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/IR/IntrinsicInst.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Target/CostTable.h"
48 #include "llvm/Target/TargetLowering.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 unsigned X86TTIImpl::getNumberOfRegisters(bool Vector) {
70   if (Vector && !ST->hasSSE1())
71     return 0;
72 
73   if (ST->is64Bit()) {
74     if (Vector && ST->hasAVX512())
75       return 32;
76     return 16;
77   }
78   return 8;
79 }
80 
81 unsigned X86TTIImpl::getRegisterBitWidth(bool Vector) {
82   if (Vector) {
83     if (ST->hasAVX512()) return 512;
84     if (ST->hasAVX()) return 256;
85     if (ST->hasSSE1()) return 128;
86     return 0;
87   }
88 
89   if (ST->is64Bit())
90     return 64;
91 
92   return 32;
93 }
94 
95 unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) {
96   // If the loop will not be vectorized, don't interleave the loop.
97   // Let regular unroll to unroll the loop, which saves the overflow
98   // check and memory check cost.
99   if (VF == 1)
100     return 1;
101 
102   if (ST->isAtom())
103     return 1;
104 
105   // Sandybridge and Haswell have multiple execution ports and pipelined
106   // vector units.
107   if (ST->hasAVX())
108     return 4;
109 
110   return 2;
111 }
112 
113 int X86TTIImpl::getArithmeticInstrCost(
114     unsigned Opcode, Type *Ty, TTI::OperandValueKind Op1Info,
115     TTI::OperandValueKind Op2Info, TTI::OperandValueProperties Opd1PropInfo,
116     TTI::OperandValueProperties Opd2PropInfo) {
117   // Legalize the type.
118   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
119 
120   int ISD = TLI->InstructionOpcodeToISD(Opcode);
121   assert(ISD && "Invalid opcode");
122 
123   if (ISD == ISD::SDIV &&
124       Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
125       Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
126     // On X86, vector signed division by constants power-of-two are
127     // normally expanded to the sequence SRA + SRL + ADD + SRA.
128     // The OperandValue properties many not be same as that of previous
129     // operation;conservatively assume OP_None.
130     int Cost = 2 * getArithmeticInstrCost(Instruction::AShr, Ty, Op1Info,
131                                           Op2Info, TargetTransformInfo::OP_None,
132                                           TargetTransformInfo::OP_None);
133     Cost += getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info,
134                                    TargetTransformInfo::OP_None,
135                                    TargetTransformInfo::OP_None);
136     Cost += getArithmeticInstrCost(Instruction::Add, Ty, Op1Info, Op2Info,
137                                    TargetTransformInfo::OP_None,
138                                    TargetTransformInfo::OP_None);
139 
140     return Cost;
141   }
142 
143   static const CostTblEntry AVX512BWUniformConstCostTable[] = {
144     { ISD::SDIV, MVT::v32i16,  6 }, // vpmulhw sequence
145     { ISD::UDIV, MVT::v32i16,  6 }, // vpmulhuw sequence
146   };
147 
148   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
149       ST->hasBWI()) {
150     if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD,
151                                             LT.second))
152       return LT.first * Entry->Cost;
153   }
154 
155   static const CostTblEntry AVX512UniformConstCostTable[] = {
156     { ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence
157     { ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence
158   };
159 
160   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
161       ST->hasAVX512()) {
162     if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD,
163                                             LT.second))
164       return LT.first * Entry->Cost;
165   }
166 
167   static const CostTblEntry AVX2UniformConstCostTable[] = {
168     { ISD::SRA,  MVT::v4i64,   4 }, // 2 x psrad + shuffle.
169 
170     { ISD::SDIV, MVT::v16i16,  6 }, // vpmulhw sequence
171     { ISD::UDIV, MVT::v16i16,  6 }, // vpmulhuw sequence
172     { ISD::SDIV, MVT::v8i32,  15 }, // vpmuldq sequence
173     { ISD::UDIV, MVT::v8i32,  15 }, // vpmuludq sequence
174   };
175 
176   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
177       ST->hasAVX2()) {
178     if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD,
179                                             LT.second))
180       return LT.first * Entry->Cost;
181   }
182 
183   static const CostTblEntry SSE2UniformConstCostTable[] = {
184     { ISD::SDIV, MVT::v16i16, 12 }, // pmulhw sequence
185     { ISD::SDIV, MVT::v8i16,   6 }, // pmulhw sequence
186     { ISD::UDIV, MVT::v16i16, 12 }, // pmulhuw sequence
187     { ISD::UDIV, MVT::v8i16,   6 }, // pmulhuw sequence
188     { ISD::SDIV, MVT::v8i32,  38 }, // pmuludq sequence
189     { ISD::SDIV, MVT::v4i32,  19 }, // pmuludq sequence
190     { ISD::UDIV, MVT::v8i32,  30 }, // pmuludq sequence
191     { ISD::UDIV, MVT::v4i32,  15 }, // pmuludq sequence
192   };
193 
194   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
195       ST->hasSSE2()) {
196     // pmuldq sequence.
197     if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX())
198       return LT.first * 30;
199     if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
200       return LT.first * 15;
201 
202     if (const auto *Entry = CostTableLookup(SSE2UniformConstCostTable, ISD,
203                                             LT.second))
204       return LT.first * Entry->Cost;
205   }
206 
207   static const CostTblEntry AVX512BWCostTable[] = {
208     // Vectorizing division is a bad idea. See the SSE2 table for more comments.
209     { ISD::SDIV,  MVT::v64i8,  64*20 },
210     { ISD::SDIV,  MVT::v32i16, 32*20 },
211     { ISD::SDIV,  MVT::v16i32, 16*20 },
212     { ISD::SDIV,  MVT::v8i64,  8*20 },
213     { ISD::UDIV,  MVT::v64i8,  64*20 },
214     { ISD::UDIV,  MVT::v32i16, 32*20 },
215     { ISD::UDIV,  MVT::v16i32, 16*20 },
216     { ISD::UDIV,  MVT::v8i64,  8*20 },
217   };
218 
219   // Look for AVX512BW lowering tricks for custom cases.
220   if (ST->hasBWI()) {
221     if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD,
222                                             LT.second))
223       return LT.first * Entry->Cost;
224   }
225 
226   static const CostTblEntry AVX512CostTable[] = {
227     { ISD::SHL,     MVT::v16i32,    1 },
228     { ISD::SRL,     MVT::v16i32,    1 },
229     { ISD::SRA,     MVT::v16i32,    1 },
230     { ISD::SHL,     MVT::v8i64,    1 },
231     { ISD::SRL,     MVT::v8i64,    1 },
232     { ISD::SRA,     MVT::v8i64,    1 },
233   };
234 
235   if (ST->hasAVX512()) {
236     if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second))
237       return LT.first * Entry->Cost;
238   }
239 
240   static const CostTblEntry AVX2CostTable[] = {
241     // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
242     // customize them to detect the cases where shift amount is a scalar one.
243     { ISD::SHL,     MVT::v4i32,    1 },
244     { ISD::SRL,     MVT::v4i32,    1 },
245     { ISD::SRA,     MVT::v4i32,    1 },
246     { ISD::SHL,     MVT::v8i32,    1 },
247     { ISD::SRL,     MVT::v8i32,    1 },
248     { ISD::SRA,     MVT::v8i32,    1 },
249     { ISD::SHL,     MVT::v2i64,    1 },
250     { ISD::SRL,     MVT::v2i64,    1 },
251     { ISD::SHL,     MVT::v4i64,    1 },
252     { ISD::SRL,     MVT::v4i64,    1 },
253   };
254 
255   // Look for AVX2 lowering tricks.
256   if (ST->hasAVX2()) {
257     if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
258         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
259          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
260       // On AVX2, a packed v16i16 shift left by a constant build_vector
261       // is lowered into a vector multiply (vpmullw).
262       return LT.first;
263 
264     if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second))
265       return LT.first * Entry->Cost;
266   }
267 
268   static const CostTblEntry XOPCostTable[] = {
269     // 128bit shifts take 1cy, but right shifts require negation beforehand.
270     { ISD::SHL,     MVT::v16i8,    1 },
271     { ISD::SRL,     MVT::v16i8,    2 },
272     { ISD::SRA,     MVT::v16i8,    2 },
273     { ISD::SHL,     MVT::v8i16,    1 },
274     { ISD::SRL,     MVT::v8i16,    2 },
275     { ISD::SRA,     MVT::v8i16,    2 },
276     { ISD::SHL,     MVT::v4i32,    1 },
277     { ISD::SRL,     MVT::v4i32,    2 },
278     { ISD::SRA,     MVT::v4i32,    2 },
279     { ISD::SHL,     MVT::v2i64,    1 },
280     { ISD::SRL,     MVT::v2i64,    2 },
281     { ISD::SRA,     MVT::v2i64,    2 },
282     // 256bit shifts require splitting if AVX2 didn't catch them above.
283     { ISD::SHL,     MVT::v32i8,    2 },
284     { ISD::SRL,     MVT::v32i8,    4 },
285     { ISD::SRA,     MVT::v32i8,    4 },
286     { ISD::SHL,     MVT::v16i16,   2 },
287     { ISD::SRL,     MVT::v16i16,   4 },
288     { ISD::SRA,     MVT::v16i16,   4 },
289     { ISD::SHL,     MVT::v8i32,    2 },
290     { ISD::SRL,     MVT::v8i32,    4 },
291     { ISD::SRA,     MVT::v8i32,    4 },
292     { ISD::SHL,     MVT::v4i64,    2 },
293     { ISD::SRL,     MVT::v4i64,    4 },
294     { ISD::SRA,     MVT::v4i64,    4 },
295   };
296 
297   // Look for XOP lowering tricks.
298   if (ST->hasXOP()) {
299     if (const auto *Entry = CostTableLookup(XOPCostTable, ISD, LT.second))
300       return LT.first * Entry->Cost;
301   }
302 
303   static const CostTblEntry AVX2CustomCostTable[] = {
304     { ISD::SHL,  MVT::v32i8,      11 }, // vpblendvb sequence.
305     { ISD::SHL,  MVT::v16i16,     10 }, // extend/vpsrlvd/pack sequence.
306 
307     { ISD::SRL,  MVT::v32i8,      11 }, // vpblendvb sequence.
308     { ISD::SRL,  MVT::v16i16,     10 }, // extend/vpsrlvd/pack sequence.
309 
310     { ISD::SRA,  MVT::v32i8,      24 }, // vpblendvb sequence.
311     { ISD::SRA,  MVT::v16i16,     10 }, // extend/vpsravd/pack sequence.
312     { ISD::SRA,  MVT::v2i64,       4 }, // srl/xor/sub sequence.
313     { ISD::SRA,  MVT::v4i64,       4 }, // srl/xor/sub sequence.
314   };
315 
316   // Look for AVX2 lowering tricks for custom cases.
317   if (ST->hasAVX2()) {
318     if (const auto *Entry = CostTableLookup(AVX2CustomCostTable, ISD,
319                                             LT.second))
320       return LT.first * Entry->Cost;
321   }
322 
323   static const CostTblEntry AVXCustomCostTable[] = {
324     // Vectorizing division is a bad idea. See the SSE2 table for more comments.
325     { ISD::SDIV,  MVT::v32i8,  32*20 },
326     { ISD::SDIV,  MVT::v16i16, 16*20 },
327     { ISD::SDIV,  MVT::v8i32,  8*20 },
328     { ISD::SDIV,  MVT::v4i64,  4*20 },
329     { ISD::UDIV,  MVT::v32i8,  32*20 },
330     { ISD::UDIV,  MVT::v16i16, 16*20 },
331     { ISD::UDIV,  MVT::v8i32,  8*20 },
332     { ISD::UDIV,  MVT::v4i64,  4*20 },
333   };
334 
335   // Look for AVX2 lowering tricks for custom cases.
336   if (ST->hasAVX()) {
337     if (const auto *Entry = CostTableLookup(AVXCustomCostTable, ISD,
338                                             LT.second))
339       return LT.first * Entry->Cost;
340   }
341 
342   static const CostTblEntry
343   SSE2UniformCostTable[] = {
344     // Uniform splats are cheaper for the following instructions.
345     { ISD::SHL,  MVT::v16i8,  1 }, // psllw.
346     { ISD::SHL,  MVT::v32i8,  2 }, // psllw.
347     { ISD::SHL,  MVT::v8i16,  1 }, // psllw.
348     { ISD::SHL,  MVT::v16i16, 2 }, // psllw.
349     { ISD::SHL,  MVT::v4i32,  1 }, // pslld
350     { ISD::SHL,  MVT::v8i32,  2 }, // pslld
351     { ISD::SHL,  MVT::v2i64,  1 }, // psllq.
352     { ISD::SHL,  MVT::v4i64,  2 }, // psllq.
353 
354     { ISD::SRL,  MVT::v16i8,  1 }, // psrlw.
355     { ISD::SRL,  MVT::v32i8,  2 }, // psrlw.
356     { ISD::SRL,  MVT::v8i16,  1 }, // psrlw.
357     { ISD::SRL,  MVT::v16i16, 2 }, // psrlw.
358     { ISD::SRL,  MVT::v4i32,  1 }, // psrld.
359     { ISD::SRL,  MVT::v8i32,  2 }, // psrld.
360     { ISD::SRL,  MVT::v2i64,  1 }, // psrlq.
361     { ISD::SRL,  MVT::v4i64,  2 }, // psrlq.
362 
363     { ISD::SRA,  MVT::v16i8,  4 }, // psrlw, pand, pxor, psubb.
364     { ISD::SRA,  MVT::v32i8,  8 }, // psrlw, pand, pxor, psubb.
365     { ISD::SRA,  MVT::v8i16,  1 }, // psraw.
366     { ISD::SRA,  MVT::v16i16, 2 }, // psraw.
367     { ISD::SRA,  MVT::v4i32,  1 }, // psrad.
368     { ISD::SRA,  MVT::v8i32,  2 }, // psrad.
369     { ISD::SRA,  MVT::v2i64,  4 }, // 2 x psrad + shuffle.
370     { ISD::SRA,  MVT::v4i64,  8 }, // 2 x psrad + shuffle.
371   };
372 
373   if (ST->hasSSE2() &&
374       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
375        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
376     if (const auto *Entry =
377             CostTableLookup(SSE2UniformCostTable, ISD, LT.second))
378       return LT.first * Entry->Cost;
379   }
380 
381   if (ISD == ISD::SHL &&
382       Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
383     MVT VT = LT.second;
384     // Vector shift left by non uniform constant can be lowered
385     // into vector multiply (pmullw/pmulld).
386     if ((VT == MVT::v8i16 && ST->hasSSE2()) ||
387         (VT == MVT::v4i32 && ST->hasSSE41()))
388       return LT.first;
389 
390     // v16i16 and v8i32 shifts by non-uniform constants are lowered into a
391     // sequence of extract + two vector multiply + insert.
392     if ((VT == MVT::v8i32 || VT == MVT::v16i16) &&
393        (ST->hasAVX() && !ST->hasAVX2()))
394       ISD = ISD::MUL;
395 
396     // A vector shift left by non uniform constant is converted
397     // into a vector multiply; the new multiply is eventually
398     // lowered into a sequence of shuffles and 2 x pmuludq.
399     if (VT == MVT::v4i32 && ST->hasSSE2())
400       ISD = ISD::MUL;
401   }
402 
403   static const CostTblEntry SSE41CostTable[] = {
404     { ISD::SHL,  MVT::v16i8,    11 }, // pblendvb sequence.
405     { ISD::SHL,  MVT::v32i8,  2*11 }, // pblendvb sequence.
406     { ISD::SHL,  MVT::v8i16,    14 }, // pblendvb sequence.
407     { ISD::SHL,  MVT::v16i16, 2*14 }, // pblendvb sequence.
408 
409     { ISD::SRL,  MVT::v16i8,    12 }, // pblendvb sequence.
410     { ISD::SRL,  MVT::v32i8,  2*12 }, // pblendvb sequence.
411     { ISD::SRL,  MVT::v8i16,    14 }, // pblendvb sequence.
412     { ISD::SRL,  MVT::v16i16, 2*14 }, // pblendvb sequence.
413     { ISD::SRL,  MVT::v4i32,    11 }, // Shift each lane + blend.
414     { ISD::SRL,  MVT::v8i32,  2*11 }, // Shift each lane + blend.
415 
416     { ISD::SRA,  MVT::v16i8,    24 }, // pblendvb sequence.
417     { ISD::SRA,  MVT::v32i8,  2*24 }, // pblendvb sequence.
418     { ISD::SRA,  MVT::v8i16,    14 }, // pblendvb sequence.
419     { ISD::SRA,  MVT::v16i16, 2*14 }, // pblendvb sequence.
420     { ISD::SRA,  MVT::v4i32,    12 }, // Shift each lane + blend.
421     { ISD::SRA,  MVT::v8i32,  2*12 }, // Shift each lane + blend.
422   };
423 
424   if (ST->hasSSE41()) {
425     if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second))
426       return LT.first * Entry->Cost;
427   }
428 
429   static const CostTblEntry SSE2CostTable[] = {
430     // We don't correctly identify costs of casts because they are marked as
431     // custom.
432     { ISD::SHL,  MVT::v16i8,    26 }, // cmpgtb sequence.
433     { ISD::SHL,  MVT::v32i8,  2*26 }, // cmpgtb sequence.
434     { ISD::SHL,  MVT::v8i16,    32 }, // cmpgtb sequence.
435     { ISD::SHL,  MVT::v16i16, 2*32 }, // cmpgtb sequence.
436     { ISD::SHL,  MVT::v4i32,   2*5 }, // We optimized this using mul.
437     { ISD::SHL,  MVT::v8i32, 2*2*5 }, // We optimized this using mul.
438     { ISD::SHL,  MVT::v2i64,     4 }, // splat+shuffle sequence.
439     { ISD::SHL,  MVT::v4i64,   2*4 }, // splat+shuffle sequence.
440 
441     { ISD::SRL,  MVT::v16i8,    26 }, // cmpgtb sequence.
442     { ISD::SRL,  MVT::v32i8,  2*26 }, // cmpgtb sequence.
443     { ISD::SRL,  MVT::v8i16,    32 }, // cmpgtb sequence.
444     { ISD::SRL,  MVT::v16i16, 2*32 }, // cmpgtb sequence.
445     { ISD::SRL,  MVT::v4i32,    16 }, // Shift each lane + blend.
446     { ISD::SRL,  MVT::v8i32,  2*16 }, // Shift each lane + blend.
447     { ISD::SRL,  MVT::v2i64,     4 }, // splat+shuffle sequence.
448     { ISD::SRL,  MVT::v4i64,   2*4 }, // splat+shuffle sequence.
449 
450     { ISD::SRA,  MVT::v16i8,    54 }, // unpacked cmpgtb sequence.
451     { ISD::SRA,  MVT::v32i8,  2*54 }, // unpacked cmpgtb sequence.
452     { ISD::SRA,  MVT::v8i16,    32 }, // cmpgtb sequence.
453     { ISD::SRA,  MVT::v16i16, 2*32 }, // cmpgtb sequence.
454     { ISD::SRA,  MVT::v4i32,    16 }, // Shift each lane + blend.
455     { ISD::SRA,  MVT::v8i32,  2*16 }, // Shift each lane + blend.
456     { ISD::SRA,  MVT::v2i64,    12 }, // srl/xor/sub sequence.
457     { ISD::SRA,  MVT::v4i64,  2*12 }, // srl/xor/sub sequence.
458 
459     // It is not a good idea to vectorize division. We have to scalarize it and
460     // in the process we will often end up having to spilling regular
461     // registers. The overhead of division is going to dominate most kernels
462     // anyways so try hard to prevent vectorization of division - it is
463     // generally a bad idea. Assume somewhat arbitrarily that we have to be able
464     // to hide "20 cycles" for each lane.
465     { ISD::SDIV,  MVT::v16i8,  16*20 },
466     { ISD::SDIV,  MVT::v8i16,  8*20 },
467     { ISD::SDIV,  MVT::v4i32,  4*20 },
468     { ISD::SDIV,  MVT::v2i64,  2*20 },
469     { ISD::UDIV,  MVT::v16i8,  16*20 },
470     { ISD::UDIV,  MVT::v8i16,  8*20 },
471     { ISD::UDIV,  MVT::v4i32,  4*20 },
472     { ISD::UDIV,  MVT::v2i64,  2*20 },
473   };
474 
475   if (ST->hasSSE2()) {
476     if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second))
477       return LT.first * Entry->Cost;
478   }
479 
480   static const CostTblEntry AVX1CostTable[] = {
481     // We don't have to scalarize unsupported ops. We can issue two half-sized
482     // operations and we only need to extract the upper YMM half.
483     // Two ops + 1 extract + 1 insert = 4.
484     { ISD::MUL,     MVT::v16i16,   4 },
485     { ISD::MUL,     MVT::v8i32,    4 },
486     { ISD::SUB,     MVT::v8i32,    4 },
487     { ISD::ADD,     MVT::v8i32,    4 },
488     { ISD::SUB,     MVT::v4i64,    4 },
489     { ISD::ADD,     MVT::v4i64,    4 },
490     // A v4i64 multiply is custom lowered as two split v2i64 vectors that then
491     // are lowered as a series of long multiplies(3), shifts(4) and adds(2)
492     // Because we believe v4i64 to be a legal type, we must also include the
493     // split factor of two in the cost table. Therefore, the cost here is 18
494     // instead of 9.
495     { ISD::MUL,     MVT::v4i64,    18 },
496   };
497 
498   // Look for AVX1 lowering tricks.
499   if (ST->hasAVX() && !ST->hasAVX2()) {
500     MVT VT = LT.second;
501 
502     if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, VT))
503       return LT.first * Entry->Cost;
504   }
505 
506   // Custom lowering of vectors.
507   static const CostTblEntry CustomLowered[] = {
508     // A v2i64/v4i64 and multiply is custom lowered as a series of long
509     // multiplies(3), shifts(4) and adds(2).
510     { ISD::MUL,     MVT::v2i64,    9 },
511     { ISD::MUL,     MVT::v4i64,    9 },
512   };
513   if (const auto *Entry = CostTableLookup(CustomLowered, ISD, LT.second))
514     return LT.first * Entry->Cost;
515 
516   // Special lowering of v4i32 mul on sse2, sse3: Lower v4i32 mul as 2x shuffle,
517   // 2x pmuludq, 2x shuffle.
518   if (ISD == ISD::MUL && LT.second == MVT::v4i32 && ST->hasSSE2() &&
519       !ST->hasSSE41())
520     return LT.first * 6;
521 
522   // Fallback to the default implementation.
523   return BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
524 }
525 
526 int X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
527                                Type *SubTp) {
528   // We only estimate the cost of reverse and alternate shuffles.
529   if (Kind != TTI::SK_Reverse && Kind != TTI::SK_Alternate)
530     return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
531 
532   if (Kind == TTI::SK_Reverse) {
533     std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
534     int Cost = 1;
535     if (LT.second.getSizeInBits() > 128)
536       Cost = 3; // Extract + insert + copy.
537 
538     // Multiple by the number of parts.
539     return Cost * LT.first;
540   }
541 
542   if (Kind == TTI::SK_Alternate) {
543     // 64-bit packed float vectors (v2f32) are widened to type v4f32.
544     // 64-bit packed integer vectors (v2i32) are promoted to type v2i64.
545     std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
546 
547     // The backend knows how to generate a single VEX.256 version of
548     // instruction VPBLENDW if the target supports AVX2.
549     if (ST->hasAVX2() && LT.second == MVT::v16i16)
550       return LT.first;
551 
552     static const CostTblEntry AVXAltShuffleTbl[] = {
553       {ISD::VECTOR_SHUFFLE, MVT::v4i64, 1},  // vblendpd
554       {ISD::VECTOR_SHUFFLE, MVT::v4f64, 1},  // vblendpd
555 
556       {ISD::VECTOR_SHUFFLE, MVT::v8i32, 1},  // vblendps
557       {ISD::VECTOR_SHUFFLE, MVT::v8f32, 1},  // vblendps
558 
559       // This shuffle is custom lowered into a sequence of:
560       //  2x  vextractf128 , 2x vpblendw , 1x vinsertf128
561       {ISD::VECTOR_SHUFFLE, MVT::v16i16, 5},
562 
563       // This shuffle is custom lowered into a long sequence of:
564       //  2x vextractf128 , 4x vpshufb , 2x vpor ,  1x vinsertf128
565       {ISD::VECTOR_SHUFFLE, MVT::v32i8, 9}
566     };
567 
568     if (ST->hasAVX())
569       if (const auto *Entry = CostTableLookup(AVXAltShuffleTbl,
570                                               ISD::VECTOR_SHUFFLE, LT.second))
571         return LT.first * Entry->Cost;
572 
573     static const CostTblEntry SSE41AltShuffleTbl[] = {
574       // These are lowered into movsd.
575       {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
576       {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
577 
578       // packed float vectors with four elements are lowered into BLENDI dag
579       // nodes. A v4i32/v4f32 BLENDI generates a single 'blendps'/'blendpd'.
580       {ISD::VECTOR_SHUFFLE, MVT::v4i32, 1},
581       {ISD::VECTOR_SHUFFLE, MVT::v4f32, 1},
582 
583       // This shuffle generates a single pshufw.
584       {ISD::VECTOR_SHUFFLE, MVT::v8i16, 1},
585 
586       // There is no instruction that matches a v16i8 alternate shuffle.
587       // The backend will expand it into the sequence 'pshufb + pshufb + or'.
588       {ISD::VECTOR_SHUFFLE, MVT::v16i8, 3}
589     };
590 
591     if (ST->hasSSE41())
592       if (const auto *Entry = CostTableLookup(SSE41AltShuffleTbl, ISD::VECTOR_SHUFFLE,
593                                               LT.second))
594         return LT.first * Entry->Cost;
595 
596     static const CostTblEntry SSSE3AltShuffleTbl[] = {
597       {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},  // movsd
598       {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},  // movsd
599 
600       // SSE3 doesn't have 'blendps'. The following shuffles are expanded into
601       // the sequence 'shufps + pshufd'
602       {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
603       {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
604 
605       {ISD::VECTOR_SHUFFLE, MVT::v8i16, 3}, // pshufb + pshufb + or
606       {ISD::VECTOR_SHUFFLE, MVT::v16i8, 3}  // pshufb + pshufb + or
607     };
608 
609     if (ST->hasSSSE3())
610       if (const auto *Entry = CostTableLookup(SSSE3AltShuffleTbl,
611                                               ISD::VECTOR_SHUFFLE, LT.second))
612         return LT.first * Entry->Cost;
613 
614     static const CostTblEntry SSEAltShuffleTbl[] = {
615       {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},  // movsd
616       {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},  // movsd
617 
618       {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2}, // shufps + pshufd
619       {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2}, // shufps + pshufd
620 
621       // This is expanded into a long sequence of four extract + four insert.
622       {ISD::VECTOR_SHUFFLE, MVT::v8i16, 8}, // 4 x pextrw + 4 pinsrw.
623 
624       // 8 x (pinsrw + pextrw + and + movb + movzb + or)
625       {ISD::VECTOR_SHUFFLE, MVT::v16i8, 48}
626     };
627 
628     // Fall-back (SSE3 and SSE2).
629     if (const auto *Entry = CostTableLookup(SSEAltShuffleTbl,
630                                             ISD::VECTOR_SHUFFLE, LT.second))
631       return LT.first * Entry->Cost;
632     return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
633   }
634 
635   return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
636 }
637 
638 int X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) {
639   int ISD = TLI->InstructionOpcodeToISD(Opcode);
640   assert(ISD && "Invalid opcode");
641 
642   // FIXME: Need a better design of the cost table to handle non-simple types of
643   // potential massive combinations (elem_num x src_type x dst_type).
644 
645   static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = {
646     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
647     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
648     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
649     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
650     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
651     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
652 
653     { ISD::FP_TO_UINT,  MVT::v2i64, MVT::v2f32, 1 },
654     { ISD::FP_TO_UINT,  MVT::v4i64, MVT::v4f32, 1 },
655     { ISD::FP_TO_UINT,  MVT::v8i64, MVT::v8f32, 1 },
656     { ISD::FP_TO_UINT,  MVT::v2i64, MVT::v2f64, 1 },
657     { ISD::FP_TO_UINT,  MVT::v4i64, MVT::v4f64, 1 },
658     { ISD::FP_TO_UINT,  MVT::v8i64, MVT::v8f64, 1 },
659   };
660 
661   // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and
662   // 256-bit wide vectors.
663 
664   static const TypeConversionCostTblEntry AVX512FConversionTbl[] = {
665     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v8f32,  1 },
666     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v16f32, 3 },
667     { ISD::FP_ROUND,  MVT::v8f32,   MVT::v8f64,  1 },
668 
669     { ISD::TRUNCATE,  MVT::v16i8,   MVT::v16i32, 1 },
670     { ISD::TRUNCATE,  MVT::v16i16,  MVT::v16i32, 1 },
671     { ISD::TRUNCATE,  MVT::v8i16,   MVT::v8i64,  1 },
672     { ISD::TRUNCATE,  MVT::v8i32,   MVT::v8i64,  1 },
673 
674     // v16i1 -> v16i32 - load + broadcast
675     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
676     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
677     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
678     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
679     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
680     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
681     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
682     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
683     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
684     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
685 
686     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
687     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
688     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i8,   2 },
689     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i8,  2 },
690     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
691     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i16, 2 },
692     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
693     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
694     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64, 26 },
695     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64, 26 },
696 
697     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
698     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
699     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i8,   2 },
700     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i8,   2 },
701     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i8,   2 },
702     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i8,   2 },
703     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i8,  2 },
704     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i16,  5 },
705     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i16,  2 },
706     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  2 },
707     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
708     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i16, 2 },
709     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  2 },
710     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i32,  1 },
711     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  1 },
712     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  1 },
713     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  1 },
714     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
715     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
716     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  5 },
717     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  5 },
718     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64, 12 },
719     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64, 26 },
720 
721     { ISD::FP_TO_UINT,  MVT::v2i32,  MVT::v2f32,  1 },
722     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  1 },
723     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f32,  1 },
724     { ISD::FP_TO_UINT,  MVT::v16i32, MVT::v16f32, 1 },
725   };
726 
727   static const TypeConversionCostTblEntry AVX2ConversionTbl[] = {
728     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
729     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
730     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
731     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
732     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,   3 },
733     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,   3 },
734     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   3 },
735     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   3 },
736     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
737     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
738     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
739     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
740     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
741     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
742     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
743     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
744 
745     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i64,  2 },
746     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i64,  2 },
747     { ISD::TRUNCATE,    MVT::v4i32,  MVT::v4i64,  2 },
748     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  2 },
749     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  2 },
750     { ISD::TRUNCATE,    MVT::v8i32,  MVT::v8i64,  4 },
751 
752     { ISD::FP_EXTEND,   MVT::v8f64,  MVT::v8f32,  3 },
753     { ISD::FP_ROUND,    MVT::v8f32,  MVT::v8f64,  3 },
754 
755     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  8 },
756   };
757 
758   static const TypeConversionCostTblEntry AVXConversionTbl[] = {
759     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,  6 },
760     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,  4 },
761     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,  7 },
762     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,  4 },
763     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,  6 },
764     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,  4 },
765     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,  7 },
766     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,  4 },
767     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
768     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
769     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16, 6 },
770     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16, 3 },
771     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16, 4 },
772     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16, 4 },
773     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32, 4 },
774     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32, 4 },
775 
776     { ISD::TRUNCATE,    MVT::v16i8, MVT::v16i16, 4 },
777     { ISD::TRUNCATE,    MVT::v8i8,  MVT::v8i32,  4 },
778     { ISD::TRUNCATE,    MVT::v8i16, MVT::v8i32,  5 },
779     { ISD::TRUNCATE,    MVT::v4i8,  MVT::v4i64,  4 },
780     { ISD::TRUNCATE,    MVT::v4i16, MVT::v4i64,  4 },
781     { ISD::TRUNCATE,    MVT::v4i32, MVT::v4i64,  4 },
782     { ISD::TRUNCATE,    MVT::v8i32, MVT::v8i64,  9 },
783 
784     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i1,  3 },
785     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i1,  3 },
786     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i1,  8 },
787     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i8,  3 },
788     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i8,  3 },
789     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i8,  8 },
790     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i16, 3 },
791     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i16, 3 },
792     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
793     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },
794     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i32, 1 },
795     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i32, 1 },
796 
797     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i1,  7 },
798     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i1,  7 },
799     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i1,  6 },
800     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i8,  2 },
801     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i8,  2 },
802     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i8,  5 },
803     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
804     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i16, 2 },
805     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
806     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i32, 6 },
807     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i32, 6 },
808     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i32, 6 },
809     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i32, 9 },
810     // The generic code to compute the scalar overhead is currently broken.
811     // Workaround this limitation by estimating the scalarization overhead
812     // here. We have roughly 10 instructions per scalar element.
813     // Multiply that by the vector width.
814     // FIXME: remove that when PR19268 is fixed.
815     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i64, 10 },
816     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i64, 20 },
817     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
818     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
819 
820     { ISD::FP_TO_SINT,  MVT::v4i8,  MVT::v4f32, 1 },
821     { ISD::FP_TO_SINT,  MVT::v8i8,  MVT::v8f32, 7 },
822     // This node is expanded into scalarized operations but BasicTTI is overly
823     // optimistic estimating its cost.  It computes 3 per element (one
824     // vector-extract, one scalar conversion and one vector-insert).  The
825     // problem is that the inserts form a read-modify-write chain so latency
826     // should be factored in too.  Inflating the cost per element by 1.
827     { ISD::FP_TO_UINT,  MVT::v8i32, MVT::v8f32, 8*4 },
828     { ISD::FP_TO_UINT,  MVT::v4i32, MVT::v4f64, 4*4 },
829 
830     { ISD::FP_EXTEND,   MVT::v4f64,  MVT::v4f32,  1 },
831     { ISD::FP_ROUND,    MVT::v4f32,  MVT::v4f64,  1 },
832   };
833 
834   static const TypeConversionCostTblEntry SSE41ConversionTbl[] = {
835     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8,    2 },
836     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8,    2 },
837     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16,   2 },
838     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16,   2 },
839     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32,   2 },
840     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32,   2 },
841 
842     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i8,   1 },
843     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i8,   2 },
844     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i8,   1 },
845     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i8,   1 },
846     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
847     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
848     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   2 },
849     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   2 },
850     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
851     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
852     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  4 },
853     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  4 },
854     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
855     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
856     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
857     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
858     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 4 },
859     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 4 },
860 
861     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  2 },
862     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  1 },
863     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i32,  1 },
864     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i32,  1 },
865     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  3 },
866     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  3 },
867     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 6 },
868 
869   };
870 
871   static const TypeConversionCostTblEntry SSE2ConversionTbl[] = {
872     // These are somewhat magic numbers justified by looking at the output of
873     // Intel's IACA, running some kernels and making sure when we take
874     // legalization into account the throughput will be overestimated.
875     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
876     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
877     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
878     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
879     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 },
880     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
881     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
882     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
883 
884     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
885     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
886     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
887     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
888     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
889     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 8 },
890     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
891     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
892 
893     { ISD::FP_TO_SINT,  MVT::v2i32,  MVT::v2f64,  3 },
894 
895     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i8,   1 },
896     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i8,   6 },
897     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i8,   2 },
898     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i8,   3 },
899     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,   4 },
900     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,   8 },
901     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
902     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i8,   2 },
903     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   6 },
904     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   6 },
905     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  3 },
906     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  4 },
907     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  9 },
908     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  12 },
909     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
910     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i16,  2 },
911     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
912     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16,  10 },
913     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  3 },
914     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  4 },
915     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 6 },
916     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 8 },
917     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  3 },
918     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  5 },
919 
920     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  4 },
921     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  2 },
922     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 3 },
923     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i32,  3 },
924     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i32,  3 },
925     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  4 },
926     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i32, 7 },
927     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  5 },
928     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 10 },
929   };
930 
931   std::pair<int, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src);
932   std::pair<int, MVT> LTDest = TLI->getTypeLegalizationCost(DL, Dst);
933 
934   if (ST->hasSSE2() && !ST->hasAVX()) {
935     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
936                                                    LTDest.second, LTSrc.second))
937       return LTSrc.first * Entry->Cost;
938   }
939 
940   EVT SrcTy = TLI->getValueType(DL, Src);
941   EVT DstTy = TLI->getValueType(DL, Dst);
942 
943   // The function getSimpleVT only handles simple value types.
944   if (!SrcTy.isSimple() || !DstTy.isSimple())
945     return BaseT::getCastInstrCost(Opcode, Dst, Src);
946 
947   if (ST->hasDQI())
948     if (const auto *Entry = ConvertCostTableLookup(AVX512DQConversionTbl, ISD,
949                                                    DstTy.getSimpleVT(),
950                                                    SrcTy.getSimpleVT()))
951       return Entry->Cost;
952 
953   if (ST->hasAVX512())
954     if (const auto *Entry = ConvertCostTableLookup(AVX512FConversionTbl, ISD,
955                                                    DstTy.getSimpleVT(),
956                                                    SrcTy.getSimpleVT()))
957       return Entry->Cost;
958 
959   if (ST->hasAVX2()) {
960     if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
961                                                    DstTy.getSimpleVT(),
962                                                    SrcTy.getSimpleVT()))
963       return Entry->Cost;
964   }
965 
966   if (ST->hasAVX()) {
967     if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
968                                                    DstTy.getSimpleVT(),
969                                                    SrcTy.getSimpleVT()))
970       return Entry->Cost;
971   }
972 
973   if (ST->hasSSE41()) {
974     if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
975                                                    DstTy.getSimpleVT(),
976                                                    SrcTy.getSimpleVT()))
977       return Entry->Cost;
978   }
979 
980   if (ST->hasSSE2()) {
981     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
982                                                    DstTy.getSimpleVT(),
983                                                    SrcTy.getSimpleVT()))
984       return Entry->Cost;
985   }
986 
987   return BaseT::getCastInstrCost(Opcode, Dst, Src);
988 }
989 
990 int X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) {
991   // Legalize the type.
992   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
993 
994   MVT MTy = LT.second;
995 
996   int ISD = TLI->InstructionOpcodeToISD(Opcode);
997   assert(ISD && "Invalid opcode");
998 
999   static const CostTblEntry SSE2CostTbl[] = {
1000     { ISD::SETCC,   MVT::v2i64,   8 },
1001     { ISD::SETCC,   MVT::v4i32,   1 },
1002     { ISD::SETCC,   MVT::v8i16,   1 },
1003     { ISD::SETCC,   MVT::v16i8,   1 },
1004   };
1005 
1006   static const CostTblEntry SSE42CostTbl[] = {
1007     { ISD::SETCC,   MVT::v2f64,   1 },
1008     { ISD::SETCC,   MVT::v4f32,   1 },
1009     { ISD::SETCC,   MVT::v2i64,   1 },
1010   };
1011 
1012   static const CostTblEntry AVX1CostTbl[] = {
1013     { ISD::SETCC,   MVT::v4f64,   1 },
1014     { ISD::SETCC,   MVT::v8f32,   1 },
1015     // AVX1 does not support 8-wide integer compare.
1016     { ISD::SETCC,   MVT::v4i64,   4 },
1017     { ISD::SETCC,   MVT::v8i32,   4 },
1018     { ISD::SETCC,   MVT::v16i16,  4 },
1019     { ISD::SETCC,   MVT::v32i8,   4 },
1020   };
1021 
1022   static const CostTblEntry AVX2CostTbl[] = {
1023     { ISD::SETCC,   MVT::v4i64,   1 },
1024     { ISD::SETCC,   MVT::v8i32,   1 },
1025     { ISD::SETCC,   MVT::v16i16,  1 },
1026     { ISD::SETCC,   MVT::v32i8,   1 },
1027   };
1028 
1029   static const CostTblEntry AVX512CostTbl[] = {
1030     { ISD::SETCC,   MVT::v8i64,   1 },
1031     { ISD::SETCC,   MVT::v16i32,  1 },
1032     { ISD::SETCC,   MVT::v8f64,   1 },
1033     { ISD::SETCC,   MVT::v16f32,  1 },
1034   };
1035 
1036   if (ST->hasAVX512())
1037     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
1038       return LT.first * Entry->Cost;
1039 
1040   if (ST->hasAVX2())
1041     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
1042       return LT.first * Entry->Cost;
1043 
1044   if (ST->hasAVX())
1045     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
1046       return LT.first * Entry->Cost;
1047 
1048   if (ST->hasSSE42())
1049     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
1050       return LT.first * Entry->Cost;
1051 
1052   if (ST->hasSSE2())
1053     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
1054       return LT.first * Entry->Cost;
1055 
1056   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy);
1057 }
1058 
1059 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1060                                       ArrayRef<Type *> Tys, FastMathFlags FMF) {
1061   // Costs should match the codegen from:
1062   // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll
1063   // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll
1064   // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll
1065   // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll
1066   // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll
1067   static const CostTblEntry XOPCostTbl[] = {
1068     { ISD::BITREVERSE, MVT::v4i64,   4 },
1069     { ISD::BITREVERSE, MVT::v8i32,   4 },
1070     { ISD::BITREVERSE, MVT::v16i16,  4 },
1071     { ISD::BITREVERSE, MVT::v32i8,   4 },
1072     { ISD::BITREVERSE, MVT::v2i64,   1 },
1073     { ISD::BITREVERSE, MVT::v4i32,   1 },
1074     { ISD::BITREVERSE, MVT::v8i16,   1 },
1075     { ISD::BITREVERSE, MVT::v16i8,   1 },
1076     { ISD::BITREVERSE, MVT::i64,     3 },
1077     { ISD::BITREVERSE, MVT::i32,     3 },
1078     { ISD::BITREVERSE, MVT::i16,     3 },
1079     { ISD::BITREVERSE, MVT::i8,      3 }
1080   };
1081   static const CostTblEntry AVX2CostTbl[] = {
1082     { ISD::BITREVERSE, MVT::v4i64,   5 },
1083     { ISD::BITREVERSE, MVT::v8i32,   5 },
1084     { ISD::BITREVERSE, MVT::v16i16,  5 },
1085     { ISD::BITREVERSE, MVT::v32i8,   5 },
1086     { ISD::BSWAP,      MVT::v4i64,   1 },
1087     { ISD::BSWAP,      MVT::v8i32,   1 },
1088     { ISD::BSWAP,      MVT::v16i16,  1 },
1089     { ISD::CTLZ,       MVT::v4i64,  23 },
1090     { ISD::CTLZ,       MVT::v8i32,  18 },
1091     { ISD::CTLZ,       MVT::v16i16, 14 },
1092     { ISD::CTLZ,       MVT::v32i8,   9 },
1093     { ISD::CTPOP,      MVT::v4i64,   7 },
1094     { ISD::CTPOP,      MVT::v8i32,  11 },
1095     { ISD::CTPOP,      MVT::v16i16,  9 },
1096     { ISD::CTPOP,      MVT::v32i8,   6 },
1097     { ISD::CTTZ,       MVT::v4i64,  10 },
1098     { ISD::CTTZ,       MVT::v8i32,  14 },
1099     { ISD::CTTZ,       MVT::v16i16, 12 },
1100     { ISD::CTTZ,       MVT::v32i8,   9 }
1101   };
1102   static const CostTblEntry AVX1CostTbl[] = {
1103     { ISD::BITREVERSE, MVT::v4i64,  10 },
1104     { ISD::BITREVERSE, MVT::v8i32,  10 },
1105     { ISD::BITREVERSE, MVT::v16i16, 10 },
1106     { ISD::BITREVERSE, MVT::v32i8,  10 },
1107     { ISD::BSWAP,      MVT::v4i64,   4 },
1108     { ISD::BSWAP,      MVT::v8i32,   4 },
1109     { ISD::BSWAP,      MVT::v16i16,  4 },
1110     { ISD::CTLZ,       MVT::v4i64,  46 },
1111     { ISD::CTLZ,       MVT::v8i32,  36 },
1112     { ISD::CTLZ,       MVT::v16i16, 28 },
1113     { ISD::CTLZ,       MVT::v32i8,  18 },
1114     { ISD::CTPOP,      MVT::v4i64,  14 },
1115     { ISD::CTPOP,      MVT::v8i32,  22 },
1116     { ISD::CTPOP,      MVT::v16i16, 18 },
1117     { ISD::CTPOP,      MVT::v32i8,  12 },
1118     { ISD::CTTZ,       MVT::v4i64,  20 },
1119     { ISD::CTTZ,       MVT::v8i32,  28 },
1120     { ISD::CTTZ,       MVT::v16i16, 24 },
1121     { ISD::CTTZ,       MVT::v32i8,  18 },
1122   };
1123   static const CostTblEntry SSSE3CostTbl[] = {
1124     { ISD::BITREVERSE, MVT::v2i64,   5 },
1125     { ISD::BITREVERSE, MVT::v4i32,   5 },
1126     { ISD::BITREVERSE, MVT::v8i16,   5 },
1127     { ISD::BITREVERSE, MVT::v16i8,   5 },
1128     { ISD::BSWAP,      MVT::v2i64,   1 },
1129     { ISD::BSWAP,      MVT::v4i32,   1 },
1130     { ISD::BSWAP,      MVT::v8i16,   1 },
1131     { ISD::CTLZ,       MVT::v2i64,  23 },
1132     { ISD::CTLZ,       MVT::v4i32,  18 },
1133     { ISD::CTLZ,       MVT::v8i16,  14 },
1134     { ISD::CTLZ,       MVT::v16i8,   9 },
1135     { ISD::CTPOP,      MVT::v2i64,   7 },
1136     { ISD::CTPOP,      MVT::v4i32,  11 },
1137     { ISD::CTPOP,      MVT::v8i16,   9 },
1138     { ISD::CTPOP,      MVT::v16i8,   6 },
1139     { ISD::CTTZ,       MVT::v2i64,  10 },
1140     { ISD::CTTZ,       MVT::v4i32,  14 },
1141     { ISD::CTTZ,       MVT::v8i16,  12 },
1142     { ISD::CTTZ,       MVT::v16i8,   9 }
1143   };
1144   static const CostTblEntry SSE2CostTbl[] = {
1145     { ISD::BSWAP,      MVT::v2i64,   7 },
1146     { ISD::BSWAP,      MVT::v4i32,   7 },
1147     { ISD::BSWAP,      MVT::v8i16,   7 },
1148     /* ISD::CTLZ - currently scalarized pre-SSSE3 */
1149     { ISD::CTPOP,      MVT::v2i64,  12 },
1150     { ISD::CTPOP,      MVT::v4i32,  15 },
1151     { ISD::CTPOP,      MVT::v8i16,  13 },
1152     { ISD::CTPOP,      MVT::v16i8,  10 },
1153     { ISD::CTTZ,       MVT::v2i64,  14 },
1154     { ISD::CTTZ,       MVT::v4i32,  18 },
1155     { ISD::CTTZ,       MVT::v8i16,  16 },
1156     { ISD::CTTZ,       MVT::v16i8,  13 }
1157   };
1158 
1159   unsigned ISD = ISD::DELETED_NODE;
1160   switch (IID) {
1161   default:
1162     break;
1163   case Intrinsic::bitreverse:
1164     ISD = ISD::BITREVERSE;
1165     break;
1166   case Intrinsic::bswap:
1167     ISD = ISD::BSWAP;
1168     break;
1169   case Intrinsic::ctlz:
1170     ISD = ISD::CTLZ;
1171     break;
1172   case Intrinsic::ctpop:
1173     ISD = ISD::CTPOP;
1174     break;
1175   case Intrinsic::cttz:
1176     ISD = ISD::CTTZ;
1177     break;
1178   }
1179 
1180   // Legalize the type.
1181   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
1182   MVT MTy = LT.second;
1183 
1184   // Attempt to lookup cost.
1185   if (ST->hasXOP())
1186     if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
1187       return LT.first * Entry->Cost;
1188 
1189   if (ST->hasAVX2())
1190     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
1191       return LT.first * Entry->Cost;
1192 
1193   if (ST->hasAVX())
1194     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
1195       return LT.first * Entry->Cost;
1196 
1197   if (ST->hasSSSE3())
1198     if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy))
1199       return LT.first * Entry->Cost;
1200 
1201   if (ST->hasSSE2())
1202     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
1203       return LT.first * Entry->Cost;
1204 
1205   return BaseT::getIntrinsicInstrCost(IID, RetTy, Tys, FMF);
1206 }
1207 
1208 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1209                                       ArrayRef<Value *> Args, FastMathFlags FMF) {
1210   return BaseT::getIntrinsicInstrCost(IID, RetTy, Args, FMF);
1211 }
1212 
1213 int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
1214   assert(Val->isVectorTy() && "This must be a vector type");
1215 
1216   Type *ScalarType = Val->getScalarType();
1217 
1218   if (Index != -1U) {
1219     // Legalize the type.
1220     std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Val);
1221 
1222     // This type is legalized to a scalar type.
1223     if (!LT.second.isVector())
1224       return 0;
1225 
1226     // The type may be split. Normalize the index to the new type.
1227     unsigned Width = LT.second.getVectorNumElements();
1228     Index = Index % Width;
1229 
1230     // Floating point scalars are already located in index #0.
1231     if (ScalarType->isFloatingPointTy() && Index == 0)
1232       return 0;
1233   }
1234 
1235   // Add to the base cost if we know that the extracted element of a vector is
1236   // destined to be moved to and used in the integer register file.
1237   int RegisterFileMoveCost = 0;
1238   if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy())
1239     RegisterFileMoveCost = 1;
1240 
1241   return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost;
1242 }
1243 
1244 int X86TTIImpl::getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
1245   assert (Ty->isVectorTy() && "Can only scalarize vectors");
1246   int Cost = 0;
1247 
1248   for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
1249     if (Insert)
1250       Cost += getVectorInstrCost(Instruction::InsertElement, Ty, i);
1251     if (Extract)
1252       Cost += getVectorInstrCost(Instruction::ExtractElement, Ty, i);
1253   }
1254 
1255   return Cost;
1256 }
1257 
1258 int X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
1259                                 unsigned AddressSpace) {
1260   // Handle non-power-of-two vectors such as <3 x float>
1261   if (VectorType *VTy = dyn_cast<VectorType>(Src)) {
1262     unsigned NumElem = VTy->getVectorNumElements();
1263 
1264     // Handle a few common cases:
1265     // <3 x float>
1266     if (NumElem == 3 && VTy->getScalarSizeInBits() == 32)
1267       // Cost = 64 bit store + extract + 32 bit store.
1268       return 3;
1269 
1270     // <3 x double>
1271     if (NumElem == 3 && VTy->getScalarSizeInBits() == 64)
1272       // Cost = 128 bit store + unpack + 64 bit store.
1273       return 3;
1274 
1275     // Assume that all other non-power-of-two numbers are scalarized.
1276     if (!isPowerOf2_32(NumElem)) {
1277       int Cost = BaseT::getMemoryOpCost(Opcode, VTy->getScalarType(), Alignment,
1278                                         AddressSpace);
1279       int SplitCost = getScalarizationOverhead(Src, Opcode == Instruction::Load,
1280                                                Opcode == Instruction::Store);
1281       return NumElem * Cost + SplitCost;
1282     }
1283   }
1284 
1285   // Legalize the type.
1286   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
1287   assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
1288          "Invalid Opcode");
1289 
1290   // Each load/store unit costs 1.
1291   int Cost = LT.first * 1;
1292 
1293   // This isn't exactly right. We're using slow unaligned 32-byte accesses as a
1294   // proxy for a double-pumped AVX memory interface such as on Sandybridge.
1295   if (LT.second.getStoreSize() == 32 && ST->isUnalignedMem32Slow())
1296     Cost *= 2;
1297 
1298   return Cost;
1299 }
1300 
1301 int X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy,
1302                                       unsigned Alignment,
1303                                       unsigned AddressSpace) {
1304   VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy);
1305   if (!SrcVTy)
1306     // To calculate scalar take the regular cost, without mask
1307     return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace);
1308 
1309   unsigned NumElem = SrcVTy->getVectorNumElements();
1310   VectorType *MaskTy =
1311     VectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem);
1312   if ((Opcode == Instruction::Load && !isLegalMaskedLoad(SrcVTy)) ||
1313       (Opcode == Instruction::Store && !isLegalMaskedStore(SrcVTy)) ||
1314       !isPowerOf2_32(NumElem)) {
1315     // Scalarization
1316     int MaskSplitCost = getScalarizationOverhead(MaskTy, false, true);
1317     int ScalarCompareCost = getCmpSelInstrCost(
1318         Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr);
1319     int BranchCost = getCFInstrCost(Instruction::Br);
1320     int MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost);
1321 
1322     int ValueSplitCost = getScalarizationOverhead(
1323         SrcVTy, Opcode == Instruction::Load, Opcode == Instruction::Store);
1324     int MemopCost =
1325         NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
1326                                          Alignment, AddressSpace);
1327     return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost;
1328   }
1329 
1330   // Legalize the type.
1331   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy);
1332   auto VT = TLI->getValueType(DL, SrcVTy);
1333   int Cost = 0;
1334   if (VT.isSimple() && LT.second != VT.getSimpleVT() &&
1335       LT.second.getVectorNumElements() == NumElem)
1336     // Promotion requires expand/truncate for data and a shuffle for mask.
1337     Cost += getShuffleCost(TTI::SK_Alternate, SrcVTy, 0, nullptr) +
1338             getShuffleCost(TTI::SK_Alternate, MaskTy, 0, nullptr);
1339 
1340   else if (LT.second.getVectorNumElements() > NumElem) {
1341     VectorType *NewMaskTy = VectorType::get(MaskTy->getVectorElementType(),
1342                                             LT.second.getVectorNumElements());
1343     // Expanding requires fill mask with zeroes
1344     Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, 0, MaskTy);
1345   }
1346   if (!ST->hasAVX512())
1347     return Cost + LT.first*4; // Each maskmov costs 4
1348 
1349   // AVX-512 masked load/store is cheapper
1350   return Cost+LT.first;
1351 }
1352 
1353 int X86TTIImpl::getAddressComputationCost(Type *Ty, bool IsComplex) {
1354   // Address computations in vectorized code with non-consecutive addresses will
1355   // likely result in more instructions compared to scalar code where the
1356   // computation can more often be merged into the index mode. The resulting
1357   // extra micro-ops can significantly decrease throughput.
1358   unsigned NumVectorInstToHideOverhead = 10;
1359 
1360   if (Ty->isVectorTy() && IsComplex)
1361     return NumVectorInstToHideOverhead;
1362 
1363   return BaseT::getAddressComputationCost(Ty, IsComplex);
1364 }
1365 
1366 int X86TTIImpl::getReductionCost(unsigned Opcode, Type *ValTy,
1367                                  bool IsPairwise) {
1368 
1369   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
1370 
1371   MVT MTy = LT.second;
1372 
1373   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1374   assert(ISD && "Invalid opcode");
1375 
1376   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
1377   // and make it as the cost.
1378 
1379   static const CostTblEntry SSE42CostTblPairWise[] = {
1380     { ISD::FADD,  MVT::v2f64,   2 },
1381     { ISD::FADD,  MVT::v4f32,   4 },
1382     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
1383     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.5".
1384     { ISD::ADD,   MVT::v8i16,   5 },
1385   };
1386 
1387   static const CostTblEntry AVX1CostTblPairWise[] = {
1388     { ISD::FADD,  MVT::v4f32,   4 },
1389     { ISD::FADD,  MVT::v4f64,   5 },
1390     { ISD::FADD,  MVT::v8f32,   7 },
1391     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
1392     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.5".
1393     { ISD::ADD,   MVT::v4i64,   5 },      // The data reported by the IACA tool is "4.8".
1394     { ISD::ADD,   MVT::v8i16,   5 },
1395     { ISD::ADD,   MVT::v8i32,   5 },
1396   };
1397 
1398   static const CostTblEntry SSE42CostTblNoPairWise[] = {
1399     { ISD::FADD,  MVT::v2f64,   2 },
1400     { ISD::FADD,  MVT::v4f32,   4 },
1401     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
1402     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.3".
1403     { ISD::ADD,   MVT::v8i16,   4 },      // The data reported by the IACA tool is "4.3".
1404   };
1405 
1406   static const CostTblEntry AVX1CostTblNoPairWise[] = {
1407     { ISD::FADD,  MVT::v4f32,   3 },
1408     { ISD::FADD,  MVT::v4f64,   3 },
1409     { ISD::FADD,  MVT::v8f32,   4 },
1410     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
1411     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "2.8".
1412     { ISD::ADD,   MVT::v4i64,   3 },
1413     { ISD::ADD,   MVT::v8i16,   4 },
1414     { ISD::ADD,   MVT::v8i32,   5 },
1415   };
1416 
1417   if (IsPairwise) {
1418     if (ST->hasAVX())
1419       if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy))
1420         return LT.first * Entry->Cost;
1421 
1422     if (ST->hasSSE42())
1423       if (const auto *Entry = CostTableLookup(SSE42CostTblPairWise, ISD, MTy))
1424         return LT.first * Entry->Cost;
1425   } else {
1426     if (ST->hasAVX())
1427       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
1428         return LT.first * Entry->Cost;
1429 
1430     if (ST->hasSSE42())
1431       if (const auto *Entry = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy))
1432         return LT.first * Entry->Cost;
1433   }
1434 
1435   return BaseT::getReductionCost(Opcode, ValTy, IsPairwise);
1436 }
1437 
1438 /// \brief Calculate the cost of materializing a 64-bit value. This helper
1439 /// method might only calculate a fraction of a larger immediate. Therefore it
1440 /// is valid to return a cost of ZERO.
1441 int X86TTIImpl::getIntImmCost(int64_t Val) {
1442   if (Val == 0)
1443     return TTI::TCC_Free;
1444 
1445   if (isInt<32>(Val))
1446     return TTI::TCC_Basic;
1447 
1448   return 2 * TTI::TCC_Basic;
1449 }
1450 
1451 int X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) {
1452   assert(Ty->isIntegerTy());
1453 
1454   unsigned BitSize = Ty->getPrimitiveSizeInBits();
1455   if (BitSize == 0)
1456     return ~0U;
1457 
1458   // Never hoist constants larger than 128bit, because this might lead to
1459   // incorrect code generation or assertions in codegen.
1460   // Fixme: Create a cost model for types larger than i128 once the codegen
1461   // issues have been fixed.
1462   if (BitSize > 128)
1463     return TTI::TCC_Free;
1464 
1465   if (Imm == 0)
1466     return TTI::TCC_Free;
1467 
1468   // Sign-extend all constants to a multiple of 64-bit.
1469   APInt ImmVal = Imm;
1470   if (BitSize & 0x3f)
1471     ImmVal = Imm.sext((BitSize + 63) & ~0x3fU);
1472 
1473   // Split the constant into 64-bit chunks and calculate the cost for each
1474   // chunk.
1475   int Cost = 0;
1476   for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
1477     APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
1478     int64_t Val = Tmp.getSExtValue();
1479     Cost += getIntImmCost(Val);
1480   }
1481   // We need at least one instruction to materialize the constant.
1482   return std::max(1, Cost);
1483 }
1484 
1485 int X86TTIImpl::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
1486                               Type *Ty) {
1487   assert(Ty->isIntegerTy());
1488 
1489   unsigned BitSize = Ty->getPrimitiveSizeInBits();
1490   // There is no cost model for constants with a bit size of 0. Return TCC_Free
1491   // here, so that constant hoisting will ignore this constant.
1492   if (BitSize == 0)
1493     return TTI::TCC_Free;
1494 
1495   unsigned ImmIdx = ~0U;
1496   switch (Opcode) {
1497   default:
1498     return TTI::TCC_Free;
1499   case Instruction::GetElementPtr:
1500     // Always hoist the base address of a GetElementPtr. This prevents the
1501     // creation of new constants for every base constant that gets constant
1502     // folded with the offset.
1503     if (Idx == 0)
1504       return 2 * TTI::TCC_Basic;
1505     return TTI::TCC_Free;
1506   case Instruction::Store:
1507     ImmIdx = 0;
1508     break;
1509   case Instruction::ICmp:
1510     // This is an imperfect hack to prevent constant hoisting of
1511     // compares that might be trying to check if a 64-bit value fits in
1512     // 32-bits. The backend can optimize these cases using a right shift by 32.
1513     // Ideally we would check the compare predicate here. There also other
1514     // similar immediates the backend can use shifts for.
1515     if (Idx == 1 && Imm.getBitWidth() == 64) {
1516       uint64_t ImmVal = Imm.getZExtValue();
1517       if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff)
1518         return TTI::TCC_Free;
1519     }
1520     ImmIdx = 1;
1521     break;
1522   case Instruction::And:
1523     // We support 64-bit ANDs with immediates with 32-bits of leading zeroes
1524     // by using a 32-bit operation with implicit zero extension. Detect such
1525     // immediates here as the normal path expects bit 31 to be sign extended.
1526     if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue()))
1527       return TTI::TCC_Free;
1528     LLVM_FALLTHROUGH;
1529   case Instruction::Add:
1530   case Instruction::Sub:
1531   case Instruction::Mul:
1532   case Instruction::UDiv:
1533   case Instruction::SDiv:
1534   case Instruction::URem:
1535   case Instruction::SRem:
1536   case Instruction::Or:
1537   case Instruction::Xor:
1538     ImmIdx = 1;
1539     break;
1540   // Always return TCC_Free for the shift value of a shift instruction.
1541   case Instruction::Shl:
1542   case Instruction::LShr:
1543   case Instruction::AShr:
1544     if (Idx == 1)
1545       return TTI::TCC_Free;
1546     break;
1547   case Instruction::Trunc:
1548   case Instruction::ZExt:
1549   case Instruction::SExt:
1550   case Instruction::IntToPtr:
1551   case Instruction::PtrToInt:
1552   case Instruction::BitCast:
1553   case Instruction::PHI:
1554   case Instruction::Call:
1555   case Instruction::Select:
1556   case Instruction::Ret:
1557   case Instruction::Load:
1558     break;
1559   }
1560 
1561   if (Idx == ImmIdx) {
1562     int NumConstants = (BitSize + 63) / 64;
1563     int Cost = X86TTIImpl::getIntImmCost(Imm, Ty);
1564     return (Cost <= NumConstants * TTI::TCC_Basic)
1565                ? static_cast<int>(TTI::TCC_Free)
1566                : Cost;
1567   }
1568 
1569   return X86TTIImpl::getIntImmCost(Imm, Ty);
1570 }
1571 
1572 int X86TTIImpl::getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
1573                               Type *Ty) {
1574   assert(Ty->isIntegerTy());
1575 
1576   unsigned BitSize = Ty->getPrimitiveSizeInBits();
1577   // There is no cost model for constants with a bit size of 0. Return TCC_Free
1578   // here, so that constant hoisting will ignore this constant.
1579   if (BitSize == 0)
1580     return TTI::TCC_Free;
1581 
1582   switch (IID) {
1583   default:
1584     return TTI::TCC_Free;
1585   case Intrinsic::sadd_with_overflow:
1586   case Intrinsic::uadd_with_overflow:
1587   case Intrinsic::ssub_with_overflow:
1588   case Intrinsic::usub_with_overflow:
1589   case Intrinsic::smul_with_overflow:
1590   case Intrinsic::umul_with_overflow:
1591     if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
1592       return TTI::TCC_Free;
1593     break;
1594   case Intrinsic::experimental_stackmap:
1595     if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
1596       return TTI::TCC_Free;
1597     break;
1598   case Intrinsic::experimental_patchpoint_void:
1599   case Intrinsic::experimental_patchpoint_i64:
1600     if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
1601       return TTI::TCC_Free;
1602     break;
1603   }
1604   return X86TTIImpl::getIntImmCost(Imm, Ty);
1605 }
1606 
1607 // Return an average cost of Gather / Scatter instruction, maybe improved later
1608 int X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy, Value *Ptr,
1609                                 unsigned Alignment, unsigned AddressSpace) {
1610 
1611   assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost");
1612   unsigned VF = SrcVTy->getVectorNumElements();
1613 
1614   // Try to reduce index size from 64 bit (default for GEP)
1615   // to 32. It is essential for VF 16. If the index can't be reduced to 32, the
1616   // operation will use 16 x 64 indices which do not fit in a zmm and needs
1617   // to split. Also check that the base pointer is the same for all lanes,
1618   // and that there's at most one variable index.
1619   auto getIndexSizeInBits = [](Value *Ptr, const DataLayout& DL) {
1620     unsigned IndexSize = DL.getPointerSizeInBits();
1621     GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
1622     if (IndexSize < 64 || !GEP)
1623       return IndexSize;
1624 
1625     unsigned NumOfVarIndices = 0;
1626     Value *Ptrs = GEP->getPointerOperand();
1627     if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs))
1628       return IndexSize;
1629     for (unsigned i = 1; i < GEP->getNumOperands(); ++i) {
1630       if (isa<Constant>(GEP->getOperand(i)))
1631         continue;
1632       Type *IndxTy = GEP->getOperand(i)->getType();
1633       if (IndxTy->isVectorTy())
1634         IndxTy = IndxTy->getVectorElementType();
1635       if ((IndxTy->getPrimitiveSizeInBits() == 64 &&
1636           !isa<SExtInst>(GEP->getOperand(i))) ||
1637          ++NumOfVarIndices > 1)
1638         return IndexSize; // 64
1639     }
1640     return (unsigned)32;
1641   };
1642 
1643 
1644   // Trying to reduce IndexSize to 32 bits for vector 16.
1645   // By default the IndexSize is equal to pointer size.
1646   unsigned IndexSize = (VF >= 16) ? getIndexSizeInBits(Ptr, DL) :
1647     DL.getPointerSizeInBits();
1648 
1649   Type *IndexVTy = VectorType::get(IntegerType::get(SrcVTy->getContext(),
1650                                                     IndexSize), VF);
1651   std::pair<int, MVT> IdxsLT = TLI->getTypeLegalizationCost(DL, IndexVTy);
1652   std::pair<int, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, SrcVTy);
1653   int SplitFactor = std::max(IdxsLT.first, SrcLT.first);
1654   if (SplitFactor > 1) {
1655     // Handle splitting of vector of pointers
1656     Type *SplitSrcTy = VectorType::get(SrcVTy->getScalarType(), VF / SplitFactor);
1657     return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment,
1658                                          AddressSpace);
1659   }
1660 
1661   // The gather / scatter cost is given by Intel architects. It is a rough
1662   // number since we are looking at one instruction in a time.
1663   const int GSOverhead = 2;
1664   return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
1665                                            Alignment, AddressSpace);
1666 }
1667 
1668 /// Return the cost of full scalarization of gather / scatter operation.
1669 ///
1670 /// Opcode - Load or Store instruction.
1671 /// SrcVTy - The type of the data vector that should be gathered or scattered.
1672 /// VariableMask - The mask is non-constant at compile time.
1673 /// Alignment - Alignment for one element.
1674 /// AddressSpace - pointer[s] address space.
1675 ///
1676 int X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy,
1677                                 bool VariableMask, unsigned Alignment,
1678                                 unsigned AddressSpace) {
1679   unsigned VF = SrcVTy->getVectorNumElements();
1680 
1681   int MaskUnpackCost = 0;
1682   if (VariableMask) {
1683     VectorType *MaskTy =
1684       VectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF);
1685     MaskUnpackCost = getScalarizationOverhead(MaskTy, false, true);
1686     int ScalarCompareCost =
1687       getCmpSelInstrCost(Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()),
1688                          nullptr);
1689     int BranchCost = getCFInstrCost(Instruction::Br);
1690     MaskUnpackCost += VF * (BranchCost + ScalarCompareCost);
1691   }
1692 
1693   // The cost of the scalar loads/stores.
1694   int MemoryOpCost = VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
1695                                           Alignment, AddressSpace);
1696 
1697   int InsertExtractCost = 0;
1698   if (Opcode == Instruction::Load)
1699     for (unsigned i = 0; i < VF; ++i)
1700       // Add the cost of inserting each scalar load into the vector
1701       InsertExtractCost +=
1702         getVectorInstrCost(Instruction::InsertElement, SrcVTy, i);
1703   else
1704     for (unsigned i = 0; i < VF; ++i)
1705       // Add the cost of extracting each element out of the data vector
1706       InsertExtractCost +=
1707         getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i);
1708 
1709   return MemoryOpCost + MaskUnpackCost + InsertExtractCost;
1710 }
1711 
1712 /// Calculate the cost of Gather / Scatter operation
1713 int X86TTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *SrcVTy,
1714                                        Value *Ptr, bool VariableMask,
1715                                        unsigned Alignment) {
1716   assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter");
1717   unsigned VF = SrcVTy->getVectorNumElements();
1718   PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
1719   if (!PtrTy && Ptr->getType()->isVectorTy())
1720     PtrTy = dyn_cast<PointerType>(Ptr->getType()->getVectorElementType());
1721   assert(PtrTy && "Unexpected type for Ptr argument");
1722   unsigned AddressSpace = PtrTy->getAddressSpace();
1723 
1724   bool Scalarize = false;
1725   if ((Opcode == Instruction::Load && !isLegalMaskedGather(SrcVTy)) ||
1726       (Opcode == Instruction::Store && !isLegalMaskedScatter(SrcVTy)))
1727     Scalarize = true;
1728   // Gather / Scatter for vector 2 is not profitable on KNL / SKX
1729   // Vector-4 of gather/scatter instruction does not exist on KNL.
1730   // We can extend it to 8 elements, but zeroing upper bits of
1731   // the mask vector will add more instructions. Right now we give the scalar
1732   // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter instruction is
1733   // better in the VariableMask case.
1734   if (VF == 2 || (VF == 4 && !ST->hasVLX()))
1735     Scalarize = true;
1736 
1737   if (Scalarize)
1738     return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment, AddressSpace);
1739 
1740   return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace);
1741 }
1742 
1743 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy) {
1744   Type *ScalarTy = DataTy->getScalarType();
1745   int DataWidth = isa<PointerType>(ScalarTy) ?
1746     DL.getPointerSizeInBits() : ScalarTy->getPrimitiveSizeInBits();
1747 
1748   return ((DataWidth == 32 || DataWidth == 64) && ST->hasAVX()) ||
1749          ((DataWidth == 8 || DataWidth == 16) && ST->hasBWI());
1750 }
1751 
1752 bool X86TTIImpl::isLegalMaskedStore(Type *DataType) {
1753   return isLegalMaskedLoad(DataType);
1754 }
1755 
1756 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy) {
1757   // This function is called now in two cases: from the Loop Vectorizer
1758   // and from the Scalarizer.
1759   // When the Loop Vectorizer asks about legality of the feature,
1760   // the vectorization factor is not calculated yet. The Loop Vectorizer
1761   // sends a scalar type and the decision is based on the width of the
1762   // scalar element.
1763   // Later on, the cost model will estimate usage this intrinsic based on
1764   // the vector type.
1765   // The Scalarizer asks again about legality. It sends a vector type.
1766   // In this case we can reject non-power-of-2 vectors.
1767   if (isa<VectorType>(DataTy) && !isPowerOf2_32(DataTy->getVectorNumElements()))
1768     return false;
1769   Type *ScalarTy = DataTy->getScalarType();
1770   int DataWidth = isa<PointerType>(ScalarTy) ?
1771     DL.getPointerSizeInBits() : ScalarTy->getPrimitiveSizeInBits();
1772 
1773   // AVX-512 allows gather and scatter
1774   return (DataWidth == 32 || DataWidth == 64) && ST->hasAVX512();
1775 }
1776 
1777 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType) {
1778   return isLegalMaskedGather(DataType);
1779 }
1780 
1781 bool X86TTIImpl::areInlineCompatible(const Function *Caller,
1782                                      const Function *Callee) const {
1783   const TargetMachine &TM = getTLI()->getTargetMachine();
1784 
1785   // Work this as a subsetting of subtarget features.
1786   const FeatureBitset &CallerBits =
1787       TM.getSubtargetImpl(*Caller)->getFeatureBits();
1788   const FeatureBitset &CalleeBits =
1789       TM.getSubtargetImpl(*Callee)->getFeatureBits();
1790 
1791   // FIXME: This is likely too limiting as it will include subtarget features
1792   // that we might not care about for inlining, but it is conservatively
1793   // correct.
1794   return (CallerBits & CalleeBits) == CalleeBits;
1795 }
1796 
1797 bool X86TTIImpl::enableInterleavedAccessVectorization() {
1798   // TODO: We expect this to be beneficial regardless of arch,
1799   // but there are currently some unexplained performance artifacts on Atom.
1800   // As a temporary solution, disable on Atom.
1801   return !(ST->isAtom() || ST->isSLM());
1802 }
1803