 4
1904 
1905   // AVX-512 masked load/store is cheapper
1906   return Cost+LT.first;
1907 }
1908 
1909 int X86TTIImpl::getAddressComputationCost(Type *Ty, ScalarEvolution *SE,
1910                                           const SCEV *Ptr) {
1911   // Address computations in vectorized code with non-consecutive addresses will
1912   // likely result in more instructions compared to scalar code where the
1913   // computation can more often be merged into the index mode. The resulting
1914   // extra micro-ops can significantly decrease throughput.
1915   unsigned NumVectorInstToHideOverhead = 10;
1916 
1917   // Cost modeling of Strided Access Computation is hidden by the indexing
1918   // modes of X86 regardless of the stride value. We dont believe that there
1919   // is a difference between constant strided access in gerenal and constant
1920   // strided value which is less than or equal to 64.
1921   // Even in the case of (loop invariant) stride whose value is not known at
1922   // compile time, the address computation will not incur more than one extra
1923   // ADD instruction.
1924   if (Ty->isVectorTy() && SE) {
1925     if (!BaseT::isStridedAccess(Ptr))
1926       return NumVectorInstToHideOverhead;
1927     if (!BaseT::getConstantStrideStep(SE, Ptr))
1928       return 1;
1929   }
1930 
1931   return BaseT::getAddressComputationCost(Ty, SE, Ptr);
1932 }
1933 
1934 int X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, Type *ValTy,
1935                                            bool IsPairwise) {
1936 
1937   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
1938 
1939   MVT MTy = LT.second;
1940 
1941   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1942   assert(ISD && "Invalid opcode");
1943 
1944   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
1945   // and make it as the cost.
1946 
1947   static const CostTblEntry SSE42CostTblPairWise[] = {
1948     { ISD::FADD,  MVT::v2f64,   2 },
1949     { ISD::FADD,  MVT::v4f32,   4 },
1950     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
1951     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.5".
1952     { ISD::ADD,   MVT::v8i16,   5 },
1953   };
1954 
1955   static const CostTblEntry AVX1CostTblPairWise[] = {
1956     { ISD::FADD,  MVT::v4f32,   4 },
1957     { ISD::FADD,  MVT::v4f64,   5 },
1958     { ISD::FADD,  MVT::v8f32,   7 },
1959     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
1960     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.5".
1961     { ISD::ADD,   MVT::v4i64,   5 },      // The data reported by the IACA tool is "4.8".
1962     { ISD::ADD,   MVT::v8i16,   5 },
1963     { ISD::ADD,   MVT::v8i32,   5 },
1964   };
1965 
1966   static const CostTblEntry SSE42CostTblNoPairWise[] = {
1967     { ISD::FADD,  MVT::v2f64,   2 },
1968     { ISD::FADD,  MVT::v4f32,   4 },
1969     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
1970     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.3".
1971     { ISD::ADD,   MVT::v8i16,   4 },      // The data reported by the IACA tool is "4.3".
1972   };
1973 
1974   static const CostTblEntry AVX1CostTblNoPairWise[] = {
1975     { ISD::FADD,  MVT::v4f32,   3 },
1976     { ISD::FADD,  MVT::v4f64,   3 },
1977     { ISD::FADD,  MVT::v8f32,   4 },
1978     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
1979     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "2.8".
1980     { ISD::ADD,   MVT::v4i64,   3 },
1981     { ISD::ADD,   MVT::v8i16,   4 },
1982     { ISD::ADD,   MVT::v8i32,   5 },
1983   };
1984 
1985   if (IsPairwise) {
1986     if (ST->hasAVX())
1987       if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy))
1988         return LT.first * Entry->Cost;
1989 
1990     if (ST->hasSSE42())
1991       if (const auto *Entry = CostTableLookup(SSE42CostTblPairWise, ISD, MTy))
1992         return LT.first * Entry->Cost;
1993   } else {
1994     if (ST->hasAVX())
1995       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
1996         return LT.first * Entry->Cost;
1997 
1998     if (ST->hasSSE42())
1999       if (const auto *Entry = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy))
2000         return LT.first * Entry->Cost;
2001   }
2002 
2003   return BaseT::getArithmeticReductionCost(Opcode, ValTy, IsPairwise);
2004 }
2005 
2006 int X86TTIImpl::getMinMaxReductionCost(Type *ValTy, Type *CondTy,
2007                                        bool IsPairwise, bool IsUnsigned) {
2008   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
2009 
2010   MVT MTy = LT.second;
2011 
2012   int ISD;
2013   if (ValTy->isIntOrIntVectorTy()) {
2014     ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN;
2015   } else {
2016     assert(ValTy->isFPOrFPVectorTy() &&
2017            "Expected float point or integer vector type.");
2018     ISD = ISD::FMINNUM;
2019   }
2020 
2021   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
2022   // and make it as the cost.
2023 
2024   static const CostTblEntry SSE42CostTblPairWise[] = {
2025       {ISD::FMINNUM, MVT::v2f64, 3},
2026       {ISD::FMINNUM, MVT::v4f32, 2},
2027       {ISD::SMIN, MVT::v2i64, 7}, // The data reported by the IACA is "6.8"
2028       {ISD::UMIN, MVT::v2i64, 8}, // The data reported by the IACA is "8.6"
2029       {ISD::SMIN, MVT::v4i32, 1}, // The data reported by the IACA is "1.5"
2030       {ISD::UMIN, MVT::v4i32, 2}, // The data reported by the IACA is "1.8"
2031       {ISD::SMIN, MVT::v8i16, 2},
2032       {ISD::UMIN, MVT::v8i16, 2},
2033   };
2034 
2035   static const CostTblEntry AVX1CostTblPairWise[] = {
2036       {ISD::FMINNUM, MVT::v4f32, 1},
2037       {ISD::FMINNUM, MVT::v4f64, 1},
2038       {ISD::FMINNUM, MVT::v8f32, 2},
2039       {ISD::SMIN, MVT::v2i64, 3},
2040       {ISD::UMIN, MVT::v2i64, 3},
2041       {ISD::SMIN, MVT::v4i32, 1},
2042       {ISD::UMIN, MVT::v4i32, 1},
2043       {ISD::SMIN, MVT::v8i16, 1},
2044       {ISD::UMIN, MVT::v8i16, 1},
2045       {ISD::SMIN, MVT::v8i32, 3},
2046       {ISD::UMIN, MVT::v8i32, 3},
2047   };
2048 
2049   static const CostTblEntry AVX2CostTblPairWise[] = {
2050       {ISD::SMIN, MVT::v4i64, 2},
2051       {ISD::UMIN, MVT::v4i64, 2},
2052       {ISD::SMIN, MVT::v8i32, 1},
2053       {ISD::UMIN, MVT::v8i32, 1},
2054       {ISD::SMIN, MVT::v16i16, 1},
2055       {ISD::UMIN, MVT::v16i16, 1},
2056       {ISD::SMIN, MVT::v32i8, 2},
2057       {ISD::UMIN, MVT::v32i8, 2},
2058   };
2059 
2060   static const CostTblEntry AVX512CostTblPairWise[] = {
2061       {ISD::FMINNUM, MVT::v8f64, 1},
2062       {ISD::FMINNUM, MVT::v16f32, 2},
2063       {ISD::SMIN, MVT::v8i64, 2},
2064       {ISD::UMIN, MVT::v8i64, 2},
2065       {ISD::SMIN, MVT::v16i32, 1},
2066       {ISD::UMIN, MVT::v16i32, 1},
2067   };
2068 
2069   static const CostTblEntry SSE42CostTblNoPairWise[] = {
2070       {ISD::FMINNUM, MVT::v2f64, 3},
2071       {ISD::FMINNUM, MVT::v4f32, 3},
2072       {ISD::SMIN, MVT::v2i64, 7}, // The data reported by the IACA is "6.8"
2073       {ISD::UMIN, MVT::v2i64, 9}, // The data reported by the IACA is "8.6"
2074       {ISD::SMIN, MVT::v4i32, 1}, // The data reported by the IACA is "1.5"
2075       {ISD::UMIN, MVT::v4i32, 2}, // The data reported by the IACA is "1.8"
2076       {ISD::SMIN, MVT::v8i16, 1}, // The data reported by the IACA is "1.5"
2077       {ISD::UMIN, MVT::v8i16, 2}, // The data reported by the IACA is "1.8"
2078   };
2079 
2080   static const CostTblEntry AVX1CostTblNoPairWise[] = {
2081       {ISD::FMINNUM, MVT::v4f32, 1},
2082       {ISD::FMINNUM, MVT::v4f64, 1},
2083       {ISD::FMINNUM, MVT::v8f32, 1},
2084       {ISD::SMIN, MVT::v2i64, 3},
2085       {ISD::UMIN, MVT::v2i64, 3},
2086       {ISD::SMIN, MVT::v4i32, 1},
2087       {ISD::UMIN, MVT::v4i32, 1},
2088       {ISD::SMIN, MVT::v8i16, 1},
2089       {ISD::UMIN, MVT::v8i16, 1},
2090       {ISD::SMIN, MVT::v8i32, 2},
2091       {ISD::UMIN, MVT::v8i32, 2},
2092   };
2093 
2094   static const CostTblEntry AVX2CostTblNoPairWise[] = {
2095       {ISD::SMIN, MVT::v4i64, 1},
2096       {ISD::UMIN, MVT::v4i64, 1},
2097       {ISD::SMIN, MVT::v8i32, 1},
2098       {ISD::UMIN, MVT::v8i32, 1},
2099       {ISD::SMIN, MVT::v16i16, 1},
2100       {ISD::UMIN, MVT::v16i16, 1},
2101       {ISD::SMIN, MVT::v32i8, 1},
2102       {ISD::UMIN, MVT::v32i8, 1},
2103   };
2104 
2105   static const CostTblEntry AVX512CostTblNoPairWise[] = {
2106       {ISD::FMINNUM, MVT::v8f64, 1},
2107       {ISD::FMINNUM, MVT::v16f32, 2},
2108       {ISD::SMIN, MVT::v8i64, 1},
2109       {ISD::UMIN, MVT::v8i64, 1},
2110       {ISD::SMIN, MVT::v16i32, 1},
2111       {ISD::UMIN, MVT::v16i32, 1},
2112   };
2113 
2114   if (IsPairwise) {
2115     if (ST->hasAVX512())
2116       if (const auto *Entry = CostTableLookup(AVX512CostTblPairWise, ISD, MTy))
2117         return LT.first * Entry->Cost;
2118 
2119     if (ST->hasAVX2())
2120       if (const auto *Entry = CostTableLookup(AVX2CostTblPairWise, ISD, MTy))
2121         return LT.first * Entry->Cost;
2122 
2123     if (ST->hasAVX())
2124       if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy))
2125         return LT.first * Entry->Cost;
2126 
2127     if (ST->hasSSE42())
2128       if (const auto *Entry = CostTableLookup(SSE42CostTblPairWise, ISD, MTy))
2129         return LT.first * Entry->Cost;
2130   } else {
2131     if (ST->hasAVX512())
2132       if (const auto *Entry =
2133               CostTableLookup(AVX512CostTblNoPairWise, ISD, MTy))
2134         return LT.first * Entry->Cost;
2135 
2136     if (ST->hasAVX2())
2137       if (const auto *Entry = CostTableLookup(AVX2CostTblNoPairWise, ISD, MTy))
2138         return LT.first * Entry->Cost;
2139 
2140     if (ST->hasAVX())
2141       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
2142         return LT.first * Entry->Cost;
2143 
2144     if (ST->hasSSE42())
2145       if (const auto *Entry = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy))
2146         return LT.first * Entry->Cost;
2147   }
2148 
2149   return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsPairwise, IsUnsigned);
2150 }
2151 
2152 /// \brief Calculate the cost of materializing a 64-bit value. This helper
2153 /// method might only calculate a fraction of a larger immediate. Therefore it
2154 /// is valid to return a cost of ZERO.
2155 int X86TTIImpl::getIntImmCost(int64_t Val) {
2156   if (Val == 0)
2157     return TTI::TCC_Free;
2158 
2159   if (isInt<32>(Val))
2160     return TTI::TCC_Basic;
2161 
2162   return 2 * TTI::TCC_Basic;
2163 }
2164 
2165 int X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) {
2166   assert(Ty->isIntegerTy());
2167 
2168   unsigned BitSize = Ty->getPrimitiveSizeInBits();
2169   if (BitSize == 0)
2170     return ~0U;
2171 
2172   // Never hoist constants larger than 128bit, because this might lead to
2173   // incorrect code generation or assertions in codegen.
2174   // Fixme: Create a cost model for types larger than i128 once the codegen
2175   // issues have been fixed.
2176   if (BitSize > 128)
2177     return TTI::TCC_Free;
2178 
2179   if (Imm == 0)
2180     return TTI::TCC_Free;
2181 
2182   // Sign-extend all constants to a multiple of 64-bit.
2183   APInt ImmVal = Imm;
2184   if (BitSize & 0x3f)
2185     ImmVal = Imm.sext((BitSize + 63) & ~0x3fU);
2186 
2187   // Split the constant into 64-bit chunks and calculate the cost for each
2188   // chunk.
2189   int Cost = 0;
2190   for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
2191     APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
2192     int64_t Val = Tmp.getSExtValue();
2193     Cost += getIntImmCost(Val);
2194   }
2195   // We need at least one instruction to materialize the constant.
2196   return std::max(1, Cost);
2197 }
2198 
2199 int X86TTIImpl::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
2200                               Type *Ty) {
2201   assert(Ty->isIntegerTy());
2202 
2203   unsigned BitSize = Ty->getPrimitiveSizeInBits();
2204   // There is no cost model for constants with a bit size of 0. Return TCC_Free
2205   // here, so that constant hoisting will ignore this constant.
2206   if (BitSize == 0)
2207     return TTI::TCC_Free;
2208 
2209   unsigned ImmIdx = ~0U;
2210   switch (Opcode) {
2211   default:
2212     return TTI::TCC_Free;
2213   case Instruction::GetElementPtr:
2214     // Always hoist the base address of a GetElementPtr. This prevents the
2215     // creation of new constants for every base constant that gets constant
2216     // folded with the offset.
2217     if (Idx == 0)
2218       return 2 * TTI::TCC_Basic;
2219     return TTI::TCC_Free;
2220   case Instruction::Store:
2221     ImmIdx = 0;
2222     break;
2223   case Instruction::ICmp:
2224     // This is an imperfect hack to prevent constant hoisting of
2225     // compares that might be trying to check if a 64-bit value fits in
2226     // 32-bits. The backend can optimize these cases using a right shift by 32.
2227     // Ideally we would check the compare predicate here. There also other
2228     // similar immediates the backend can use shifts for.
2229     if (Idx == 1 && Imm.getBitWidth() == 64) {
2230       uint64_t ImmVal = Imm.getZExtValue();
2231       if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff)
2232         return TTI::TCC_Free;
2233     }
2234     ImmIdx = 1;
2235     break;
2236   case Instruction::And:
2237     // We support 64-bit ANDs with immediates with 32-bits of leading zeroes
2238     // by using a 32-bit operation with implicit zero extension. Detect such
2239     // immediates here as the normal path expects bit 31 to be sign extended.
2240     if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue()))
2241       return TTI::TCC_Free;
2242     LLVM_FALLTHROUGH;
2243   case Instruction::Add:
2244   case Instruction::Sub:
2245   case Instruction::Mul:
2246   case Instruction::UDiv:
2247   case Instruction::SDiv:
2248   case Instruction::URem:
2249   case Instruction::SRem:
2250   case Instruction::Or:
2251   case Instruction::Xor:
2252     ImmIdx = 1;
2253     break;
2254   // Always return TCC_Free for the shift value of a shift instruction.
2255   case Instruction::Shl:
2256   case Instruction::LShr:
2257   case Instruction::AShr:
2258     if (Idx == 1)
2259       return TTI::TCC_Free;
2260     break;
2261   case Instruction::Trunc:
2262   case Instruction::ZExt:
2263   case Instruction::SExt:
2264   case Instruction::IntToPtr:
2265   case Instruction::PtrToInt:
2266   case Instruction::BitCast:
2267   case Instruction::PHI:
2268   case Instruction::Call:
2269   case Instruction::Select:
2270   case Instruction::Ret:
2271   case Instruction::Load:
2272     break;
2273   }
2274 
2275   if (Idx == ImmIdx) {
2276     int NumConstants = (BitSize + 63) / 64;
2277     int Cost = X86TTIImpl::getIntImmCost(Imm, Ty);
2278     return (Cost <= NumConstants * TTI::TCC_Basic)
2279                ? static_cast<int>(TTI::TCC_Free)
2280                : Cost;
2281   }
2282 
2283   return X86TTIImpl::getIntImmCost(Imm, Ty);
2284 }
2285 
2286 int X86TTIImpl::getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
2287                               Type *Ty) {
2288   assert(Ty->isIntegerTy());
2289 
2290   unsigned BitSize = Ty->getPrimitiveSizeInBits();
2291   // There is no cost model for constants with a bit size of 0. Return TCC_Free
2292   // here, so that constant hoisting will ignore this constant.
2293   if (BitSize == 0)
2294     return TTI::TCC_Free;
2295 
2296   switch (IID) {
2297   default:
2298     return TTI::TCC_Free;
2299   case Intrinsic::sadd_with_overflow:
2300   case Intrinsic::uadd_with_overflow:
2301   case Intrinsic::ssub_with_overflow:
2302   case Intrinsic::usub_with_overflow:
2303   case Intrinsic::smul_with_overflow:
2304   case Intrinsic::umul_with_overflow:
2305     if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
2306       return TTI::TCC_Free;
2307     break;
2308   case Intrinsic::experimental_stackmap:
2309     if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
2310       return TTI::TCC_Free;
2311     break;
2312   case Intrinsic::experimental_patchpoint_void:
2313   case Intrinsic::experimental_patchpoint_i64:
2314     if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
2315       return TTI::TCC_Free;
2316     break;
2317   }
2318   return X86TTIImpl::getIntImmCost(Imm, Ty);
2319 }
2320 
2321 unsigned X86TTIImpl::getUserCost(const User *U,
2322                                  ArrayRef<const Value *> Operands) {
2323   if (isa<StoreInst>(U)) {
2324     Value *Ptr = U->getOperand(1);
2325     // Store instruction with index and scale costs 2 Uops.
2326     // Check the preceding GEP to identify non-const indices.
2327     if (auto GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
2328       if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); }))
2329         return TTI::TCC_Basic * 2;
2330     }
2331     return TTI::TCC_Basic;
2332   }
2333   return BaseT::getUserCost(U, Operands);
2334 }
2335 
2336 // Return an average cost of Gather / Scatter instruction, maybe improved later
2337 int X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy, Value *Ptr,
2338                                 unsigned Alignment, unsigned AddressSpace) {
2339 
2340   assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost");
2341   unsigned VF = SrcVTy->getVectorNumElements();
2342 
2343   // Try to reduce index size from 64 bit (default for GEP)
2344   // to 32. It is essential for VF 16. If the index can't be reduced to 32, the
2345   // operation will use 16 x 64 indices which do not fit in a zmm and needs
2346   // to split. Also check that the base pointer is the same for all lanes,
2347   // and that there's at most one variable index.
2348   auto getIndexSizeInBits = [](Value *Ptr, const DataLayout& DL) {
2349     unsigned IndexSize = DL.getPointerSizeInBits();
2350     GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
2351     if (IndexSize < 64 || !GEP)
2352       return IndexSize;
2353 
2354     unsigned NumOfVarIndices = 0;
2355     Value *Ptrs = GEP->getPointerOperand();
2356     if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs))
2357       return IndexSize;
2358     for (unsigned i = 1; i < GEP->getNumOperands(); ++i) {
2359       if (isa<Constant>(GEP->getOperand(i)))
2360         continue;
2361       Type *IndxTy = GEP->getOperand(i)->getType();
2362       if (IndxTy->isVectorTy())
2363         IndxTy = IndxTy->getVectorElementType();
2364       if ((IndxTy->getPrimitiveSizeInBits() == 64 &&
2365           !isa<SExtInst>(GEP->getOperand(i))) ||
2366          ++NumOfVarIndices > 1)
2367         return IndexSize; // 64
2368     }
2369     return (unsigned)32;
2370   };
2371 
2372 
2373   // Trying to reduce IndexSize to 32 bits for vector 16.
2374   // By default the IndexSize is equal to pointer size.
2375   unsigned IndexSize = (ST->hasAVX512() && VF >= 16)
2376                            ? getIndexSizeInBits(Ptr, DL)
2377                            : DL.getPointerSizeInBits();
2378 
2379   Type *IndexVTy = VectorType::get(IntegerType::get(SrcVTy->getContext(),
2380                                                     IndexSize), VF);
2381   std::pair<int, MVT> IdxsLT = TLI->getTypeLegalizationCost(DL, IndexVTy);
2382   std::pair<int, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, SrcVTy);
2383   int SplitFactor = std::max(IdxsLT.first, SrcLT.first);
2384   if (SplitFactor > 1) {
2385     // Handle splitting of vector of pointers
2386     Type *SplitSrcTy = VectorType::get(SrcVTy->getScalarType(), VF / SplitFactor);
2387     return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment,
2388                                          AddressSpace);
2389   }
2390 
2391   // The gather / scatter cost is given by Intel architects. It is a rough
2392   // number since we are looking at one instruction in a time.
2393   const int GSOverhead = (Opcode == Instruction::Load)
2394                              ? ST->getGatherOverhead()
2395                              : ST->getScatterOverhead();
2396   return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
2397                                            Alignment, AddressSpace);
2398 }
2399 
2400 /// Return the cost of full scalarization of gather / scatter operation.
2401 ///
2402 /// Opcode - Load or Store instruction.
2403 /// SrcVTy - The type of the data vector that should be gathered or scattered.
2404 /// VariableMask - The mask is non-constant at compile time.
2405 /// Alignment - Alignment for one element.
2406 /// AddressSpace - pointer[s] address space.
2407 ///
2408 int X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy,
2409                                 bool VariableMask, unsigned Alignment,
2410                                 unsigned AddressSpace) {
2411   unsigned VF = SrcVTy->getVectorNumElements();
2412 
2413   int MaskUnpackCost = 0;
2414   if (VariableMask) {
2415     VectorType *MaskTy =
2416       VectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF);
2417     MaskUnpackCost = getScalarizationOverhead(MaskTy, false, true);
2418     int ScalarCompareCost =
2419       getCmpSelInstrCost(Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()),
2420                          nullptr);
2421     int BranchCost = getCFInstrCost(Instruction::Br);
2422     MaskUnpackCost += VF * (BranchCost + ScalarCompareCost);
2423   }
2424 
2425   // The cost of the scalar loads/stores.
2426   int MemoryOpCost = VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
2427                                           Alignment, AddressSpace);
2428 
2429   int InsertExtractCost = 0;
2430   if (Opcode == Instruction::Load)
2431     for (unsigned i = 0; i < VF; ++i)
2432       // Add the cost of inserting each scalar load into the vector
2433       InsertExtractCost +=
2434         getVectorInstrCost(Instruction::InsertElement, SrcVTy, i);
2435   else
2436     for (unsigned i = 0; i < VF; ++i)
2437       // Add the cost of extracting each element out of the data vector
2438       InsertExtractCost +=
2439         getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i);
2440 
2441   return MemoryOpCost + MaskUnpackCost + InsertExtractCost;
2442 }
2443 
2444 /// Calculate the cost of Gather / Scatter operation
2445 int X86TTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *SrcVTy,
2446                                        Value *Ptr, bool VariableMask,
2447                                        unsigned Alignment) {
2448   assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter");
2449   unsigned VF = SrcVTy->getVectorNumElements();
2450   PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
2451   if (!PtrTy && Ptr->getType()->isVectorTy())
2452     PtrTy = dyn_cast<PointerType>(Ptr->getType()->getVectorElementType());
2453   assert(PtrTy && "Unexpected type for Ptr argument");
2454   unsigned AddressSpace = PtrTy->getAddressSpace();
2455 
2456   bool Scalarize = false;
2457   if ((Opcode == Instruction::Load && !isLegalMaskedGather(SrcVTy)) ||
2458       (Opcode == Instruction::Store && !isLegalMaskedScatter(SrcVTy)))
2459     Scalarize = true;
2460   // Gather / Scatter for vector 2 is not profitable on KNL / SKX
2461   // Vector-4 of gather/scatter instruction does not exist on KNL.
2462   // We can extend it to 8 elements, but zeroing upper bits of
2463   // the mask vector will add more instructions. Right now we give the scalar
2464   // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter instruction
2465   // is better in the VariableMask case.
2466   if (ST->hasAVX512() && (VF == 2 || (VF == 4 && !ST->hasVLX())))
2467     Scalarize = true;
2468 
2469   if (Scalarize)
2470     return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment,
2471                            AddressSpace);
2472 
2473   return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace);
2474 }
2475 
2476 bool X86TTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1,
2477                                TargetTransformInfo::LSRCost &C2) {
2478     // X86 specific here are "instruction number 1st priority".
2479     return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost,
2480                     C1.NumIVMuls, C1.NumBaseAdds,
2481                     C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
2482            std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost,
2483                     C2.NumIVMuls, C2.NumBaseAdds,
2484                     C2.ScaleCost, C2.ImmCost, C2.SetupCost);
2485 }
2486 
2487 bool X86TTIImpl::canMacroFuseCmp() {
2488   return ST->hasMacroFusion();
2489 }
2490 
2491 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy) {
2492   // The backend can't handle a single element vector.
2493   if (isa<VectorType>(DataTy) && DataTy->getVectorNumElements() == 1)
2494     return false;
2495   Type *ScalarTy = DataTy->getScalarType();
2496   int DataWidth = isa<PointerType>(ScalarTy) ?
2497     DL.getPointerSizeInBits() : ScalarTy->getPrimitiveSizeInBits();
2498 
2499   return ((DataWidth == 32 || DataWidth == 64) && ST->hasAVX()) ||
2500          ((DataWidth == 8 || DataWidth == 16) && ST->hasBWI());
2501 }
2502 
2503 bool X86TTIImpl::isLegalMaskedStore(Type *DataType) {
2504   return isLegalMaskedLoad(DataType);
2505 }
2506 
2507 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy) {
2508   // This function is called now in two cases: from the Loop Vectorizer
2509   // and from the Scalarizer.
2510   // When the Loop Vectorizer asks about legality of the feature,
2511   // the vectorization factor is not calculated yet. The Loop Vectorizer
2512   // sends a scalar type and the decision is based on the width of the
2513   // scalar element.
2514   // Later on, the cost model will estimate usage this intrinsic based on
2515   // the vector type.
2516   // The Scalarizer asks again about legality. It sends a vector type.
2517   // In this case we can reject non-power-of-2 vectors.
2518   // We also reject single element vectors as the type legalizer can't
2519   // scalarize it.
2520   if (isa<VectorType>(DataTy)) {
2521     unsigned NumElts = DataTy->getVectorNumElements();
2522     if (NumElts == 1 || !isPowerOf2_32(NumElts))
2523       return false;
2524   }
2525   Type *ScalarTy = DataTy->getScalarType();
2526   int DataWidth = isa<PointerType>(ScalarTy) ?
2527     DL.getPointerSizeInBits() : ScalarTy->getPrimitiveSizeInBits();
2528 
2529   // Some CPUs have better gather performance than others.
2530   // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only
2531   // enable gather with a -march.
2532   return (DataWidth == 32 || DataWidth == 64) &&
2533          (ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2()));
2534 }
2535 
2536 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType) {
2537   // AVX2 doesn't support scatter
2538   if (!ST->hasAVX512())
2539     return false;
2540   return isLegalMaskedGather(DataType);
2541 }
2542 
2543 bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) {
2544   EVT VT = TLI->getValueType(DL, DataType);
2545   return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT);
2546 }
2547 
2548 bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
2549   return false;
2550 }
2551 
2552 bool X86TTIImpl::areInlineCompatible(const Function *Caller,
2553                                      const Function *Callee) const {
2554   const TargetMachine &TM = getTLI()->getTargetMachine();
2555 
2556   // Work this as a subsetting of subtarget features.
2557   const FeatureBitset &CallerBits =
2558       TM.getSubtargetImpl(*Caller)->getFeatureBits();
2559   const FeatureBitset &CalleeBits =
2560       TM.getSubtargetImpl(*Callee)->getFeatureBits();
2561 
2562   // FIXME: This is likely too limiting as it will include subtarget features
2563   // that we might not care about for inlining, but it is conservatively
2564   // correct.
2565   return (CallerBits & CalleeBits) == CalleeBits;
2566 }
2567 
2568 const X86TTIImpl::TTI::MemCmpExpansionOptions *
2569 X86TTIImpl::enableMemCmpExpansion(bool IsZeroCmp) const {
2570   // Only enable vector loads for equality comparison.
2571   // Right now the vector version is not as fast, see #33329.
2572   static const auto ThreeWayOptions = [this]() {
2573     TTI::MemCmpExpansionOptions Options;
2574     if (ST->is64Bit()) {
2575       Options.LoadSizes.push_back(8);
2576     }
2577     Options.LoadSizes.push_back(4);
2578     Options.LoadSizes.push_back(2);
2579     Options.LoadSizes.push_back(1);
2580     return Options;
2581   }();
2582   static const auto EqZeroOptions = [this]() {
2583     TTI::MemCmpExpansionOptions Options;
2584     // TODO: enable AVX512 when the DAG is ready.
2585     // if (ST->hasAVX512()) Options.LoadSizes.push_back(64);
2586     if (ST->hasAVX2()) Options.LoadSizes.push_back(32);
2587     if (ST->hasSSE2()) Options.LoadSizes.push_back(16);
2588     if (ST->is64Bit()) {
2589       Options.LoadSizes.push_back(8);
2590     }
2591     Options.LoadSizes.push_back(4);
2592     Options.LoadSizes.push_back(2);
2593     Options.LoadSizes.push_back(1);
2594     return Options;
2595   }();
2596   return IsZeroCmp ? &EqZeroOptions : &ThreeWayOptions;
2597 }
2598 
2599 bool X86TTIImpl::enableInterleavedAccessVectorization() {
2600   // TODO: We expect this to be beneficial regardless of arch,
2601   // but there are currently some unexplained performance artifacts on Atom.
2602   // As a temporary solution, disable on Atom.
2603   return !(ST->isAtom());
2604 }
2605 
2606 // Get estimation for interleaved load/store operations for AVX2.
2607 // \p Factor is the interleaved-access factor (stride) - number of
2608 // (interleaved) elements in the group.
2609 // \p Indices contains the indices for a strided load: when the
2610 // interleaved load has gaps they indicate which elements are used.
2611 // If Indices is empty (or if the number of indices is equal to the size
2612 // of the interleaved-access as given in \p Factor) the access has no gaps.
2613 //
2614 // As opposed to AVX-512, AVX2 does not have generic shuffles that allow
2615 // computing the cost using a generic formula as a function of generic
2616 // shuffles. We therefore use a lookup table instead, filled according to
2617 // the instruction sequences that codegen currently generates.
2618 int X86TTIImpl::getInterleavedMemoryOpCostAVX2(unsigned Opcode, Type *VecTy,
2619                                                unsigned Factor,
2620                                                ArrayRef<unsigned> Indices,
2621                                                unsigned Alignment,
2622                                                unsigned AddressSpace) {
2623 
2624   // We currently Support only fully-interleaved groups, with no gaps.
2625   // TODO: Support also strided loads (interleaved-groups with gaps).
2626   if (Indices.size() && Indices.size() != Factor)
2627     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2628                                              Alignment, AddressSpace);
2629 
2630   // VecTy for interleave memop is <VF*Factor x Elt>.
2631   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
2632   // VecTy = <12 x i32>.
2633   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
2634 
2635   // This function can be called with VecTy=<6xi128>, Factor=3, in which case
2636   // the VF=2, while v2i128 is an unsupported MVT vector type
2637   // (see MachineValueType.h::getVectorVT()).
2638   if (!LegalVT.isVector())
2639     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2640                                              Alignment, AddressSpace);
2641 
2642   unsigned VF = VecTy->getVectorNumElements() / Factor;
2643   Type *ScalarTy = VecTy->getVectorElementType();
2644 
2645   // Calculate the number of memory operations (NumOfMemOps), required
2646   // for load/store the VecTy.
2647   unsigned VecTySize = DL.getTypeStoreSize(VecTy);
2648   unsigned LegalVTSize = LegalVT.getStoreSize();
2649   unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize;
2650 
2651   // Get the cost of one memory operation.
2652   Type *SingleMemOpTy = VectorType::get(VecTy->getVectorElementType(),
2653                                         LegalVT.getVectorNumElements());
2654   unsigned MemOpCost =
2655       getMemoryOpCost(Opcode, SingleMemOpTy, Alignment, AddressSpace);
2656 
2657   VectorType *VT = VectorType::get(ScalarTy, VF);
2658   EVT ETy = TLI->getValueType(DL, VT);
2659   if (!ETy.isSimple())
2660     return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2661                                              Alignment, AddressSpace);
2662 
2663   // TODO: Complete for other data-types and strides.
2664   // Each combination of Stride, ElementTy and VF results in a different
2665   // sequence; The cost tables are therefore accessed with:
2666   // Factor (stride) and VectorType=VFxElemType.
2667   // The Cost accounts only for the shuffle sequence;
2668   // The cost of the loads/stores is accounted for separately.
2669   //
2670   static const CostTblEntry AVX2InterleavedLoadTbl[] = {
2671     { 2, MVT::v4i64, 6 }, //(load 8i64 and) deinterleave into 2 x 4i64
2672     { 2, MVT::v4f64, 6 }, //(load 8f64 and) deinterleave into 2 x 4f64
2673 
2674     { 3, MVT::v2i8,  10 }, //(load 6i8 and)  deinterleave into 3 x 2i8
2675     { 3, MVT::v4i8,  4 },  //(load 12i8 and) deinterleave into 3 x 4i8
2676     { 3, MVT::v8i8,  9 },  //(load 24i8 and) deinterleave into 3 x 8i8
2677     { 3, MVT::v16i8, 11},  //(load 48i8 and) deinterleave into 3 x 16i8
2678     { 3, MVT::v32i8, 13},  //(load 96i8 and) deinterleave into 3 x 32i8
2679     { 3, MVT::v8f32, 17 }, //(load 24f32 and)deinterleave into 3 x 8f32
2680 
2681     { 4, MVT::v2i8,  12 }, //(load 8i8 and)   deinterleave into 4 x 2i8
2682     { 4, MVT::v4i8,  4 },  //(load 16i8 and)  deinterleave into 4 x 4i8
2683     { 4, MVT::v8i8,  20 }, //(load 32i8 and)  deinterleave into 4 x 8i8
2684     { 4, MVT::v16i8, 39 }, //(load 64i8 and)  deinterleave into 4 x 16i8
2685     { 4, MVT::v32i8, 80 }, //(load 128i8 and) deinterleave into 4 x 32i8
2686 
2687     { 8, MVT::v8f32, 40 }  //(load 64f32 and)deinterleave into 8 x 8f32
2688   };
2689 
2690   static const CostTblEntry AVX2InterleavedStoreTbl[] = {
2691     { 2, MVT::v4i64, 6 }, //interleave into 2 x 4i64 into 8i64 (and store)
2692     { 2, MVT::v4f64, 6 }, //interleave into 2 x 4f64 into 8f64 (and store)
2693 
2694     { 3, MVT::v2i8,  7 },  //interleave 3 x 2i8  into 6i8 (and store)
2695     { 3, MVT::v4i8,  8 },  //interleave 3 x 4i8  into 12i8 (and store)
2696     { 3, MVT::v8i8,  11 }, //interleave 3 x 8i8  into 24i8 (and store)
2697     { 3, MVT::v16i8, 11 }, //interleave 3 x 16i8 into 48i8 (and store)
2698     { 3, MVT::v32i8, 13 }, //interleave 3 x 32i8 into 96i8 (and store)
2699 
2700     { 4, MVT::v2i8,  12 }, //interleave 4 x 2i8  into 8i8 (and store)
2701     { 4, MVT::v4i8,  9 },  //interleave 4 x 4i8  into 16i8 (and store)
2702     { 4, MVT::v8i8,  10 }, //interleave 4 x 8i8  into 32i8 (and store)
2703     { 4, MVT::v16i8, 10 }, //interleave 4 x 16i8 into 64i8 (and store)
2704     { 4, MVT::v32i8, 12 }  //interleave 4 x 32i8 into 128i8 (and store)
2705   };
2706 
2707   if (Opcode == Instruction::Load) {
2708     if (const auto *Entry =
2709             CostTableLookup(AVX2InterleavedLoadTbl, Factor, ETy.getSimpleVT()))
2710       return NumOfMemOps * MemOpCost + Entry->Cost;
2711   } else {
2712     assert(Opcode == Instruction::Store &&
2713            "Expected Store Instruction at this  point");
2714     if (const auto *Entry =
2715             CostTableLookup(AVX2InterleavedStoreTbl, Factor, ETy.getSimpleVT()))
2716       return NumOfMemOps * MemOpCost + Entry->Cost;
2717   }
2718 
2719   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2720                                            Alignment, AddressSpace);
2721 }
2722 
2723 // Get estimation for interleaved load/store operations and strided load.
2724 // \p Indices contains indices for strided load.
2725 // \p Factor - the factor of interleaving.
2726 // AVX-512 provides 3-src shuffles that significantly reduces the cost.
2727 int X86TTIImpl::getInterleavedMemoryOpCostAVX512(unsigned Opcode, Type *VecTy,
2728                                                  unsigned Factor,
2729                                                  ArrayRef<unsigned> Indices,
2730                                                  unsigned Alignment,
2731                                                  unsigned AddressSpace) {
2732 
2733   // VecTy for interleave memop is <VF*Factor x Elt>.
2734   // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
2735   // VecTy = <12 x i32>.
2736 
2737   // Calculate the number of memory operations (NumOfMemOps), required
2738   // for load/store the VecTy.
2739   MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
2740   unsigned VecTySize = DL.getTypeStoreSize(VecTy);
2741   unsigned LegalVTSize = LegalVT.getStoreSize();
2742   unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize;
2743 
2744   // Get the cost of one memory operation.
2745   Type *SingleMemOpTy = VectorType::get(VecTy->getVectorElementType(),
2746                                         LegalVT.getVectorNumElements());
2747   unsigned MemOpCost =
2748       getMemoryOpCost(Opcode, SingleMemOpTy, Alignment, AddressSpace);
2749 
2750   unsigned VF = VecTy->getVectorNumElements() / Factor;
2751   MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF);
2752 
2753   if (Opcode == Instruction::Load) {
2754     // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl)
2755     // contain the cost of the optimized shuffle sequence that the
2756     // X86InterleavedAccess pass will generate.
2757     // The cost of loads and stores are computed separately from the table.
2758 
2759     // X86InterleavedAccess support only the following interleaved-access group.
2760     static const CostTblEntry AVX512InterleavedLoadTbl[] = {
2761         {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8
2762         {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8
2763         {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8
2764     };
2765 
2766     if (const auto *Entry =
2767             CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT))
2768       return NumOfMemOps * MemOpCost + Entry->Cost;
2769     //If an entry does not exist, fallback to the default implementation.
2770 
2771     // Kind of shuffle depends on number of loaded values.
2772     // If we load the entire data in one register, we can use a 1-src shuffle.
2773     // Otherwise, we'll merge 2 sources in each operation.
2774     TTI::ShuffleKind ShuffleKind =
2775         (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc;
2776 
2777     unsigned ShuffleCost =
2778         getShuffleCost(ShuffleKind, SingleMemOpTy, 0, nullptr);
2779 
2780     unsigned NumOfLoadsInInterleaveGrp =
2781         Indices.size() ? Indices.size() : Factor;
2782     Type *ResultTy = VectorType::get(VecTy->getVectorElementType(),
2783                                      VecTy->getVectorNumElements() / Factor);
2784     unsigned NumOfResults =
2785         getTLI()->getTypeLegalizationCost(DL, ResultTy).first *
2786         NumOfLoadsInInterleaveGrp;
2787 
2788     // About a half of the loads may be folded in shuffles when we have only
2789     // one result. If we have more than one result, we do not fold loads at all.
2790     unsigned NumOfUnfoldedLoads =
2791         NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2;
2792 
2793     // Get a number of shuffle operations per result.
2794     unsigned NumOfShufflesPerResult =
2795         std::max((unsigned)1, (unsigned)(NumOfMemOps - 1));
2796 
2797     // The SK_MergeTwoSrc shuffle clobbers one of src operands.
2798     // When we have more than one destination, we need additional instructions
2799     // to keep sources.
2800     unsigned NumOfMoves = 0;
2801     if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc)
2802       NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2;
2803 
2804     int Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost +
2805                NumOfUnfoldedLoads * MemOpCost + NumOfMoves;
2806 
2807     return Cost;
2808   }
2809 
2810   // Store.
2811   assert(Opcode == Instruction::Store &&
2812          "Expected Store Instruction at this  point");
2813   // X86InterleavedAccess support only the following interleaved-access group.
2814   static const CostTblEntry AVX512InterleavedStoreTbl[] = {
2815       {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store)
2816       {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store)
2817       {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store)
2818 
2819       {4, MVT::v8i8, 10},  // interleave 4 x 8i8  into 32i8  (and store)
2820       {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8  (and store)
2821       {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store)
2822       {4, MVT::v64i8, 24}  // interleave 4 x 32i8 into 256i8 (and store)
2823   };
2824 
2825   if (const auto *Entry =
2826           CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT))
2827     return NumOfMemOps * MemOpCost + Entry->Cost;
2828   //If an entry does not exist, fallback to the default implementation.
2829 
2830   // There is no strided stores meanwhile. And store can't be folded in
2831   // shuffle.
2832   unsigned NumOfSources = Factor; // The number of values to be merged.
2833   unsigned ShuffleCost =
2834       getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, 0, nullptr);
2835   unsigned NumOfShufflesPerStore = NumOfSources - 1;
2836 
2837   // The SK_MergeTwoSrc shuffle clobbers one of src operands.
2838   // We need additional instructions to keep sources.
2839   unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2;
2840   int Cost = NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) +
2841              NumOfMoves;
2842   return Cost;
2843 }
2844 
2845 int X86TTIImpl::getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
2846                                            unsigned Factor,
2847                                            ArrayRef<unsigned> Indices,
2848                                            unsigned Alignment,
2849                                            unsigned AddressSpace) {
2850   auto isSupportedOnAVX512 = [](Type *VecTy, bool HasBW) {
2851     Type *EltTy = VecTy->getVectorElementType();
2852     if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) ||
2853         EltTy->isIntegerTy(32) || EltTy->isPointerTy())
2854       return true;
2855     if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8))
2856       return HasBW;
2857     return false;
2858   };
2859   if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI()))
2860     return getInterleavedMemoryOpCostAVX512(Opcode, VecTy, Factor, Indices,
2861                                             Alignment, AddressSpace);
2862   if (ST->hasAVX2())
2863     return getInterleavedMemoryOpCostAVX2(Opcode, VecTy, Factor, Indices,
2864                                           Alignment, AddressSpace);
2865 
2866   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
2867                                            Alignment, AddressSpace);
2868 }
2869