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 
17 #include "X86TargetTransformInfo.h"
18 #include "llvm/Analysis/TargetTransformInfo.h"
19 #include "llvm/CodeGen/BasicTTIImpl.h"
20 #include "llvm/IR/IntrinsicInst.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Target/CostTable.h"
23 #include "llvm/Target/TargetLowering.h"
24 
25 using namespace llvm;
26 
27 #define DEBUG_TYPE "x86tti"
28 
29 //===----------------------------------------------------------------------===//
30 //
31 // X86 cost model.
32 //
33 //===----------------------------------------------------------------------===//
34 
35 TargetTransformInfo::PopcntSupportKind
36 X86TTIImpl::getPopcntSupport(unsigned TyWidth) {
37   assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
38   // TODO: Currently the __builtin_popcount() implementation using SSE3
39   //   instructions is inefficient. Once the problem is fixed, we should
40   //   call ST->hasSSE3() instead of ST->hasPOPCNT().
41   return ST->hasPOPCNT() ? TTI::PSK_FastHardware : TTI::PSK_Software;
42 }
43 
44 unsigned X86TTIImpl::getNumberOfRegisters(bool Vector) {
45   if (Vector && !ST->hasSSE1())
46     return 0;
47 
48   if (ST->is64Bit()) {
49     if (Vector && ST->hasAVX512())
50       return 32;
51     return 16;
52   }
53   return 8;
54 }
55 
56 unsigned X86TTIImpl::getRegisterBitWidth(bool Vector) {
57   if (Vector) {
58     if (ST->hasAVX512()) return 512;
59     if (ST->hasAVX()) return 256;
60     if (ST->hasSSE1()) return 128;
61     return 0;
62   }
63 
64   if (ST->is64Bit())
65     return 64;
66 
67   return 32;
68 }
69 
70 unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) {
71   // If the loop will not be vectorized, don't interleave the loop.
72   // Let regular unroll to unroll the loop, which saves the overflow
73   // check and memory check cost.
74   if (VF == 1)
75     return 1;
76 
77   if (ST->isAtom())
78     return 1;
79 
80   // Sandybridge and Haswell have multiple execution ports and pipelined
81   // vector units.
82   if (ST->hasAVX())
83     return 4;
84 
85   return 2;
86 }
87 
88 int X86TTIImpl::getArithmeticInstrCost(
89     unsigned Opcode, Type *Ty, TTI::OperandValueKind Op1Info,
90     TTI::OperandValueKind Op2Info, TTI::OperandValueProperties Opd1PropInfo,
91     TTI::OperandValueProperties Opd2PropInfo) {
92   // Legalize the type.
93   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
94 
95   int ISD = TLI->InstructionOpcodeToISD(Opcode);
96   assert(ISD && "Invalid opcode");
97 
98   if (ISD == ISD::SDIV &&
99       Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
100       Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
101     // On X86, vector signed division by constants power-of-two are
102     // normally expanded to the sequence SRA + SRL + ADD + SRA.
103     // The OperandValue properties many not be same as that of previous
104     // operation;conservatively assume OP_None.
105     int Cost = 2 * getArithmeticInstrCost(Instruction::AShr, Ty, Op1Info,
106                                           Op2Info, TargetTransformInfo::OP_None,
107                                           TargetTransformInfo::OP_None);
108     Cost += getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info,
109                                    TargetTransformInfo::OP_None,
110                                    TargetTransformInfo::OP_None);
111     Cost += getArithmeticInstrCost(Instruction::Add, Ty, Op1Info, Op2Info,
112                                    TargetTransformInfo::OP_None,
113                                    TargetTransformInfo::OP_None);
114 
115     return Cost;
116   }
117 
118   static const CostTblEntry AVX2UniformConstCostTable[] = {
119     { ISD::SRA,  MVT::v4i64,   4 }, // 2 x psrad + shuffle.
120 
121     { ISD::SDIV, MVT::v16i16,  6 }, // vpmulhw sequence
122     { ISD::UDIV, MVT::v16i16,  6 }, // vpmulhuw sequence
123     { ISD::SDIV, MVT::v8i32,  15 }, // vpmuldq sequence
124     { ISD::UDIV, MVT::v8i32,  15 }, // vpmuludq sequence
125   };
126 
127   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
128       ST->hasAVX2()) {
129     if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD,
130                                             LT.second))
131       return LT.first * Entry->Cost;
132   }
133 
134   static const CostTblEntry AVX512CostTable[] = {
135     { ISD::SHL,     MVT::v16i32,    1 },
136     { ISD::SRL,     MVT::v16i32,    1 },
137     { ISD::SRA,     MVT::v16i32,    1 },
138     { ISD::SHL,     MVT::v8i64,    1 },
139     { ISD::SRL,     MVT::v8i64,    1 },
140     { ISD::SRA,     MVT::v8i64,    1 },
141   };
142 
143   if (ST->hasAVX512()) {
144     if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second))
145       return LT.first * Entry->Cost;
146   }
147 
148   static const CostTblEntry AVX2CostTable[] = {
149     // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
150     // customize them to detect the cases where shift amount is a scalar one.
151     { ISD::SHL,     MVT::v4i32,    1 },
152     { ISD::SRL,     MVT::v4i32,    1 },
153     { ISD::SRA,     MVT::v4i32,    1 },
154     { ISD::SHL,     MVT::v8i32,    1 },
155     { ISD::SRL,     MVT::v8i32,    1 },
156     { ISD::SRA,     MVT::v8i32,    1 },
157     { ISD::SHL,     MVT::v2i64,    1 },
158     { ISD::SRL,     MVT::v2i64,    1 },
159     { ISD::SHL,     MVT::v4i64,    1 },
160     { ISD::SRL,     MVT::v4i64,    1 },
161   };
162 
163   // Look for AVX2 lowering tricks.
164   if (ST->hasAVX2()) {
165     if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
166         (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
167          Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
168       // On AVX2, a packed v16i16 shift left by a constant build_vector
169       // is lowered into a vector multiply (vpmullw).
170       return LT.first;
171 
172     if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second))
173       return LT.first * Entry->Cost;
174   }
175 
176   static const CostTblEntry XOPCostTable[] = {
177     // 128bit shifts take 1cy, but right shifts require negation beforehand.
178     { ISD::SHL,     MVT::v16i8,    1 },
179     { ISD::SRL,     MVT::v16i8,    2 },
180     { ISD::SRA,     MVT::v16i8,    2 },
181     { ISD::SHL,     MVT::v8i16,    1 },
182     { ISD::SRL,     MVT::v8i16,    2 },
183     { ISD::SRA,     MVT::v8i16,    2 },
184     { ISD::SHL,     MVT::v4i32,    1 },
185     { ISD::SRL,     MVT::v4i32,    2 },
186     { ISD::SRA,     MVT::v4i32,    2 },
187     { ISD::SHL,     MVT::v2i64,    1 },
188     { ISD::SRL,     MVT::v2i64,    2 },
189     { ISD::SRA,     MVT::v2i64,    2 },
190     // 256bit shifts require splitting if AVX2 didn't catch them above.
191     { ISD::SHL,     MVT::v32i8,    2 },
192     { ISD::SRL,     MVT::v32i8,    4 },
193     { ISD::SRA,     MVT::v32i8,    4 },
194     { ISD::SHL,     MVT::v16i16,   2 },
195     { ISD::SRL,     MVT::v16i16,   4 },
196     { ISD::SRA,     MVT::v16i16,   4 },
197     { ISD::SHL,     MVT::v8i32,    2 },
198     { ISD::SRL,     MVT::v8i32,    4 },
199     { ISD::SRA,     MVT::v8i32,    4 },
200     { ISD::SHL,     MVT::v4i64,    2 },
201     { ISD::SRL,     MVT::v4i64,    4 },
202     { ISD::SRA,     MVT::v4i64,    4 },
203   };
204 
205   // Look for XOP lowering tricks.
206   if (ST->hasXOP()) {
207     if (const auto *Entry = CostTableLookup(XOPCostTable, ISD, LT.second))
208       return LT.first * Entry->Cost;
209   }
210 
211   static const CostTblEntry AVX2CustomCostTable[] = {
212     { ISD::SHL,  MVT::v32i8,      11 }, // vpblendvb sequence.
213     { ISD::SHL,  MVT::v16i16,     10 }, // extend/vpsrlvd/pack sequence.
214 
215     { ISD::SRL,  MVT::v32i8,      11 }, // vpblendvb sequence.
216     { ISD::SRL,  MVT::v16i16,     10 }, // extend/vpsrlvd/pack sequence.
217 
218     { ISD::SRA,  MVT::v32i8,      24 }, // vpblendvb sequence.
219     { ISD::SRA,  MVT::v16i16,     10 }, // extend/vpsravd/pack sequence.
220     { ISD::SRA,  MVT::v2i64,       4 }, // srl/xor/sub sequence.
221     { ISD::SRA,  MVT::v4i64,       4 }, // srl/xor/sub sequence.
222 
223     // Vectorizing division is a bad idea. See the SSE2 table for more comments.
224     { ISD::SDIV,  MVT::v32i8,  32*20 },
225     { ISD::SDIV,  MVT::v16i16, 16*20 },
226     { ISD::SDIV,  MVT::v8i32,  8*20 },
227     { ISD::SDIV,  MVT::v4i64,  4*20 },
228     { ISD::UDIV,  MVT::v32i8,  32*20 },
229     { ISD::UDIV,  MVT::v16i16, 16*20 },
230     { ISD::UDIV,  MVT::v8i32,  8*20 },
231     { ISD::UDIV,  MVT::v4i64,  4*20 },
232   };
233 
234   // Look for AVX2 lowering tricks for custom cases.
235   if (ST->hasAVX2()) {
236     if (const auto *Entry = CostTableLookup(AVX2CustomCostTable, ISD,
237                                             LT.second))
238       return LT.first * Entry->Cost;
239   }
240 
241   static const CostTblEntry
242   SSE2UniformConstCostTable[] = {
243     // We don't correctly identify costs of casts because they are marked as
244     // custom.
245     // Constant splats are cheaper for the following instructions.
246     { ISD::SHL,  MVT::v16i8,  1 }, // psllw.
247     { ISD::SHL,  MVT::v32i8,  2 }, // psllw.
248     { ISD::SHL,  MVT::v8i16,  1 }, // psllw.
249     { ISD::SHL,  MVT::v16i16, 2 }, // psllw.
250     { ISD::SHL,  MVT::v4i32,  1 }, // pslld
251     { ISD::SHL,  MVT::v8i32,  2 }, // pslld
252     { ISD::SHL,  MVT::v2i64,  1 }, // psllq.
253     { ISD::SHL,  MVT::v4i64,  2 }, // psllq.
254 
255     { ISD::SRL,  MVT::v16i8,  1 }, // psrlw.
256     { ISD::SRL,  MVT::v32i8,  2 }, // psrlw.
257     { ISD::SRL,  MVT::v8i16,  1 }, // psrlw.
258     { ISD::SRL,  MVT::v16i16, 2 }, // psrlw.
259     { ISD::SRL,  MVT::v4i32,  1 }, // psrld.
260     { ISD::SRL,  MVT::v8i32,  2 }, // psrld.
261     { ISD::SRL,  MVT::v2i64,  1 }, // psrlq.
262     { ISD::SRL,  MVT::v4i64,  2 }, // psrlq.
263 
264     { ISD::SRA,  MVT::v16i8,  4 }, // psrlw, pand, pxor, psubb.
265     { ISD::SRA,  MVT::v32i8,  8 }, // psrlw, pand, pxor, psubb.
266     { ISD::SRA,  MVT::v8i16,  1 }, // psraw.
267     { ISD::SRA,  MVT::v16i16, 2 }, // psraw.
268     { ISD::SRA,  MVT::v4i32,  1 }, // psrad.
269     { ISD::SRA,  MVT::v8i32,  2 }, // psrad.
270     { ISD::SRA,  MVT::v2i64,  4 }, // 2 x psrad + shuffle.
271     { ISD::SRA,  MVT::v4i64,  8 }, // 2 x psrad + shuffle.
272 
273     { ISD::SDIV, MVT::v8i16,  6 }, // pmulhw sequence
274     { ISD::UDIV, MVT::v8i16,  6 }, // pmulhuw sequence
275     { ISD::SDIV, MVT::v4i32, 19 }, // pmuludq sequence
276     { ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence
277   };
278 
279   if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
280       ST->hasSSE2()) {
281     // pmuldq sequence.
282     if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
283       return LT.first * 15;
284 
285     if (const auto *Entry = CostTableLookup(SSE2UniformConstCostTable, ISD,
286                                             LT.second))
287       return LT.first * Entry->Cost;
288   }
289 
290   if (ISD == ISD::SHL &&
291       Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
292     MVT VT = LT.second;
293     // Vector shift left by non uniform constant can be lowered
294     // into vector multiply (pmullw/pmulld).
295     if ((VT == MVT::v8i16 && ST->hasSSE2()) ||
296         (VT == MVT::v4i32 && ST->hasSSE41()))
297       return LT.first;
298 
299     // v16i16 and v8i32 shifts by non-uniform constants are lowered into a
300     // sequence of extract + two vector multiply + insert.
301     if ((VT == MVT::v8i32 || VT == MVT::v16i16) &&
302        (ST->hasAVX() && !ST->hasAVX2()))
303       ISD = ISD::MUL;
304 
305     // A vector shift left by non uniform constant is converted
306     // into a vector multiply; the new multiply is eventually
307     // lowered into a sequence of shuffles and 2 x pmuludq.
308     if (VT == MVT::v4i32 && ST->hasSSE2())
309       ISD = ISD::MUL;
310   }
311 
312   static const CostTblEntry SSE2CostTable[] = {
313     // We don't correctly identify costs of casts because they are marked as
314     // custom.
315     // For some cases, where the shift amount is a scalar we would be able
316     // to generate better code. Unfortunately, when this is the case the value
317     // (the splat) will get hoisted out of the loop, thereby making it invisible
318     // to ISel. The cost model must return worst case assumptions because it is
319     // used for vectorization and we don't want to make vectorized code worse
320     // than scalar code.
321     { ISD::SHL,  MVT::v16i8,    26 }, // cmpgtb sequence.
322     { ISD::SHL,  MVT::v32i8,  2*26 }, // cmpgtb sequence.
323     { ISD::SHL,  MVT::v8i16,    32 }, // cmpgtb sequence.
324     { ISD::SHL,  MVT::v16i16, 2*32 }, // cmpgtb sequence.
325     { ISD::SHL,  MVT::v4i32,   2*5 }, // We optimized this using mul.
326     { ISD::SHL,  MVT::v8i32, 2*2*5 }, // We optimized this using mul.
327     { ISD::SHL,  MVT::v2i64,     4 }, // splat+shuffle sequence.
328     { ISD::SHL,  MVT::v4i64,   2*4 }, // splat+shuffle sequence.
329 
330     { ISD::SRL,  MVT::v16i8,    26 }, // cmpgtb sequence.
331     { ISD::SRL,  MVT::v32i8,  2*26 }, // cmpgtb sequence.
332     { ISD::SRL,  MVT::v8i16,    32 }, // cmpgtb sequence.
333     { ISD::SRL,  MVT::v16i16, 2*32 }, // cmpgtb sequence.
334     { ISD::SRL,  MVT::v4i32,    16 }, // Shift each lane + blend.
335     { ISD::SRL,  MVT::v8i32,  2*16 }, // Shift each lane + blend.
336     { ISD::SRL,  MVT::v2i64,     4 }, // splat+shuffle sequence.
337     { ISD::SRL,  MVT::v4i64,   2*4 }, // splat+shuffle sequence.
338 
339     { ISD::SRA,  MVT::v16i8,    54 }, // unpacked cmpgtb sequence.
340     { ISD::SRA,  MVT::v32i8,  2*54 }, // unpacked cmpgtb sequence.
341     { ISD::SRA,  MVT::v8i16,    32 }, // cmpgtb sequence.
342     { ISD::SRA,  MVT::v16i16, 2*32 }, // cmpgtb sequence.
343     { ISD::SRA,  MVT::v4i32,    16 }, // Shift each lane + blend.
344     { ISD::SRA,  MVT::v8i32,  2*16 }, // Shift each lane + blend.
345     { ISD::SRA,  MVT::v2i64,    12 }, // srl/xor/sub sequence.
346     { ISD::SRA,  MVT::v4i64,  2*12 }, // srl/xor/sub sequence.
347 
348     // It is not a good idea to vectorize division. We have to scalarize it and
349     // in the process we will often end up having to spilling regular
350     // registers. The overhead of division is going to dominate most kernels
351     // anyways so try hard to prevent vectorization of division - it is
352     // generally a bad idea. Assume somewhat arbitrarily that we have to be able
353     // to hide "20 cycles" for each lane.
354     { ISD::SDIV,  MVT::v16i8,  16*20 },
355     { ISD::SDIV,  MVT::v8i16,  8*20 },
356     { ISD::SDIV,  MVT::v4i32,  4*20 },
357     { ISD::SDIV,  MVT::v2i64,  2*20 },
358     { ISD::UDIV,  MVT::v16i8,  16*20 },
359     { ISD::UDIV,  MVT::v8i16,  8*20 },
360     { ISD::UDIV,  MVT::v4i32,  4*20 },
361     { ISD::UDIV,  MVT::v2i64,  2*20 },
362   };
363 
364   if (ST->hasSSE2()) {
365     if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second))
366       return LT.first * Entry->Cost;
367   }
368 
369   static const CostTblEntry AVX1CostTable[] = {
370     // We don't have to scalarize unsupported ops. We can issue two half-sized
371     // operations and we only need to extract the upper YMM half.
372     // Two ops + 1 extract + 1 insert = 4.
373     { ISD::MUL,     MVT::v16i16,   4 },
374     { ISD::MUL,     MVT::v8i32,    4 },
375     { ISD::SUB,     MVT::v8i32,    4 },
376     { ISD::ADD,     MVT::v8i32,    4 },
377     { ISD::SUB,     MVT::v4i64,    4 },
378     { ISD::ADD,     MVT::v4i64,    4 },
379     // A v4i64 multiply is custom lowered as two split v2i64 vectors that then
380     // are lowered as a series of long multiplies(3), shifts(4) and adds(2)
381     // Because we believe v4i64 to be a legal type, we must also include the
382     // split factor of two in the cost table. Therefore, the cost here is 18
383     // instead of 9.
384     { ISD::MUL,     MVT::v4i64,    18 },
385   };
386 
387   // Look for AVX1 lowering tricks.
388   if (ST->hasAVX() && !ST->hasAVX2()) {
389     MVT VT = LT.second;
390 
391     if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, VT))
392       return LT.first * Entry->Cost;
393   }
394 
395   // Custom lowering of vectors.
396   static const CostTblEntry CustomLowered[] = {
397     // A v2i64/v4i64 and multiply is custom lowered as a series of long
398     // multiplies(3), shifts(4) and adds(2).
399     { ISD::MUL,     MVT::v2i64,    9 },
400     { ISD::MUL,     MVT::v4i64,    9 },
401   };
402   if (const auto *Entry = CostTableLookup(CustomLowered, ISD, LT.second))
403     return LT.first * Entry->Cost;
404 
405   // Special lowering of v4i32 mul on sse2, sse3: Lower v4i32 mul as 2x shuffle,
406   // 2x pmuludq, 2x shuffle.
407   if (ISD == ISD::MUL && LT.second == MVT::v4i32 && ST->hasSSE2() &&
408       !ST->hasSSE41())
409     return LT.first * 6;
410 
411   // Fallback to the default implementation.
412   return BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
413 }
414 
415 int X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
416                                Type *SubTp) {
417   // We only estimate the cost of reverse and alternate shuffles.
418   if (Kind != TTI::SK_Reverse && Kind != TTI::SK_Alternate)
419     return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
420 
421   if (Kind == TTI::SK_Reverse) {
422     std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
423     int Cost = 1;
424     if (LT.second.getSizeInBits() > 128)
425       Cost = 3; // Extract + insert + copy.
426 
427     // Multiple by the number of parts.
428     return Cost * LT.first;
429   }
430 
431   if (Kind == TTI::SK_Alternate) {
432     // 64-bit packed float vectors (v2f32) are widened to type v4f32.
433     // 64-bit packed integer vectors (v2i32) are promoted to type v2i64.
434     std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
435 
436     // The backend knows how to generate a single VEX.256 version of
437     // instruction VPBLENDW if the target supports AVX2.
438     if (ST->hasAVX2() && LT.second == MVT::v16i16)
439       return LT.first;
440 
441     static const CostTblEntry AVXAltShuffleTbl[] = {
442       {ISD::VECTOR_SHUFFLE, MVT::v4i64, 1},  // vblendpd
443       {ISD::VECTOR_SHUFFLE, MVT::v4f64, 1},  // vblendpd
444 
445       {ISD::VECTOR_SHUFFLE, MVT::v8i32, 1},  // vblendps
446       {ISD::VECTOR_SHUFFLE, MVT::v8f32, 1},  // vblendps
447 
448       // This shuffle is custom lowered into a sequence of:
449       //  2x  vextractf128 , 2x vpblendw , 1x vinsertf128
450       {ISD::VECTOR_SHUFFLE, MVT::v16i16, 5},
451 
452       // This shuffle is custom lowered into a long sequence of:
453       //  2x vextractf128 , 4x vpshufb , 2x vpor ,  1x vinsertf128
454       {ISD::VECTOR_SHUFFLE, MVT::v32i8, 9}
455     };
456 
457     if (ST->hasAVX())
458       if (const auto *Entry = CostTableLookup(AVXAltShuffleTbl,
459                                               ISD::VECTOR_SHUFFLE, LT.second))
460         return LT.first * Entry->Cost;
461 
462     static const CostTblEntry SSE41AltShuffleTbl[] = {
463       // These are lowered into movsd.
464       {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
465       {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
466 
467       // packed float vectors with four elements are lowered into BLENDI dag
468       // nodes. A v4i32/v4f32 BLENDI generates a single 'blendps'/'blendpd'.
469       {ISD::VECTOR_SHUFFLE, MVT::v4i32, 1},
470       {ISD::VECTOR_SHUFFLE, MVT::v4f32, 1},
471 
472       // This shuffle generates a single pshufw.
473       {ISD::VECTOR_SHUFFLE, MVT::v8i16, 1},
474 
475       // There is no instruction that matches a v16i8 alternate shuffle.
476       // The backend will expand it into the sequence 'pshufb + pshufb + or'.
477       {ISD::VECTOR_SHUFFLE, MVT::v16i8, 3}
478     };
479 
480     if (ST->hasSSE41())
481       if (const auto *Entry = CostTableLookup(SSE41AltShuffleTbl, ISD::VECTOR_SHUFFLE,
482                                               LT.second))
483         return LT.first * Entry->Cost;
484 
485     static const CostTblEntry SSSE3AltShuffleTbl[] = {
486       {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},  // movsd
487       {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},  // movsd
488 
489       // SSE3 doesn't have 'blendps'. The following shuffles are expanded into
490       // the sequence 'shufps + pshufd'
491       {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
492       {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
493 
494       {ISD::VECTOR_SHUFFLE, MVT::v8i16, 3}, // pshufb + pshufb + or
495       {ISD::VECTOR_SHUFFLE, MVT::v16i8, 3}  // pshufb + pshufb + or
496     };
497 
498     if (ST->hasSSSE3())
499       if (const auto *Entry = CostTableLookup(SSSE3AltShuffleTbl,
500                                               ISD::VECTOR_SHUFFLE, LT.second))
501         return LT.first * Entry->Cost;
502 
503     static const CostTblEntry SSEAltShuffleTbl[] = {
504       {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},  // movsd
505       {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},  // movsd
506 
507       {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2}, // shufps + pshufd
508       {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2}, // shufps + pshufd
509 
510       // This is expanded into a long sequence of four extract + four insert.
511       {ISD::VECTOR_SHUFFLE, MVT::v8i16, 8}, // 4 x pextrw + 4 pinsrw.
512 
513       // 8 x (pinsrw + pextrw + and + movb + movzb + or)
514       {ISD::VECTOR_SHUFFLE, MVT::v16i8, 48}
515     };
516 
517     // Fall-back (SSE3 and SSE2).
518     if (const auto *Entry = CostTableLookup(SSEAltShuffleTbl,
519                                             ISD::VECTOR_SHUFFLE, LT.second))
520       return LT.first * Entry->Cost;
521     return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
522   }
523 
524   return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
525 }
526 
527 int X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) {
528   int ISD = TLI->InstructionOpcodeToISD(Opcode);
529   assert(ISD && "Invalid opcode");
530 
531   // FIXME: Need a better design of the cost table to handle non-simple types of
532   // potential massive combinations (elem_num x src_type x dst_type).
533 
534   static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = {
535     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  1 },
536     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  1 },
537     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i64,  1 },
538     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64,  1 },
539     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64,  1 },
540     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64,  1 },
541 
542     { ISD::FP_TO_UINT,  MVT::v2i64, MVT::v2f32, 1 },
543     { ISD::FP_TO_UINT,  MVT::v4i64, MVT::v4f32, 1 },
544     { ISD::FP_TO_UINT,  MVT::v8i64, MVT::v8f32, 1 },
545     { ISD::FP_TO_UINT,  MVT::v2i64, MVT::v2f64, 1 },
546     { ISD::FP_TO_UINT,  MVT::v4i64, MVT::v4f64, 1 },
547     { ISD::FP_TO_UINT,  MVT::v8i64, MVT::v8f64, 1 },
548   };
549 
550   // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and
551   // 256-bit wide vectors.
552 
553   static const TypeConversionCostTblEntry AVX512FConversionTbl[] = {
554     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v8f32,  1 },
555     { ISD::FP_EXTEND, MVT::v8f64,   MVT::v16f32, 3 },
556     { ISD::FP_ROUND,  MVT::v8f32,   MVT::v8f64,  1 },
557 
558     { ISD::TRUNCATE,  MVT::v16i8,   MVT::v16i32, 1 },
559     { ISD::TRUNCATE,  MVT::v16i16,  MVT::v16i32, 1 },
560     { ISD::TRUNCATE,  MVT::v8i16,   MVT::v8i64,  1 },
561     { ISD::TRUNCATE,  MVT::v8i32,   MVT::v8i64,  1 },
562 
563     // v16i1 -> v16i32 - load + broadcast
564     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
565     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1,  2 },
566     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
567     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  1 },
568     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
569     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
570     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
571     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i16,  1 },
572     { ISD::SIGN_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
573     { ISD::ZERO_EXTEND, MVT::v8i64,  MVT::v8i32,  1 },
574 
575     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
576     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
577     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i8,   2 },
578     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i8,  2 },
579     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
580     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i16, 2 },
581     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
582     { ISD::SINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
583     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i64, 26 },
584     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64, 26 },
585 
586     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i1,   4 },
587     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i1,  3 },
588     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i8,   2 },
589     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i8,   2 },
590     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i8,   2 },
591     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i8,   2 },
592     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i8,  2 },
593     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i16,  5 },
594     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i16,  2 },
595     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i16,  2 },
596     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i16,  2 },
597     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i16, 2 },
598     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i32,  2 },
599     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i32,  1 },
600     { ISD::UINT_TO_FP,  MVT::v4f32,  MVT::v4i32,  1 },
601     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i32,  1 },
602     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  1 },
603     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i32,  1 },
604     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i32, 1 },
605     { ISD::UINT_TO_FP,  MVT::v2f32,  MVT::v2i64,  5 },
606     { ISD::UINT_TO_FP,  MVT::v2f64,  MVT::v2i64,  5 },
607     { ISD::UINT_TO_FP,  MVT::v4f64,  MVT::v4i64, 12 },
608     { ISD::UINT_TO_FP,  MVT::v8f64,  MVT::v8i64, 26 },
609 
610     { ISD::FP_TO_UINT,  MVT::v2i32,  MVT::v2f32,  1 },
611     { ISD::FP_TO_UINT,  MVT::v4i32,  MVT::v4f32,  1 },
612     { ISD::FP_TO_UINT,  MVT::v8i32,  MVT::v8f32,  1 },
613     { ISD::FP_TO_UINT,  MVT::v16i32, MVT::v16f32, 1 },
614   };
615 
616   static const TypeConversionCostTblEntry AVX2ConversionTbl[] = {
617     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
618     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,   3 },
619     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
620     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,   3 },
621     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,   3 },
622     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,   3 },
623     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   3 },
624     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   3 },
625     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
626     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  1 },
627     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
628     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
629     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
630     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  1 },
631     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
632     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  1 },
633 
634     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i64,  2 },
635     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i64,  2 },
636     { ISD::TRUNCATE,    MVT::v4i32,  MVT::v4i64,  2 },
637     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  2 },
638     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  2 },
639     { ISD::TRUNCATE,    MVT::v8i32,  MVT::v8i64,  4 },
640 
641     { ISD::FP_EXTEND,   MVT::v8f64,  MVT::v8f32,  3 },
642     { ISD::FP_ROUND,    MVT::v8f32,  MVT::v8f64,  3 },
643 
644     { ISD::UINT_TO_FP,  MVT::v8f32,  MVT::v8i32,  8 },
645   };
646 
647   static const TypeConversionCostTblEntry AVXConversionTbl[] = {
648     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i1,  6 },
649     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i1,  4 },
650     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i1,  7 },
651     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i1,  4 },
652     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,  6 },
653     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,  4 },
654     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,  7 },
655     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,  4 },
656     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
657     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
658     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16, 6 },
659     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16, 3 },
660     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16, 4 },
661     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16, 4 },
662     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32, 4 },
663     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32, 4 },
664 
665     { ISD::TRUNCATE,    MVT::v16i8, MVT::v16i16, 4 },
666     { ISD::TRUNCATE,    MVT::v8i8,  MVT::v8i32,  4 },
667     { ISD::TRUNCATE,    MVT::v8i16, MVT::v8i32,  5 },
668     { ISD::TRUNCATE,    MVT::v4i8,  MVT::v4i64,  4 },
669     { ISD::TRUNCATE,    MVT::v4i16, MVT::v4i64,  4 },
670     { ISD::TRUNCATE,    MVT::v4i32, MVT::v4i64,  4 },
671     { ISD::TRUNCATE,    MVT::v8i32, MVT::v8i64,  9 },
672 
673     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i1,  3 },
674     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i1,  3 },
675     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i1,  8 },
676     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i8,  3 },
677     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i8,  3 },
678     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i8,  8 },
679     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i16, 3 },
680     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i16, 3 },
681     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
682     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },
683     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i32, 1 },
684     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i32, 1 },
685 
686     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i1,  7 },
687     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i1,  7 },
688     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i1,  6 },
689     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i8,  2 },
690     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i8,  2 },
691     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i8,  5 },
692     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
693     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i16, 2 },
694     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i16, 5 },
695     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i32, 6 },
696     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i32, 6 },
697     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i32, 6 },
698     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i32, 9 },
699     // The generic code to compute the scalar overhead is currently broken.
700     // Workaround this limitation by estimating the scalarization overhead
701     // here. We have roughly 10 instructions per scalar element.
702     // Multiply that by the vector width.
703     // FIXME: remove that when PR19268 is fixed.
704     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i64, 10 },
705     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i64, 20 },
706     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
707     { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
708 
709     { ISD::FP_TO_SINT,  MVT::v4i8,  MVT::v4f32, 1 },
710     { ISD::FP_TO_SINT,  MVT::v8i8,  MVT::v8f32, 7 },
711     // This node is expanded into scalarized operations but BasicTTI is overly
712     // optimistic estimating its cost.  It computes 3 per element (one
713     // vector-extract, one scalar conversion and one vector-insert).  The
714     // problem is that the inserts form a read-modify-write chain so latency
715     // should be factored in too.  Inflating the cost per element by 1.
716     { ISD::FP_TO_UINT,  MVT::v8i32, MVT::v8f32, 8*4 },
717     { ISD::FP_TO_UINT,  MVT::v4i32, MVT::v4f64, 4*4 },
718 
719     { ISD::FP_EXTEND,   MVT::v4f64,  MVT::v4f32,  1 },
720     { ISD::FP_ROUND,    MVT::v4f32,  MVT::v4f64,  1 },
721   };
722 
723   static const TypeConversionCostTblEntry SSE41ConversionTbl[] = {
724     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8,    2 },
725     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8,    2 },
726     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16,   2 },
727     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16,   2 },
728     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32,   2 },
729     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32,   2 },
730 
731     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i8,   1 },
732     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i8,   2 },
733     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i8,   1 },
734     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i8,   1 },
735     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
736     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
737     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   2 },
738     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   2 },
739     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
740     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  2 },
741     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  4 },
742     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  4 },
743     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
744     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
745     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
746     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  2 },
747     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 4 },
748     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 4 },
749 
750     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  2 },
751     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  1 },
752     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i32,  1 },
753     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i32,  1 },
754     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  3 },
755     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  3 },
756     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 6 },
757 
758   };
759 
760   static const TypeConversionCostTblEntry SSE2ConversionTbl[] = {
761     // These are somewhat magic numbers justified by looking at the output of
762     // Intel's IACA, running some kernels and making sure when we take
763     // legalization into account the throughput will be overestimated.
764     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
765     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
766     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
767     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
768     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 },
769     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
770     { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
771     { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
772 
773     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
774     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
775     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
776     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
777     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
778     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 8 },
779     { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
780     { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
781 
782     { ISD::ZERO_EXTEND, MVT::v4i16,  MVT::v4i8,   1 },
783     { ISD::SIGN_EXTEND, MVT::v4i16,  MVT::v4i8,   6 },
784     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i8,   2 },
785     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i8,   3 },
786     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i8,   4 },
787     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i8,   8 },
788     { ISD::ZERO_EXTEND, MVT::v8i16,  MVT::v8i8,   1 },
789     { ISD::SIGN_EXTEND, MVT::v8i16,  MVT::v8i8,   2 },
790     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i8,   6 },
791     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i8,   6 },
792     { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8,  3 },
793     { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8,  4 },
794     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8,  9 },
795     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8,  12 },
796     { ISD::ZERO_EXTEND, MVT::v4i32,  MVT::v4i16,  1 },
797     { ISD::SIGN_EXTEND, MVT::v4i32,  MVT::v4i16,  2 },
798     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i16,  3 },
799     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i16,  10 },
800     { ISD::ZERO_EXTEND, MVT::v8i32,  MVT::v8i16,  3 },
801     { ISD::SIGN_EXTEND, MVT::v8i32,  MVT::v8i16,  4 },
802     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 6 },
803     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 8 },
804     { ISD::ZERO_EXTEND, MVT::v4i64,  MVT::v4i32,  3 },
805     { ISD::SIGN_EXTEND, MVT::v4i64,  MVT::v4i32,  5 },
806 
807     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i16,  4 },
808     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i16,  2 },
809     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i16, 3 },
810     { ISD::TRUNCATE,    MVT::v4i8,   MVT::v4i32,  3 },
811     { ISD::TRUNCATE,    MVT::v4i16,  MVT::v4i32,  3 },
812     { ISD::TRUNCATE,    MVT::v8i8,   MVT::v8i32,  4 },
813     { ISD::TRUNCATE,    MVT::v16i8,  MVT::v16i32, 7 },
814     { ISD::TRUNCATE,    MVT::v8i16,  MVT::v8i32,  5 },
815     { ISD::TRUNCATE,    MVT::v16i16, MVT::v16i32, 10 },
816   };
817 
818   std::pair<int, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src);
819   std::pair<int, MVT> LTDest = TLI->getTypeLegalizationCost(DL, Dst);
820 
821   if (ST->hasSSE2() && !ST->hasAVX()) {
822     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
823                                                    LTDest.second, LTSrc.second))
824       return LTSrc.first * Entry->Cost;
825   }
826 
827   EVT SrcTy = TLI->getValueType(DL, Src);
828   EVT DstTy = TLI->getValueType(DL, Dst);
829 
830   // The function getSimpleVT only handles simple value types.
831   if (!SrcTy.isSimple() || !DstTy.isSimple())
832     return BaseT::getCastInstrCost(Opcode, Dst, Src);
833 
834   if (ST->hasDQI())
835     if (const auto *Entry = ConvertCostTableLookup(AVX512DQConversionTbl, ISD,
836                                                    DstTy.getSimpleVT(),
837                                                    SrcTy.getSimpleVT()))
838       return Entry->Cost;
839 
840   if (ST->hasAVX512())
841     if (const auto *Entry = ConvertCostTableLookup(AVX512FConversionTbl, ISD,
842                                                    DstTy.getSimpleVT(),
843                                                    SrcTy.getSimpleVT()))
844       return Entry->Cost;
845 
846   if (ST->hasAVX2()) {
847     if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
848                                                    DstTy.getSimpleVT(),
849                                                    SrcTy.getSimpleVT()))
850       return Entry->Cost;
851   }
852 
853   if (ST->hasAVX()) {
854     if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
855                                                    DstTy.getSimpleVT(),
856                                                    SrcTy.getSimpleVT()))
857       return Entry->Cost;
858   }
859 
860   if (ST->hasSSE41()) {
861     if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
862                                                    DstTy.getSimpleVT(),
863                                                    SrcTy.getSimpleVT()))
864       return Entry->Cost;
865   }
866 
867   if (ST->hasSSE2()) {
868     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
869                                                    DstTy.getSimpleVT(),
870                                                    SrcTy.getSimpleVT()))
871       return Entry->Cost;
872   }
873 
874   return BaseT::getCastInstrCost(Opcode, Dst, Src);
875 }
876 
877 int X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) {
878   // Legalize the type.
879   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
880 
881   MVT MTy = LT.second;
882 
883   int ISD = TLI->InstructionOpcodeToISD(Opcode);
884   assert(ISD && "Invalid opcode");
885 
886   static const CostTblEntry SSE2CostTbl[] = {
887     { ISD::SETCC,   MVT::v2i64,   8 },
888     { ISD::SETCC,   MVT::v4i32,   1 },
889     { ISD::SETCC,   MVT::v8i16,   1 },
890     { ISD::SETCC,   MVT::v16i8,   1 },
891   };
892 
893   static const CostTblEntry SSE42CostTbl[] = {
894     { ISD::SETCC,   MVT::v2f64,   1 },
895     { ISD::SETCC,   MVT::v4f32,   1 },
896     { ISD::SETCC,   MVT::v2i64,   1 },
897   };
898 
899   static const CostTblEntry AVX1CostTbl[] = {
900     { ISD::SETCC,   MVT::v4f64,   1 },
901     { ISD::SETCC,   MVT::v8f32,   1 },
902     // AVX1 does not support 8-wide integer compare.
903     { ISD::SETCC,   MVT::v4i64,   4 },
904     { ISD::SETCC,   MVT::v8i32,   4 },
905     { ISD::SETCC,   MVT::v16i16,  4 },
906     { ISD::SETCC,   MVT::v32i8,   4 },
907   };
908 
909   static const CostTblEntry AVX2CostTbl[] = {
910     { ISD::SETCC,   MVT::v4i64,   1 },
911     { ISD::SETCC,   MVT::v8i32,   1 },
912     { ISD::SETCC,   MVT::v16i16,  1 },
913     { ISD::SETCC,   MVT::v32i8,   1 },
914   };
915 
916   static const CostTblEntry AVX512CostTbl[] = {
917     { ISD::SETCC,   MVT::v8i64,   1 },
918     { ISD::SETCC,   MVT::v16i32,  1 },
919     { ISD::SETCC,   MVT::v8f64,   1 },
920     { ISD::SETCC,   MVT::v16f32,  1 },
921   };
922 
923   if (ST->hasAVX512())
924     if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
925       return LT.first * Entry->Cost;
926 
927   if (ST->hasAVX2())
928     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
929       return LT.first * Entry->Cost;
930 
931   if (ST->hasAVX())
932     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
933       return LT.first * Entry->Cost;
934 
935   if (ST->hasSSE42())
936     if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
937       return LT.first * Entry->Cost;
938 
939   if (ST->hasSSE2())
940     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
941       return LT.first * Entry->Cost;
942 
943   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy);
944 }
945 
946 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
947                                       ArrayRef<Type *> Tys, FastMathFlags FMF) {
948   // Costs should match the codegen from:
949   // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll
950   // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll
951   // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll
952   static const CostTblEntry XOPCostTbl[] = {
953     { ISD::BITREVERSE, MVT::v4i64,   4 },
954     { ISD::BITREVERSE, MVT::v8i32,   4 },
955     { ISD::BITREVERSE, MVT::v16i16,  4 },
956     { ISD::BITREVERSE, MVT::v32i8,   4 },
957     { ISD::BITREVERSE, MVT::v2i64,   1 },
958     { ISD::BITREVERSE, MVT::v4i32,   1 },
959     { ISD::BITREVERSE, MVT::v8i16,   1 },
960     { ISD::BITREVERSE, MVT::v16i8,   1 },
961     { ISD::BITREVERSE, MVT::i64,     3 },
962     { ISD::BITREVERSE, MVT::i32,     3 },
963     { ISD::BITREVERSE, MVT::i16,     3 },
964     { ISD::BITREVERSE, MVT::i8,      3 }
965   };
966   static const CostTblEntry AVX2CostTbl[] = {
967     { ISD::BITREVERSE, MVT::v4i64,   5 },
968     { ISD::BITREVERSE, MVT::v8i32,   5 },
969     { ISD::BITREVERSE, MVT::v16i16,  5 },
970     { ISD::BITREVERSE, MVT::v32i8,   5 },
971     { ISD::BSWAP,      MVT::v4i64,   1 },
972     { ISD::BSWAP,      MVT::v8i32,   1 },
973     { ISD::BSWAP,      MVT::v16i16,  1 },
974     { ISD::CTPOP,      MVT::v4i64,   7 },
975     { ISD::CTPOP,      MVT::v8i32,  11 },
976     { ISD::CTPOP,      MVT::v16i16,  9 },
977     { ISD::CTPOP,      MVT::v32i8,   6 }
978   };
979   static const CostTblEntry AVX1CostTbl[] = {
980     { ISD::BITREVERSE, MVT::v4i64,  10 },
981     { ISD::BITREVERSE, MVT::v8i32,  10 },
982     { ISD::BITREVERSE, MVT::v16i16, 10 },
983     { ISD::BITREVERSE, MVT::v32i8,  10 },
984     { ISD::BSWAP,      MVT::v4i64,   4 },
985     { ISD::BSWAP,      MVT::v8i32,   4 },
986     { ISD::BSWAP,      MVT::v16i16,  4 },
987     { ISD::CTPOP,      MVT::v4i64,  14 },
988     { ISD::CTPOP,      MVT::v8i32,  22 },
989     { ISD::CTPOP,      MVT::v16i16, 18 },
990     { ISD::CTPOP,      MVT::v32i8,  12 }
991   };
992   static const CostTblEntry SSSE3CostTbl[] = {
993     { ISD::BITREVERSE, MVT::v2i64,   5 },
994     { ISD::BITREVERSE, MVT::v4i32,   5 },
995     { ISD::BITREVERSE, MVT::v8i16,   5 },
996     { ISD::BITREVERSE, MVT::v16i8,   5 },
997     { ISD::BSWAP,      MVT::v2i64,   1 },
998     { ISD::BSWAP,      MVT::v4i32,   1 },
999     { ISD::BSWAP,      MVT::v8i16,   1 },
1000     { ISD::CTPOP,      MVT::v2i64,   7 },
1001     { ISD::CTPOP,      MVT::v4i32,  11 },
1002     { ISD::CTPOP,      MVT::v8i16,   9 },
1003     { ISD::CTPOP,      MVT::v16i8,   6 }
1004   };
1005   static const CostTblEntry SSE2CostTbl[] = {
1006     { ISD::BSWAP,      MVT::v2i64,   7 },
1007     { ISD::BSWAP,      MVT::v4i32,   7 },
1008     { ISD::BSWAP,      MVT::v8i16,   7 },
1009     { ISD::CTPOP,      MVT::v2i64,  12 },
1010     { ISD::CTPOP,      MVT::v4i32,  15 },
1011     { ISD::CTPOP,      MVT::v8i16,  13 },
1012     { ISD::CTPOP,      MVT::v16i8,  10 }
1013   };
1014 
1015   unsigned ISD = ISD::DELETED_NODE;
1016   switch (IID) {
1017   default:
1018     break;
1019   case Intrinsic::bitreverse:
1020     ISD = ISD::BITREVERSE;
1021     break;
1022   case Intrinsic::bswap:
1023     ISD = ISD::BSWAP;
1024     break;
1025   case Intrinsic::ctpop:
1026     ISD = ISD::CTPOP;
1027     break;
1028   }
1029 
1030   // Legalize the type.
1031   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
1032   MVT MTy = LT.second;
1033 
1034   // Attempt to lookup cost.
1035   if (ST->hasXOP())
1036     if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
1037       return LT.first * Entry->Cost;
1038 
1039   if (ST->hasAVX2())
1040     if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
1041       return LT.first * Entry->Cost;
1042 
1043   if (ST->hasAVX())
1044     if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
1045       return LT.first * Entry->Cost;
1046 
1047   if (ST->hasSSSE3())
1048     if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy))
1049       return LT.first * Entry->Cost;
1050 
1051   if (ST->hasSSE2())
1052     if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
1053       return LT.first * Entry->Cost;
1054 
1055   return BaseT::getIntrinsicInstrCost(IID, RetTy, Tys, FMF);
1056 }
1057 
1058 int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1059                                       ArrayRef<Value *> Args, FastMathFlags FMF) {
1060   return BaseT::getIntrinsicInstrCost(IID, RetTy, Args, FMF);
1061 }
1062 
1063 int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
1064   assert(Val->isVectorTy() && "This must be a vector type");
1065 
1066   Type *ScalarType = Val->getScalarType();
1067 
1068   if (Index != -1U) {
1069     // Legalize the type.
1070     std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Val);
1071 
1072     // This type is legalized to a scalar type.
1073     if (!LT.second.isVector())
1074       return 0;
1075 
1076     // The type may be split. Normalize the index to the new type.
1077     unsigned Width = LT.second.getVectorNumElements();
1078     Index = Index % Width;
1079 
1080     // Floating point scalars are already located in index #0.
1081     if (ScalarType->isFloatingPointTy() && Index == 0)
1082       return 0;
1083   }
1084 
1085   // Add to the base cost if we know that the extracted element of a vector is
1086   // destined to be moved to and used in the integer register file.
1087   int RegisterFileMoveCost = 0;
1088   if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy())
1089     RegisterFileMoveCost = 1;
1090 
1091   return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost;
1092 }
1093 
1094 int X86TTIImpl::getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
1095   assert (Ty->isVectorTy() && "Can only scalarize vectors");
1096   int Cost = 0;
1097 
1098   for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
1099     if (Insert)
1100       Cost += getVectorInstrCost(Instruction::InsertElement, Ty, i);
1101     if (Extract)
1102       Cost += getVectorInstrCost(Instruction::ExtractElement, Ty, i);
1103   }
1104 
1105   return Cost;
1106 }
1107 
1108 int X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
1109                                 unsigned AddressSpace) {
1110   // Handle non-power-of-two vectors such as <3 x float>
1111   if (VectorType *VTy = dyn_cast<VectorType>(Src)) {
1112     unsigned NumElem = VTy->getVectorNumElements();
1113 
1114     // Handle a few common cases:
1115     // <3 x float>
1116     if (NumElem == 3 && VTy->getScalarSizeInBits() == 32)
1117       // Cost = 64 bit store + extract + 32 bit store.
1118       return 3;
1119 
1120     // <3 x double>
1121     if (NumElem == 3 && VTy->getScalarSizeInBits() == 64)
1122       // Cost = 128 bit store + unpack + 64 bit store.
1123       return 3;
1124 
1125     // Assume that all other non-power-of-two numbers are scalarized.
1126     if (!isPowerOf2_32(NumElem)) {
1127       int Cost = BaseT::getMemoryOpCost(Opcode, VTy->getScalarType(), Alignment,
1128                                         AddressSpace);
1129       int SplitCost = getScalarizationOverhead(Src, Opcode == Instruction::Load,
1130                                                Opcode == Instruction::Store);
1131       return NumElem * Cost + SplitCost;
1132     }
1133   }
1134 
1135   // Legalize the type.
1136   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
1137   assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
1138          "Invalid Opcode");
1139 
1140   // Each load/store unit costs 1.
1141   int Cost = LT.first * 1;
1142 
1143   // This isn't exactly right. We're using slow unaligned 32-byte accesses as a
1144   // proxy for a double-pumped AVX memory interface such as on Sandybridge.
1145   if (LT.second.getStoreSize() == 32 && ST->isUnalignedMem32Slow())
1146     Cost *= 2;
1147 
1148   return Cost;
1149 }
1150 
1151 int X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy,
1152                                       unsigned Alignment,
1153                                       unsigned AddressSpace) {
1154   VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy);
1155   if (!SrcVTy)
1156     // To calculate scalar take the regular cost, without mask
1157     return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace);
1158 
1159   unsigned NumElem = SrcVTy->getVectorNumElements();
1160   VectorType *MaskTy =
1161     VectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem);
1162   if ((Opcode == Instruction::Load && !isLegalMaskedLoad(SrcVTy)) ||
1163       (Opcode == Instruction::Store && !isLegalMaskedStore(SrcVTy)) ||
1164       !isPowerOf2_32(NumElem)) {
1165     // Scalarization
1166     int MaskSplitCost = getScalarizationOverhead(MaskTy, false, true);
1167     int ScalarCompareCost = getCmpSelInstrCost(
1168         Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr);
1169     int BranchCost = getCFInstrCost(Instruction::Br);
1170     int MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost);
1171 
1172     int ValueSplitCost = getScalarizationOverhead(
1173         SrcVTy, Opcode == Instruction::Load, Opcode == Instruction::Store);
1174     int MemopCost =
1175         NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
1176                                          Alignment, AddressSpace);
1177     return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost;
1178   }
1179 
1180   // Legalize the type.
1181   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy);
1182   auto VT = TLI->getValueType(DL, SrcVTy);
1183   int Cost = 0;
1184   if (VT.isSimple() && LT.second != VT.getSimpleVT() &&
1185       LT.second.getVectorNumElements() == NumElem)
1186     // Promotion requires expand/truncate for data and a shuffle for mask.
1187     Cost += getShuffleCost(TTI::SK_Alternate, SrcVTy, 0, nullptr) +
1188             getShuffleCost(TTI::SK_Alternate, MaskTy, 0, nullptr);
1189 
1190   else if (LT.second.getVectorNumElements() > NumElem) {
1191     VectorType *NewMaskTy = VectorType::get(MaskTy->getVectorElementType(),
1192                                             LT.second.getVectorNumElements());
1193     // Expanding requires fill mask with zeroes
1194     Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, 0, MaskTy);
1195   }
1196   if (!ST->hasAVX512())
1197     return Cost + LT.first*4; // Each maskmov costs 4
1198 
1199   // AVX-512 masked load/store is cheapper
1200   return Cost+LT.first;
1201 }
1202 
1203 int X86TTIImpl::getAddressComputationCost(Type *Ty, bool IsComplex) {
1204   // Address computations in vectorized code with non-consecutive addresses will
1205   // likely result in more instructions compared to scalar code where the
1206   // computation can more often be merged into the index mode. The resulting
1207   // extra micro-ops can significantly decrease throughput.
1208   unsigned NumVectorInstToHideOverhead = 10;
1209 
1210   if (Ty->isVectorTy() && IsComplex)
1211     return NumVectorInstToHideOverhead;
1212 
1213   return BaseT::getAddressComputationCost(Ty, IsComplex);
1214 }
1215 
1216 int X86TTIImpl::getReductionCost(unsigned Opcode, Type *ValTy,
1217                                  bool IsPairwise) {
1218 
1219   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
1220 
1221   MVT MTy = LT.second;
1222 
1223   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1224   assert(ISD && "Invalid opcode");
1225 
1226   // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
1227   // and make it as the cost.
1228 
1229   static const CostTblEntry SSE42CostTblPairWise[] = {
1230     { ISD::FADD,  MVT::v2f64,   2 },
1231     { ISD::FADD,  MVT::v4f32,   4 },
1232     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
1233     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.5".
1234     { ISD::ADD,   MVT::v8i16,   5 },
1235   };
1236 
1237   static const CostTblEntry AVX1CostTblPairWise[] = {
1238     { ISD::FADD,  MVT::v4f32,   4 },
1239     { ISD::FADD,  MVT::v4f64,   5 },
1240     { ISD::FADD,  MVT::v8f32,   7 },
1241     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
1242     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.5".
1243     { ISD::ADD,   MVT::v4i64,   5 },      // The data reported by the IACA tool is "4.8".
1244     { ISD::ADD,   MVT::v8i16,   5 },
1245     { ISD::ADD,   MVT::v8i32,   5 },
1246   };
1247 
1248   static const CostTblEntry SSE42CostTblNoPairWise[] = {
1249     { ISD::FADD,  MVT::v2f64,   2 },
1250     { ISD::FADD,  MVT::v4f32,   4 },
1251     { ISD::ADD,   MVT::v2i64,   2 },      // The data reported by the IACA tool is "1.6".
1252     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "3.3".
1253     { ISD::ADD,   MVT::v8i16,   4 },      // The data reported by the IACA tool is "4.3".
1254   };
1255 
1256   static const CostTblEntry AVX1CostTblNoPairWise[] = {
1257     { ISD::FADD,  MVT::v4f32,   3 },
1258     { ISD::FADD,  MVT::v4f64,   3 },
1259     { ISD::FADD,  MVT::v8f32,   4 },
1260     { ISD::ADD,   MVT::v2i64,   1 },      // The data reported by the IACA tool is "1.5".
1261     { ISD::ADD,   MVT::v4i32,   3 },      // The data reported by the IACA tool is "2.8".
1262     { ISD::ADD,   MVT::v4i64,   3 },
1263     { ISD::ADD,   MVT::v8i16,   4 },
1264     { ISD::ADD,   MVT::v8i32,   5 },
1265   };
1266 
1267   if (IsPairwise) {
1268     if (ST->hasAVX())
1269       if (const auto *Entry = CostTableLookup(AVX1CostTblPairWise, ISD, MTy))
1270         return LT.first * Entry->Cost;
1271 
1272     if (ST->hasSSE42())
1273       if (const auto *Entry = CostTableLookup(SSE42CostTblPairWise, ISD, MTy))
1274         return LT.first * Entry->Cost;
1275   } else {
1276     if (ST->hasAVX())
1277       if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
1278         return LT.first * Entry->Cost;
1279 
1280     if (ST->hasSSE42())
1281       if (const auto *Entry = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy))
1282         return LT.first * Entry->Cost;
1283   }
1284 
1285   return BaseT::getReductionCost(Opcode, ValTy, IsPairwise);
1286 }
1287 
1288 /// \brief Calculate the cost of materializing a 64-bit value. This helper
1289 /// method might only calculate a fraction of a larger immediate. Therefore it
1290 /// is valid to return a cost of ZERO.
1291 int X86TTIImpl::getIntImmCost(int64_t Val) {
1292   if (Val == 0)
1293     return TTI::TCC_Free;
1294 
1295   if (isInt<32>(Val))
1296     return TTI::TCC_Basic;
1297 
1298   return 2 * TTI::TCC_Basic;
1299 }
1300 
1301 int X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) {
1302   assert(Ty->isIntegerTy());
1303 
1304   unsigned BitSize = Ty->getPrimitiveSizeInBits();
1305   if (BitSize == 0)
1306     return ~0U;
1307 
1308   // Never hoist constants larger than 128bit, because this might lead to
1309   // incorrect code generation or assertions in codegen.
1310   // Fixme: Create a cost model for types larger than i128 once the codegen
1311   // issues have been fixed.
1312   if (BitSize > 128)
1313     return TTI::TCC_Free;
1314 
1315   if (Imm == 0)
1316     return TTI::TCC_Free;
1317 
1318   // Sign-extend all constants to a multiple of 64-bit.
1319   APInt ImmVal = Imm;
1320   if (BitSize & 0x3f)
1321     ImmVal = Imm.sext((BitSize + 63) & ~0x3fU);
1322 
1323   // Split the constant into 64-bit chunks and calculate the cost for each
1324   // chunk.
1325   int Cost = 0;
1326   for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
1327     APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
1328     int64_t Val = Tmp.getSExtValue();
1329     Cost += getIntImmCost(Val);
1330   }
1331   // We need at least one instruction to materialize the constant.
1332   return std::max(1, Cost);
1333 }
1334 
1335 int X86TTIImpl::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
1336                               Type *Ty) {
1337   assert(Ty->isIntegerTy());
1338 
1339   unsigned BitSize = Ty->getPrimitiveSizeInBits();
1340   // There is no cost model for constants with a bit size of 0. Return TCC_Free
1341   // here, so that constant hoisting will ignore this constant.
1342   if (BitSize == 0)
1343     return TTI::TCC_Free;
1344 
1345   unsigned ImmIdx = ~0U;
1346   switch (Opcode) {
1347   default:
1348     return TTI::TCC_Free;
1349   case Instruction::GetElementPtr:
1350     // Always hoist the base address of a GetElementPtr. This prevents the
1351     // creation of new constants for every base constant that gets constant
1352     // folded with the offset.
1353     if (Idx == 0)
1354       return 2 * TTI::TCC_Basic;
1355     return TTI::TCC_Free;
1356   case Instruction::Store:
1357     ImmIdx = 0;
1358     break;
1359   case Instruction::ICmp:
1360     // This is an imperfect hack to prevent constant hoisting of
1361     // compares that might be trying to check if a 64-bit value fits in
1362     // 32-bits. The backend can optimize these cases using a right shift by 32.
1363     // Ideally we would check the compare predicate here. There also other
1364     // similar immediates the backend can use shifts for.
1365     if (Idx == 1 && Imm.getBitWidth() == 64) {
1366       uint64_t ImmVal = Imm.getZExtValue();
1367       if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff)
1368         return TTI::TCC_Free;
1369     }
1370     ImmIdx = 1;
1371     break;
1372   case Instruction::And:
1373     // We support 64-bit ANDs with immediates with 32-bits of leading zeroes
1374     // by using a 32-bit operation with implicit zero extension. Detect such
1375     // immediates here as the normal path expects bit 31 to be sign extended.
1376     if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue()))
1377       return TTI::TCC_Free;
1378     // Fallthrough
1379   case Instruction::Add:
1380   case Instruction::Sub:
1381   case Instruction::Mul:
1382   case Instruction::UDiv:
1383   case Instruction::SDiv:
1384   case Instruction::URem:
1385   case Instruction::SRem:
1386   case Instruction::Or:
1387   case Instruction::Xor:
1388     ImmIdx = 1;
1389     break;
1390   // Always return TCC_Free for the shift value of a shift instruction.
1391   case Instruction::Shl:
1392   case Instruction::LShr:
1393   case Instruction::AShr:
1394     if (Idx == 1)
1395       return TTI::TCC_Free;
1396     break;
1397   case Instruction::Trunc:
1398   case Instruction::ZExt:
1399   case Instruction::SExt:
1400   case Instruction::IntToPtr:
1401   case Instruction::PtrToInt:
1402   case Instruction::BitCast:
1403   case Instruction::PHI:
1404   case Instruction::Call:
1405   case Instruction::Select:
1406   case Instruction::Ret:
1407   case Instruction::Load:
1408     break;
1409   }
1410 
1411   if (Idx == ImmIdx) {
1412     int NumConstants = (BitSize + 63) / 64;
1413     int Cost = X86TTIImpl::getIntImmCost(Imm, Ty);
1414     return (Cost <= NumConstants * TTI::TCC_Basic)
1415                ? static_cast<int>(TTI::TCC_Free)
1416                : Cost;
1417   }
1418 
1419   return X86TTIImpl::getIntImmCost(Imm, Ty);
1420 }
1421 
1422 int X86TTIImpl::getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
1423                               Type *Ty) {
1424   assert(Ty->isIntegerTy());
1425 
1426   unsigned BitSize = Ty->getPrimitiveSizeInBits();
1427   // There is no cost model for constants with a bit size of 0. Return TCC_Free
1428   // here, so that constant hoisting will ignore this constant.
1429   if (BitSize == 0)
1430     return TTI::TCC_Free;
1431 
1432   switch (IID) {
1433   default:
1434     return TTI::TCC_Free;
1435   case Intrinsic::sadd_with_overflow:
1436   case Intrinsic::uadd_with_overflow:
1437   case Intrinsic::ssub_with_overflow:
1438   case Intrinsic::usub_with_overflow:
1439   case Intrinsic::smul_with_overflow:
1440   case Intrinsic::umul_with_overflow:
1441     if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
1442       return TTI::TCC_Free;
1443     break;
1444   case Intrinsic::experimental_stackmap:
1445     if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
1446       return TTI::TCC_Free;
1447     break;
1448   case Intrinsic::experimental_patchpoint_void:
1449   case Intrinsic::experimental_patchpoint_i64:
1450     if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
1451       return TTI::TCC_Free;
1452     break;
1453   }
1454   return X86TTIImpl::getIntImmCost(Imm, Ty);
1455 }
1456 
1457 // Return an average cost of Gather / Scatter instruction, maybe improved later
1458 int X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy, Value *Ptr,
1459                                 unsigned Alignment, unsigned AddressSpace) {
1460 
1461   assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost");
1462   unsigned VF = SrcVTy->getVectorNumElements();
1463 
1464   // Try to reduce index size from 64 bit (default for GEP)
1465   // to 32. It is essential for VF 16. If the index can't be reduced to 32, the
1466   // operation will use 16 x 64 indices which do not fit in a zmm and needs
1467   // to split. Also check that the base pointer is the same for all lanes,
1468   // and that there's at most one variable index.
1469   auto getIndexSizeInBits = [](Value *Ptr, const DataLayout& DL) {
1470     unsigned IndexSize = DL.getPointerSizeInBits();
1471     GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
1472     if (IndexSize < 64 || !GEP)
1473       return IndexSize;
1474 
1475     unsigned NumOfVarIndices = 0;
1476     Value *Ptrs = GEP->getPointerOperand();
1477     if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs))
1478       return IndexSize;
1479     for (unsigned i = 1; i < GEP->getNumOperands(); ++i) {
1480       if (isa<Constant>(GEP->getOperand(i)))
1481         continue;
1482       Type *IndxTy = GEP->getOperand(i)->getType();
1483       if (IndxTy->isVectorTy())
1484         IndxTy = IndxTy->getVectorElementType();
1485       if ((IndxTy->getPrimitiveSizeInBits() == 64 &&
1486           !isa<SExtInst>(GEP->getOperand(i))) ||
1487          ++NumOfVarIndices > 1)
1488         return IndexSize; // 64
1489     }
1490     return (unsigned)32;
1491   };
1492 
1493 
1494   // Trying to reduce IndexSize to 32 bits for vector 16.
1495   // By default the IndexSize is equal to pointer size.
1496   unsigned IndexSize = (VF >= 16) ? getIndexSizeInBits(Ptr, DL) :
1497     DL.getPointerSizeInBits();
1498 
1499   Type *IndexVTy = VectorType::get(IntegerType::get(SrcVTy->getContext(),
1500                                                     IndexSize), VF);
1501   std::pair<int, MVT> IdxsLT = TLI->getTypeLegalizationCost(DL, IndexVTy);
1502   std::pair<int, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, SrcVTy);
1503   int SplitFactor = std::max(IdxsLT.first, SrcLT.first);
1504   if (SplitFactor > 1) {
1505     // Handle splitting of vector of pointers
1506     Type *SplitSrcTy = VectorType::get(SrcVTy->getScalarType(), VF / SplitFactor);
1507     return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment,
1508                                          AddressSpace);
1509   }
1510 
1511   // The gather / scatter cost is given by Intel architects. It is a rough
1512   // number since we are looking at one instruction in a time.
1513   const int GSOverhead = 2;
1514   return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
1515                                            Alignment, AddressSpace);
1516 }
1517 
1518 /// Return the cost of full scalarization of gather / scatter operation.
1519 ///
1520 /// Opcode - Load or Store instruction.
1521 /// SrcVTy - The type of the data vector that should be gathered or scattered.
1522 /// VariableMask - The mask is non-constant at compile time.
1523 /// Alignment - Alignment for one element.
1524 /// AddressSpace - pointer[s] address space.
1525 ///
1526 int X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy,
1527                                 bool VariableMask, unsigned Alignment,
1528                                 unsigned AddressSpace) {
1529   unsigned VF = SrcVTy->getVectorNumElements();
1530 
1531   int MaskUnpackCost = 0;
1532   if (VariableMask) {
1533     VectorType *MaskTy =
1534       VectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF);
1535     MaskUnpackCost = getScalarizationOverhead(MaskTy, false, true);
1536     int ScalarCompareCost =
1537       getCmpSelInstrCost(Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()),
1538                          nullptr);
1539     int BranchCost = getCFInstrCost(Instruction::Br);
1540     MaskUnpackCost += VF * (BranchCost + ScalarCompareCost);
1541   }
1542 
1543   // The cost of the scalar loads/stores.
1544   int MemoryOpCost = VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
1545                                           Alignment, AddressSpace);
1546 
1547   int InsertExtractCost = 0;
1548   if (Opcode == Instruction::Load)
1549     for (unsigned i = 0; i < VF; ++i)
1550       // Add the cost of inserting each scalar load into the vector
1551       InsertExtractCost +=
1552         getVectorInstrCost(Instruction::InsertElement, SrcVTy, i);
1553   else
1554     for (unsigned i = 0; i < VF; ++i)
1555       // Add the cost of extracting each element out of the data vector
1556       InsertExtractCost +=
1557         getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i);
1558 
1559   return MemoryOpCost + MaskUnpackCost + InsertExtractCost;
1560 }
1561 
1562 /// Calculate the cost of Gather / Scatter operation
1563 int X86TTIImpl::getGatherScatterOpCost(unsigned Opcode, Type *SrcVTy,
1564                                        Value *Ptr, bool VariableMask,
1565                                        unsigned Alignment) {
1566   assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter");
1567   unsigned VF = SrcVTy->getVectorNumElements();
1568   PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
1569   if (!PtrTy && Ptr->getType()->isVectorTy())
1570     PtrTy = dyn_cast<PointerType>(Ptr->getType()->getVectorElementType());
1571   assert(PtrTy && "Unexpected type for Ptr argument");
1572   unsigned AddressSpace = PtrTy->getAddressSpace();
1573 
1574   bool Scalarize = false;
1575   if ((Opcode == Instruction::Load && !isLegalMaskedGather(SrcVTy)) ||
1576       (Opcode == Instruction::Store && !isLegalMaskedScatter(SrcVTy)))
1577     Scalarize = true;
1578   // Gather / Scatter for vector 2 is not profitable on KNL / SKX
1579   // Vector-4 of gather/scatter instruction does not exist on KNL.
1580   // We can extend it to 8 elements, but zeroing upper bits of
1581   // the mask vector will add more instructions. Right now we give the scalar
1582   // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter instruction is
1583   // better in the VariableMask case.
1584   if (VF == 2 || (VF == 4 && !ST->hasVLX()))
1585     Scalarize = true;
1586 
1587   if (Scalarize)
1588     return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment, AddressSpace);
1589 
1590   return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace);
1591 }
1592 
1593 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy) {
1594   Type *ScalarTy = DataTy->getScalarType();
1595   int DataWidth = isa<PointerType>(ScalarTy) ?
1596     DL.getPointerSizeInBits() : ScalarTy->getPrimitiveSizeInBits();
1597 
1598   return ((DataWidth == 32 || DataWidth == 64) && ST->hasAVX()) ||
1599          ((DataWidth == 8 || DataWidth == 16) && ST->hasBWI());
1600 }
1601 
1602 bool X86TTIImpl::isLegalMaskedStore(Type *DataType) {
1603   return isLegalMaskedLoad(DataType);
1604 }
1605 
1606 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy) {
1607   // This function is called now in two cases: from the Loop Vectorizer
1608   // and from the Scalarizer.
1609   // When the Loop Vectorizer asks about legality of the feature,
1610   // the vectorization factor is not calculated yet. The Loop Vectorizer
1611   // sends a scalar type and the decision is based on the width of the
1612   // scalar element.
1613   // Later on, the cost model will estimate usage this intrinsic based on
1614   // the vector type.
1615   // The Scalarizer asks again about legality. It sends a vector type.
1616   // In this case we can reject non-power-of-2 vectors.
1617   if (isa<VectorType>(DataTy) && !isPowerOf2_32(DataTy->getVectorNumElements()))
1618     return false;
1619   Type *ScalarTy = DataTy->getScalarType();
1620   int DataWidth = isa<PointerType>(ScalarTy) ?
1621     DL.getPointerSizeInBits() : ScalarTy->getPrimitiveSizeInBits();
1622 
1623   // AVX-512 allows gather and scatter
1624   return (DataWidth == 32 || DataWidth == 64) && ST->hasAVX512();
1625 }
1626 
1627 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType) {
1628   return isLegalMaskedGather(DataType);
1629 }
1630 
1631 bool X86TTIImpl::areInlineCompatible(const Function *Caller,
1632                                      const Function *Callee) const {
1633   const TargetMachine &TM = getTLI()->getTargetMachine();
1634 
1635   // Work this as a subsetting of subtarget features.
1636   const FeatureBitset &CallerBits =
1637       TM.getSubtargetImpl(*Caller)->getFeatureBits();
1638   const FeatureBitset &CalleeBits =
1639       TM.getSubtargetImpl(*Callee)->getFeatureBits();
1640 
1641   // FIXME: This is likely too limiting as it will include subtarget features
1642   // that we might not care about for inlining, but it is conservatively
1643   // correct.
1644   return (CallerBits & CalleeBits) == CalleeBits;
1645 }
1646