This repository was archived by the owner on Feb 5, 2019. It is now read-only.
forked from luqmana/llvm
-
Notifications
You must be signed in to change notification settings - Fork 63
/
Copy pathX86TargetTransformInfo.cpp
3057 lines (2652 loc) · 127 KB
/
X86TargetTransformInfo.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
//===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
/// \file
/// This file implements a TargetTransformInfo analysis pass specific to the
/// X86 target machine. It uses the target's detailed information to provide
/// more precise answers to certain TTI queries, while letting the target
/// independent and default TTI implementations handle the rest.
///
//===----------------------------------------------------------------------===//
/// About Cost Model numbers used below it's necessary to say the following:
/// the numbers correspond to some "generic" X86 CPU instead of usage of
/// concrete CPU model. Usually the numbers correspond to CPU where the feature
/// apeared at the first time. For example, if we do Subtarget.hasSSE42() in
/// the lookups below the cost is based on Nehalem as that was the first CPU
/// to support that feature level and thus has most likely the worst case cost.
/// Some examples of other technologies/CPUs:
/// SSE 3 - Pentium4 / Athlon64
/// SSE 4.1 - Penryn
/// SSE 4.2 - Nehalem
/// AVX - Sandy Bridge
/// AVX2 - Haswell
/// AVX-512 - Xeon Phi / Skylake
/// And some examples of instruction target dependent costs (latency)
/// divss sqrtss rsqrtss
/// AMD K7 11-16 19 3
/// Piledriver 9-24 13-15 5
/// Jaguar 14 16 2
/// Pentium II,III 18 30 2
/// Nehalem 7-14 7-18 3
/// Haswell 10-13 11 5
/// TODO: Develop and implement the target dependent cost model and
/// specialize cost numbers for different Cost Model Targets such as throughput,
/// code size, latency and uop count.
//===----------------------------------------------------------------------===//
#include "X86TargetTransformInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/CodeGen/BasicTTIImpl.h"
#include "llvm/CodeGen/CostTable.h"
#include "llvm/CodeGen/TargetLowering.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/Support/Debug.h"
using namespace llvm;
#define DEBUG_TYPE "x86tti"
//===----------------------------------------------------------------------===//
//
// X86 cost model.
//
//===----------------------------------------------------------------------===//
TargetTransformInfo::PopcntSupportKind
X86TTIImpl::getPopcntSupport(unsigned TyWidth) {
assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
// TODO: Currently the __builtin_popcount() implementation using SSE3
// instructions is inefficient. Once the problem is fixed, we should
// call ST->hasSSE3() instead of ST->hasPOPCNT().
return ST->hasPOPCNT() ? TTI::PSK_FastHardware : TTI::PSK_Software;
}
llvm::Optional<unsigned> X86TTIImpl::getCacheSize(
TargetTransformInfo::CacheLevel Level) const {
switch (Level) {
case TargetTransformInfo::CacheLevel::L1D:
// - Penryn
// - Nehalem
// - Westmere
// - Sandy Bridge
// - Ivy Bridge
// - Haswell
// - Broadwell
// - Skylake
// - Kabylake
return 32 * 1024; // 32 KByte
case TargetTransformInfo::CacheLevel::L2D:
// - Penryn
// - Nehalem
// - Westmere
// - Sandy Bridge
// - Ivy Bridge
// - Haswell
// - Broadwell
// - Skylake
// - Kabylake
return 256 * 1024; // 256 KByte
}
llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
}
llvm::Optional<unsigned> X86TTIImpl::getCacheAssociativity(
TargetTransformInfo::CacheLevel Level) const {
// - Penryn
// - Nehalem
// - Westmere
// - Sandy Bridge
// - Ivy Bridge
// - Haswell
// - Broadwell
// - Skylake
// - Kabylake
switch (Level) {
case TargetTransformInfo::CacheLevel::L1D:
LLVM_FALLTHROUGH;
case TargetTransformInfo::CacheLevel::L2D:
return 8;
}
llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
}
unsigned X86TTIImpl::getNumberOfRegisters(bool Vector) {
if (Vector && !ST->hasSSE1())
return 0;
if (ST->is64Bit()) {
if (Vector && ST->hasAVX512())
return 32;
return 16;
}
return 8;
}
unsigned X86TTIImpl::getRegisterBitWidth(bool Vector) const {
unsigned PreferVectorWidth = ST->getPreferVectorWidth();
if (Vector) {
if (ST->hasAVX512() && PreferVectorWidth >= 512)
return 512;
if (ST->hasAVX() && PreferVectorWidth >= 256)
return 256;
if (ST->hasSSE1() && PreferVectorWidth >= 128)
return 128;
return 0;
}
if (ST->is64Bit())
return 64;
return 32;
}
unsigned X86TTIImpl::getLoadStoreVecRegBitWidth(unsigned) const {
return getRegisterBitWidth(true);
}
unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) {
// If the loop will not be vectorized, don't interleave the loop.
// Let regular unroll to unroll the loop, which saves the overflow
// check and memory check cost.
if (VF == 1)
return 1;
if (ST->isAtom())
return 1;
// Sandybridge and Haswell have multiple execution ports and pipelined
// vector units.
if (ST->hasAVX())
return 4;
return 2;
}
int X86TTIImpl::getArithmeticInstrCost(
unsigned Opcode, Type *Ty,
TTI::OperandValueKind Op1Info, TTI::OperandValueKind Op2Info,
TTI::OperandValueProperties Opd1PropInfo,
TTI::OperandValueProperties Opd2PropInfo,
ArrayRef<const Value *> Args) {
// Legalize the type.
std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
int ISD = TLI->InstructionOpcodeToISD(Opcode);
assert(ISD && "Invalid opcode");
static const CostTblEntry GLMCostTable[] = {
{ ISD::FDIV, MVT::f32, 18 }, // divss
{ ISD::FDIV, MVT::v4f32, 35 }, // divps
{ ISD::FDIV, MVT::f64, 33 }, // divsd
{ ISD::FDIV, MVT::v2f64, 65 }, // divpd
};
if (ST->isGLM())
if (const auto *Entry = CostTableLookup(GLMCostTable, ISD,
LT.second))
return LT.first * Entry->Cost;
static const CostTblEntry SLMCostTable[] = {
{ ISD::MUL, MVT::v4i32, 11 }, // pmulld
{ ISD::MUL, MVT::v8i16, 2 }, // pmullw
{ ISD::MUL, MVT::v16i8, 14 }, // extend/pmullw/trunc sequence.
{ ISD::FMUL, MVT::f64, 2 }, // mulsd
{ ISD::FMUL, MVT::v2f64, 4 }, // mulpd
{ ISD::FMUL, MVT::v4f32, 2 }, // mulps
{ ISD::FDIV, MVT::f32, 17 }, // divss
{ ISD::FDIV, MVT::v4f32, 39 }, // divps
{ ISD::FDIV, MVT::f64, 32 }, // divsd
{ ISD::FDIV, MVT::v2f64, 69 }, // divpd
{ ISD::FADD, MVT::v2f64, 2 }, // addpd
{ ISD::FSUB, MVT::v2f64, 2 }, // subpd
// v2i64/v4i64 mul is custom lowered as a series of long:
// multiplies(3), shifts(3) and adds(2)
// slm muldq version throughput is 2 and addq throughput 4
// thus: 3X2 (muldq throughput) + 3X1 (shift throughput) +
// 3X4 (addq throughput) = 17
{ ISD::MUL, MVT::v2i64, 17 },
// slm addq\subq throughput is 4
{ ISD::ADD, MVT::v2i64, 4 },
{ ISD::SUB, MVT::v2i64, 4 },
};
if (ST->isSLM()) {
if (Args.size() == 2 && ISD == ISD::MUL && LT.second == MVT::v4i32) {
// Check if the operands can be shrinked into a smaller datatype.
bool Op1Signed = false;
unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed);
bool Op2Signed = false;
unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed);
bool signedMode = Op1Signed | Op2Signed;
unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize);
if (OpMinSize <= 7)
return LT.first * 3; // pmullw/sext
if (!signedMode && OpMinSize <= 8)
return LT.first * 3; // pmullw/zext
if (OpMinSize <= 15)
return LT.first * 5; // pmullw/pmulhw/pshuf
if (!signedMode && OpMinSize <= 16)
return LT.first * 5; // pmullw/pmulhw/pshuf
}
if (const auto *Entry = CostTableLookup(SLMCostTable, ISD,
LT.second)) {
return LT.first * Entry->Cost;
}
}
if ((ISD == ISD::SDIV || ISD == ISD::SREM || ISD == ISD::UDIV ||
ISD == ISD::UREM) &&
(Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
if (ISD == ISD::SDIV || ISD == ISD::SREM) {
// On X86, vector signed division by constants power-of-two are
// normally expanded to the sequence SRA + SRL + ADD + SRA.
// The OperandValue properties may not be the same as that of the previous
// operation; conservatively assume OP_None.
int Cost =
2 * getArithmeticInstrCost(Instruction::AShr, Ty, Op1Info, Op2Info,
TargetTransformInfo::OP_None,
TargetTransformInfo::OP_None);
Cost += getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info,
TargetTransformInfo::OP_None,
TargetTransformInfo::OP_None);
Cost += getArithmeticInstrCost(Instruction::Add, Ty, Op1Info, Op2Info,
TargetTransformInfo::OP_None,
TargetTransformInfo::OP_None);
if (ISD == ISD::SREM) {
// For SREM: (X % C) is the equivalent of (X - (X/C)*C)
Cost += getArithmeticInstrCost(Instruction::Mul, Ty, Op1Info, Op2Info);
Cost += getArithmeticInstrCost(Instruction::Sub, Ty, Op1Info, Op2Info);
}
return Cost;
}
// Vector unsigned division/remainder will be simplified to shifts/masks.
if (ISD == ISD::UDIV)
return getArithmeticInstrCost(Instruction::LShr, Ty, Op1Info, Op2Info,
TargetTransformInfo::OP_None,
TargetTransformInfo::OP_None);
if (ISD == ISD::UREM)
return getArithmeticInstrCost(Instruction::And, Ty, Op1Info, Op2Info,
TargetTransformInfo::OP_None,
TargetTransformInfo::OP_None);
}
static const CostTblEntry AVX512BWUniformConstCostTable[] = {
{ ISD::SHL, MVT::v64i8, 2 }, // psllw + pand.
{ ISD::SRL, MVT::v64i8, 2 }, // psrlw + pand.
{ ISD::SRA, MVT::v64i8, 4 }, // psrlw, pand, pxor, psubb.
};
if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
ST->hasBWI()) {
if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD,
LT.second))
return LT.first * Entry->Cost;
}
static const CostTblEntry AVX512UniformConstCostTable[] = {
{ ISD::SRA, MVT::v2i64, 1 },
{ ISD::SRA, MVT::v4i64, 1 },
{ ISD::SRA, MVT::v8i64, 1 },
};
if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
ST->hasAVX512()) {
if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD,
LT.second))
return LT.first * Entry->Cost;
}
static const CostTblEntry AVX2UniformConstCostTable[] = {
{ ISD::SHL, MVT::v32i8, 2 }, // psllw + pand.
{ ISD::SRL, MVT::v32i8, 2 }, // psrlw + pand.
{ ISD::SRA, MVT::v32i8, 4 }, // psrlw, pand, pxor, psubb.
{ ISD::SRA, MVT::v4i64, 4 }, // 2 x psrad + shuffle.
};
if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
ST->hasAVX2()) {
if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD,
LT.second))
return LT.first * Entry->Cost;
}
static const CostTblEntry SSE2UniformConstCostTable[] = {
{ ISD::SHL, MVT::v16i8, 2 }, // psllw + pand.
{ ISD::SRL, MVT::v16i8, 2 }, // psrlw + pand.
{ ISD::SRA, MVT::v16i8, 4 }, // psrlw, pand, pxor, psubb.
{ ISD::SHL, MVT::v32i8, 4+2 }, // 2*(psllw + pand) + split.
{ ISD::SRL, MVT::v32i8, 4+2 }, // 2*(psrlw + pand) + split.
{ ISD::SRA, MVT::v32i8, 8+2 }, // 2*(psrlw, pand, pxor, psubb) + split.
};
// XOP has faster vXi8 shifts.
if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
ST->hasSSE2() && !ST->hasXOP()) {
if (const auto *Entry =
CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second))
return LT.first * Entry->Cost;
}
static const CostTblEntry AVX512BWConstCostTable[] = {
{ ISD::SDIV, MVT::v64i8, 14 }, // 2*ext+2*pmulhw sequence
{ ISD::SREM, MVT::v64i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence
{ ISD::UDIV, MVT::v64i8, 14 }, // 2*ext+2*pmulhw sequence
{ ISD::UREM, MVT::v64i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence
{ ISD::SDIV, MVT::v32i16, 6 }, // vpmulhw sequence
{ ISD::SREM, MVT::v32i16, 8 }, // vpmulhw+mul+sub sequence
{ ISD::UDIV, MVT::v32i16, 6 }, // vpmulhuw sequence
{ ISD::UREM, MVT::v32i16, 8 }, // vpmulhuw+mul+sub sequence
};
if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
ST->hasBWI()) {
if (const auto *Entry =
CostTableLookup(AVX512BWConstCostTable, ISD, LT.second))
return LT.first * Entry->Cost;
}
static const CostTblEntry AVX512ConstCostTable[] = {
{ ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence
{ ISD::SREM, MVT::v16i32, 17 }, // vpmuldq+mul+sub sequence
{ ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence
{ ISD::UREM, MVT::v16i32, 17 }, // vpmuludq+mul+sub sequence
};
if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
ST->hasAVX512()) {
if (const auto *Entry =
CostTableLookup(AVX512ConstCostTable, ISD, LT.second))
return LT.first * Entry->Cost;
}
static const CostTblEntry AVX2ConstCostTable[] = {
{ ISD::SDIV, MVT::v32i8, 14 }, // 2*ext+2*pmulhw sequence
{ ISD::SREM, MVT::v32i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence
{ ISD::UDIV, MVT::v32i8, 14 }, // 2*ext+2*pmulhw sequence
{ ISD::UREM, MVT::v32i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence
{ ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence
{ ISD::SREM, MVT::v16i16, 8 }, // vpmulhw+mul+sub sequence
{ ISD::UDIV, MVT::v16i16, 6 }, // vpmulhuw sequence
{ ISD::UREM, MVT::v16i16, 8 }, // vpmulhuw+mul+sub sequence
{ ISD::SDIV, MVT::v8i32, 15 }, // vpmuldq sequence
{ ISD::SREM, MVT::v8i32, 19 }, // vpmuldq+mul+sub sequence
{ ISD::UDIV, MVT::v8i32, 15 }, // vpmuludq sequence
{ ISD::UREM, MVT::v8i32, 19 }, // vpmuludq+mul+sub sequence
};
if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
ST->hasAVX2()) {
if (const auto *Entry = CostTableLookup(AVX2ConstCostTable, ISD, LT.second))
return LT.first * Entry->Cost;
}
static const CostTblEntry SSE2ConstCostTable[] = {
{ ISD::SDIV, MVT::v32i8, 28+2 }, // 4*ext+4*pmulhw sequence + split.
{ ISD::SREM, MVT::v32i8, 32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split.
{ ISD::SDIV, MVT::v16i8, 14 }, // 2*ext+2*pmulhw sequence
{ ISD::SREM, MVT::v16i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence
{ ISD::UDIV, MVT::v32i8, 28+2 }, // 4*ext+4*pmulhw sequence + split.
{ ISD::UREM, MVT::v32i8, 32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split.
{ ISD::UDIV, MVT::v16i8, 14 }, // 2*ext+2*pmulhw sequence
{ ISD::UREM, MVT::v16i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence
{ ISD::SDIV, MVT::v16i16, 12+2 }, // 2*pmulhw sequence + split.
{ ISD::SREM, MVT::v16i16, 16+2 }, // 2*pmulhw+mul+sub sequence + split.
{ ISD::SDIV, MVT::v8i16, 6 }, // pmulhw sequence
{ ISD::SREM, MVT::v8i16, 8 }, // pmulhw+mul+sub sequence
{ ISD::UDIV, MVT::v16i16, 12+2 }, // 2*pmulhuw sequence + split.
{ ISD::UREM, MVT::v16i16, 16+2 }, // 2*pmulhuw+mul+sub sequence + split.
{ ISD::UDIV, MVT::v8i16, 6 }, // pmulhuw sequence
{ ISD::UREM, MVT::v8i16, 8 }, // pmulhuw+mul+sub sequence
{ ISD::SDIV, MVT::v8i32, 38+2 }, // 2*pmuludq sequence + split.
{ ISD::SREM, MVT::v8i32, 48+2 }, // 2*pmuludq+mul+sub sequence + split.
{ ISD::SDIV, MVT::v4i32, 19 }, // pmuludq sequence
{ ISD::SREM, MVT::v4i32, 24 }, // pmuludq+mul+sub sequence
{ ISD::UDIV, MVT::v8i32, 30+2 }, // 2*pmuludq sequence + split.
{ ISD::UREM, MVT::v8i32, 40+2 }, // 2*pmuludq+mul+sub sequence + split.
{ ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence
{ ISD::UREM, MVT::v4i32, 20 }, // pmuludq+mul+sub sequence
};
if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
ST->hasSSE2()) {
// pmuldq sequence.
if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX())
return LT.first * 32;
if (ISD == ISD::SREM && LT.second == MVT::v8i32 && ST->hasAVX())
return LT.first * 38;
if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
return LT.first * 15;
if (ISD == ISD::SREM && LT.second == MVT::v4i32 && ST->hasSSE41())
return LT.first * 20;
if (const auto *Entry = CostTableLookup(SSE2ConstCostTable, ISD, LT.second))
return LT.first * Entry->Cost;
}
static const CostTblEntry AVX2UniformCostTable[] = {
// Uniform splats are cheaper for the following instructions.
{ ISD::SHL, MVT::v16i16, 1 }, // psllw.
{ ISD::SRL, MVT::v16i16, 1 }, // psrlw.
{ ISD::SRA, MVT::v16i16, 1 }, // psraw.
};
if (ST->hasAVX2() &&
((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
(Op2Info == TargetTransformInfo::OK_UniformValue))) {
if (const auto *Entry =
CostTableLookup(AVX2UniformCostTable, ISD, LT.second))
return LT.first * Entry->Cost;
}
static const CostTblEntry SSE2UniformCostTable[] = {
// Uniform splats are cheaper for the following instructions.
{ ISD::SHL, MVT::v8i16, 1 }, // psllw.
{ ISD::SHL, MVT::v4i32, 1 }, // pslld
{ ISD::SHL, MVT::v2i64, 1 }, // psllq.
{ ISD::SRL, MVT::v8i16, 1 }, // psrlw.
{ ISD::SRL, MVT::v4i32, 1 }, // psrld.
{ ISD::SRL, MVT::v2i64, 1 }, // psrlq.
{ ISD::SRA, MVT::v8i16, 1 }, // psraw.
{ ISD::SRA, MVT::v4i32, 1 }, // psrad.
};
if (ST->hasSSE2() &&
((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
(Op2Info == TargetTransformInfo::OK_UniformValue))) {
if (const auto *Entry =
CostTableLookup(SSE2UniformCostTable, ISD, LT.second))
return LT.first * Entry->Cost;
}
static const CostTblEntry AVX512DQCostTable[] = {
{ ISD::MUL, MVT::v2i64, 1 },
{ ISD::MUL, MVT::v4i64, 1 },
{ ISD::MUL, MVT::v8i64, 1 }
};
// Look for AVX512DQ lowering tricks for custom cases.
if (ST->hasDQI())
if (const auto *Entry = CostTableLookup(AVX512DQCostTable, ISD, LT.second))
return LT.first * Entry->Cost;
static const CostTblEntry AVX512BWCostTable[] = {
{ ISD::SHL, MVT::v8i16, 1 }, // vpsllvw
{ ISD::SRL, MVT::v8i16, 1 }, // vpsrlvw
{ ISD::SRA, MVT::v8i16, 1 }, // vpsravw
{ ISD::SHL, MVT::v16i16, 1 }, // vpsllvw
{ ISD::SRL, MVT::v16i16, 1 }, // vpsrlvw
{ ISD::SRA, MVT::v16i16, 1 }, // vpsravw
{ ISD::SHL, MVT::v32i16, 1 }, // vpsllvw
{ ISD::SRL, MVT::v32i16, 1 }, // vpsrlvw
{ ISD::SRA, MVT::v32i16, 1 }, // vpsravw
{ ISD::SHL, MVT::v64i8, 11 }, // vpblendvb sequence.
{ ISD::SRL, MVT::v64i8, 11 }, // vpblendvb sequence.
{ ISD::SRA, MVT::v64i8, 24 }, // vpblendvb sequence.
{ ISD::MUL, MVT::v64i8, 11 }, // extend/pmullw/trunc sequence.
{ ISD::MUL, MVT::v32i8, 4 }, // extend/pmullw/trunc sequence.
{ ISD::MUL, MVT::v16i8, 4 }, // extend/pmullw/trunc sequence.
};
// Look for AVX512BW lowering tricks for custom cases.
if (ST->hasBWI())
if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD, LT.second))
return LT.first * Entry->Cost;
static const CostTblEntry AVX512CostTable[] = {
{ ISD::SHL, MVT::v16i32, 1 },
{ ISD::SRL, MVT::v16i32, 1 },
{ ISD::SRA, MVT::v16i32, 1 },
{ ISD::SHL, MVT::v8i64, 1 },
{ ISD::SRL, MVT::v8i64, 1 },
{ ISD::SRA, MVT::v2i64, 1 },
{ ISD::SRA, MVT::v4i64, 1 },
{ ISD::SRA, MVT::v8i64, 1 },
{ ISD::MUL, MVT::v32i8, 13 }, // extend/pmullw/trunc sequence.
{ ISD::MUL, MVT::v16i8, 5 }, // extend/pmullw/trunc sequence.
{ ISD::MUL, MVT::v16i32, 1 }, // pmulld (Skylake from agner.org)
{ ISD::MUL, MVT::v8i32, 1 }, // pmulld (Skylake from agner.org)
{ ISD::MUL, MVT::v4i32, 1 }, // pmulld (Skylake from agner.org)
{ ISD::MUL, MVT::v8i64, 8 }, // 3*pmuludq/3*shift/2*add
{ ISD::FADD, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/
{ ISD::FSUB, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/
{ ISD::FMUL, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/
{ ISD::FADD, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/
{ ISD::FSUB, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/
{ ISD::FMUL, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/
};
if (ST->hasAVX512())
if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second))
return LT.first * Entry->Cost;
static const CostTblEntry AVX2ShiftCostTable[] = {
// Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
// customize them to detect the cases where shift amount is a scalar one.
{ ISD::SHL, MVT::v4i32, 1 },
{ ISD::SRL, MVT::v4i32, 1 },
{ ISD::SRA, MVT::v4i32, 1 },
{ ISD::SHL, MVT::v8i32, 1 },
{ ISD::SRL, MVT::v8i32, 1 },
{ ISD::SRA, MVT::v8i32, 1 },
{ ISD::SHL, MVT::v2i64, 1 },
{ ISD::SRL, MVT::v2i64, 1 },
{ ISD::SHL, MVT::v4i64, 1 },
{ ISD::SRL, MVT::v4i64, 1 },
};
// Look for AVX2 lowering tricks.
if (ST->hasAVX2()) {
if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
(Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
// On AVX2, a packed v16i16 shift left by a constant build_vector
// is lowered into a vector multiply (vpmullw).
return getArithmeticInstrCost(Instruction::Mul, Ty, Op1Info, Op2Info,
TargetTransformInfo::OP_None,
TargetTransformInfo::OP_None);
if (const auto *Entry = CostTableLookup(AVX2ShiftCostTable, ISD, LT.second))
return LT.first * Entry->Cost;
}
static const CostTblEntry XOPShiftCostTable[] = {
// 128bit shifts take 1cy, but right shifts require negation beforehand.
{ ISD::SHL, MVT::v16i8, 1 },
{ ISD::SRL, MVT::v16i8, 2 },
{ ISD::SRA, MVT::v16i8, 2 },
{ ISD::SHL, MVT::v8i16, 1 },
{ ISD::SRL, MVT::v8i16, 2 },
{ ISD::SRA, MVT::v8i16, 2 },
{ ISD::SHL, MVT::v4i32, 1 },
{ ISD::SRL, MVT::v4i32, 2 },
{ ISD::SRA, MVT::v4i32, 2 },
{ ISD::SHL, MVT::v2i64, 1 },
{ ISD::SRL, MVT::v2i64, 2 },
{ ISD::SRA, MVT::v2i64, 2 },
// 256bit shifts require splitting if AVX2 didn't catch them above.
{ ISD::SHL, MVT::v32i8, 2+2 },
{ ISD::SRL, MVT::v32i8, 4+2 },
{ ISD::SRA, MVT::v32i8, 4+2 },
{ ISD::SHL, MVT::v16i16, 2+2 },
{ ISD::SRL, MVT::v16i16, 4+2 },
{ ISD::SRA, MVT::v16i16, 4+2 },
{ ISD::SHL, MVT::v8i32, 2+2 },
{ ISD::SRL, MVT::v8i32, 4+2 },
{ ISD::SRA, MVT::v8i32, 4+2 },
{ ISD::SHL, MVT::v4i64, 2+2 },
{ ISD::SRL, MVT::v4i64, 4+2 },
{ ISD::SRA, MVT::v4i64, 4+2 },
};
// Look for XOP lowering tricks.
if (ST->hasXOP())
if (const auto *Entry = CostTableLookup(XOPShiftCostTable, ISD, LT.second))
return LT.first * Entry->Cost;
static const CostTblEntry SSE2UniformShiftCostTable[] = {
// Uniform splats are cheaper for the following instructions.
{ ISD::SHL, MVT::v16i16, 2+2 }, // 2*psllw + split.
{ ISD::SHL, MVT::v8i32, 2+2 }, // 2*pslld + split.
{ ISD::SHL, MVT::v4i64, 2+2 }, // 2*psllq + split.
{ ISD::SRL, MVT::v16i16, 2+2 }, // 2*psrlw + split.
{ ISD::SRL, MVT::v8i32, 2+2 }, // 2*psrld + split.
{ ISD::SRL, MVT::v4i64, 2+2 }, // 2*psrlq + split.
{ ISD::SRA, MVT::v16i16, 2+2 }, // 2*psraw + split.
{ ISD::SRA, MVT::v8i32, 2+2 }, // 2*psrad + split.
{ ISD::SRA, MVT::v2i64, 4 }, // 2*psrad + shuffle.
{ ISD::SRA, MVT::v4i64, 8+2 }, // 2*(2*psrad + shuffle) + split.
};
if (ST->hasSSE2() &&
((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
(Op2Info == TargetTransformInfo::OK_UniformValue))) {
// Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table.
if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2())
return LT.first * 4; // 2*psrad + shuffle.
if (const auto *Entry =
CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second))
return LT.first * Entry->Cost;
}
if (ISD == ISD::SHL &&
Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
MVT VT = LT.second;
// Vector shift left by non uniform constant can be lowered
// into vector multiply.
if (((VT == MVT::v8i16 || VT == MVT::v4i32) && ST->hasSSE2()) ||
((VT == MVT::v16i16 || VT == MVT::v8i32) && ST->hasAVX()))
ISD = ISD::MUL;
}
static const CostTblEntry AVX2CostTable[] = {
{ ISD::SHL, MVT::v32i8, 11 }, // vpblendvb sequence.
{ ISD::SHL, MVT::v16i16, 10 }, // extend/vpsrlvd/pack sequence.
{ ISD::SRL, MVT::v32i8, 11 }, // vpblendvb sequence.
{ ISD::SRL, MVT::v16i16, 10 }, // extend/vpsrlvd/pack sequence.
{ ISD::SRA, MVT::v32i8, 24 }, // vpblendvb sequence.
{ ISD::SRA, MVT::v16i16, 10 }, // extend/vpsravd/pack sequence.
{ ISD::SRA, MVT::v2i64, 4 }, // srl/xor/sub sequence.
{ ISD::SRA, MVT::v4i64, 4 }, // srl/xor/sub sequence.
{ ISD::SUB, MVT::v32i8, 1 }, // psubb
{ ISD::ADD, MVT::v32i8, 1 }, // paddb
{ ISD::SUB, MVT::v16i16, 1 }, // psubw
{ ISD::ADD, MVT::v16i16, 1 }, // paddw
{ ISD::SUB, MVT::v8i32, 1 }, // psubd
{ ISD::ADD, MVT::v8i32, 1 }, // paddd
{ ISD::SUB, MVT::v4i64, 1 }, // psubq
{ ISD::ADD, MVT::v4i64, 1 }, // paddq
{ ISD::MUL, MVT::v32i8, 17 }, // extend/pmullw/trunc sequence.
{ ISD::MUL, MVT::v16i8, 7 }, // extend/pmullw/trunc sequence.
{ ISD::MUL, MVT::v16i16, 1 }, // pmullw
{ ISD::MUL, MVT::v8i32, 2 }, // pmulld (Haswell from agner.org)
{ ISD::MUL, MVT::v4i64, 8 }, // 3*pmuludq/3*shift/2*add
{ ISD::FADD, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/
{ ISD::FADD, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/
{ ISD::FSUB, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/
{ ISD::FSUB, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/
{ ISD::FMUL, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/
{ ISD::FMUL, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/
{ ISD::FDIV, MVT::f32, 7 }, // Haswell from http://www.agner.org/
{ ISD::FDIV, MVT::v4f32, 7 }, // Haswell from http://www.agner.org/
{ ISD::FDIV, MVT::v8f32, 14 }, // Haswell from http://www.agner.org/
{ ISD::FDIV, MVT::f64, 14 }, // Haswell from http://www.agner.org/
{ ISD::FDIV, MVT::v2f64, 14 }, // Haswell from http://www.agner.org/
{ ISD::FDIV, MVT::v4f64, 28 }, // Haswell from http://www.agner.org/
};
// Look for AVX2 lowering tricks for custom cases.
if (ST->hasAVX2())
if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second))
return LT.first * Entry->Cost;
static const CostTblEntry AVX1CostTable[] = {
// We don't have to scalarize unsupported ops. We can issue two half-sized
// operations and we only need to extract the upper YMM half.
// Two ops + 1 extract + 1 insert = 4.
{ ISD::MUL, MVT::v16i16, 4 },
{ ISD::MUL, MVT::v8i32, 4 },
{ ISD::SUB, MVT::v32i8, 4 },
{ ISD::ADD, MVT::v32i8, 4 },
{ ISD::SUB, MVT::v16i16, 4 },
{ ISD::ADD, MVT::v16i16, 4 },
{ ISD::SUB, MVT::v8i32, 4 },
{ ISD::ADD, MVT::v8i32, 4 },
{ ISD::SUB, MVT::v4i64, 4 },
{ ISD::ADD, MVT::v4i64, 4 },
// A v4i64 multiply is custom lowered as two split v2i64 vectors that then
// are lowered as a series of long multiplies(3), shifts(3) and adds(2)
// Because we believe v4i64 to be a legal type, we must also include the
// extract+insert in the cost table. Therefore, the cost here is 18
// instead of 8.
{ ISD::MUL, MVT::v4i64, 18 },
{ ISD::MUL, MVT::v32i8, 26 }, // extend/pmullw/trunc sequence.
{ ISD::FDIV, MVT::f32, 14 }, // SNB from http://www.agner.org/
{ ISD::FDIV, MVT::v4f32, 14 }, // SNB from http://www.agner.org/
{ ISD::FDIV, MVT::v8f32, 28 }, // SNB from http://www.agner.org/
{ ISD::FDIV, MVT::f64, 22 }, // SNB from http://www.agner.org/
{ ISD::FDIV, MVT::v2f64, 22 }, // SNB from http://www.agner.org/
{ ISD::FDIV, MVT::v4f64, 44 }, // SNB from http://www.agner.org/
};
if (ST->hasAVX())
if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, LT.second))
return LT.first * Entry->Cost;
static const CostTblEntry SSE42CostTable[] = {
{ ISD::FADD, MVT::f64, 1 }, // Nehalem from http://www.agner.org/
{ ISD::FADD, MVT::f32, 1 }, // Nehalem from http://www.agner.org/
{ ISD::FADD, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/
{ ISD::FADD, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/
{ ISD::FSUB, MVT::f64, 1 }, // Nehalem from http://www.agner.org/
{ ISD::FSUB, MVT::f32 , 1 }, // Nehalem from http://www.agner.org/
{ ISD::FSUB, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/
{ ISD::FSUB, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/
{ ISD::FMUL, MVT::f64, 1 }, // Nehalem from http://www.agner.org/
{ ISD::FMUL, MVT::f32, 1 }, // Nehalem from http://www.agner.org/
{ ISD::FMUL, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/
{ ISD::FMUL, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/
{ ISD::FDIV, MVT::f32, 14 }, // Nehalem from http://www.agner.org/
{ ISD::FDIV, MVT::v4f32, 14 }, // Nehalem from http://www.agner.org/
{ ISD::FDIV, MVT::f64, 22 }, // Nehalem from http://www.agner.org/
{ ISD::FDIV, MVT::v2f64, 22 }, // Nehalem from http://www.agner.org/
};
if (ST->hasSSE42())
if (const auto *Entry = CostTableLookup(SSE42CostTable, ISD, LT.second))
return LT.first * Entry->Cost;
static const CostTblEntry SSE41CostTable[] = {
{ ISD::SHL, MVT::v16i8, 11 }, // pblendvb sequence.
{ ISD::SHL, MVT::v32i8, 2*11+2 }, // pblendvb sequence + split.
{ ISD::SHL, MVT::v8i16, 14 }, // pblendvb sequence.
{ ISD::SHL, MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
{ ISD::SHL, MVT::v4i32, 4 }, // pslld/paddd/cvttps2dq/pmulld
{ ISD::SHL, MVT::v8i32, 2*4+2 }, // pslld/paddd/cvttps2dq/pmulld + split
{ ISD::SRL, MVT::v16i8, 12 }, // pblendvb sequence.
{ ISD::SRL, MVT::v32i8, 2*12+2 }, // pblendvb sequence + split.
{ ISD::SRL, MVT::v8i16, 14 }, // pblendvb sequence.
{ ISD::SRL, MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
{ ISD::SRL, MVT::v4i32, 11 }, // Shift each lane + blend.
{ ISD::SRL, MVT::v8i32, 2*11+2 }, // Shift each lane + blend + split.
{ ISD::SRA, MVT::v16i8, 24 }, // pblendvb sequence.
{ ISD::SRA, MVT::v32i8, 2*24+2 }, // pblendvb sequence + split.
{ ISD::SRA, MVT::v8i16, 14 }, // pblendvb sequence.
{ ISD::SRA, MVT::v16i16, 2*14+2 }, // pblendvb sequence + split.
{ ISD::SRA, MVT::v4i32, 12 }, // Shift each lane + blend.
{ ISD::SRA, MVT::v8i32, 2*12+2 }, // Shift each lane + blend + split.
{ ISD::MUL, MVT::v4i32, 2 } // pmulld (Nehalem from agner.org)
};
if (ST->hasSSE41())
if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second))
return LT.first * Entry->Cost;
static const CostTblEntry SSE2CostTable[] = {
// We don't correctly identify costs of casts because they are marked as
// custom.
{ ISD::SHL, MVT::v16i8, 26 }, // cmpgtb sequence.
{ ISD::SHL, MVT::v8i16, 32 }, // cmpgtb sequence.
{ ISD::SHL, MVT::v4i32, 2*5 }, // We optimized this using mul.
{ ISD::SHL, MVT::v2i64, 4 }, // splat+shuffle sequence.
{ ISD::SHL, MVT::v4i64, 2*4+2 }, // splat+shuffle sequence + split.
{ ISD::SRL, MVT::v16i8, 26 }, // cmpgtb sequence.
{ ISD::SRL, MVT::v8i16, 32 }, // cmpgtb sequence.
{ ISD::SRL, MVT::v4i32, 16 }, // Shift each lane + blend.
{ ISD::SRL, MVT::v2i64, 4 }, // splat+shuffle sequence.
{ ISD::SRL, MVT::v4i64, 2*4+2 }, // splat+shuffle sequence + split.
{ ISD::SRA, MVT::v16i8, 54 }, // unpacked cmpgtb sequence.
{ ISD::SRA, MVT::v8i16, 32 }, // cmpgtb sequence.
{ ISD::SRA, MVT::v4i32, 16 }, // Shift each lane + blend.
{ ISD::SRA, MVT::v2i64, 12 }, // srl/xor/sub sequence.
{ ISD::SRA, MVT::v4i64, 2*12+2 }, // srl/xor/sub sequence+split.
{ ISD::MUL, MVT::v16i8, 12 }, // extend/pmullw/trunc sequence.
{ ISD::MUL, MVT::v8i16, 1 }, // pmullw
{ ISD::MUL, MVT::v4i32, 6 }, // 3*pmuludq/4*shuffle
{ ISD::MUL, MVT::v2i64, 8 }, // 3*pmuludq/3*shift/2*add
{ ISD::FDIV, MVT::f32, 23 }, // Pentium IV from http://www.agner.org/
{ ISD::FDIV, MVT::v4f32, 39 }, // Pentium IV from http://www.agner.org/
{ ISD::FDIV, MVT::f64, 38 }, // Pentium IV from http://www.agner.org/
{ ISD::FDIV, MVT::v2f64, 69 }, // Pentium IV from http://www.agner.org/
};
if (ST->hasSSE2())
if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second))
return LT.first * Entry->Cost;
static const CostTblEntry SSE1CostTable[] = {
{ ISD::FDIV, MVT::f32, 17 }, // Pentium III from http://www.agner.org/
{ ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/
};
if (ST->hasSSE1())
if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second))
return LT.first * Entry->Cost;
// It is not a good idea to vectorize division. We have to scalarize it and
// in the process we will often end up having to spilling regular
// registers. The overhead of division is going to dominate most kernels
// anyways so try hard to prevent vectorization of division - it is
// generally a bad idea. Assume somewhat arbitrarily that we have to be able
// to hide "20 cycles" for each lane.
if (LT.second.isVector() && (ISD == ISD::SDIV || ISD == ISD::SREM ||
ISD == ISD::UDIV || ISD == ISD::UREM)) {
int ScalarCost = getArithmeticInstrCost(
Opcode, Ty->getScalarType(), Op1Info, Op2Info,
TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
return 20 * LT.first * LT.second.getVectorNumElements() * ScalarCost;
}
// Fallback to the default implementation.
return BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
}
int X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
Type *SubTp) {
// 64-bit packed float vectors (v2f32) are widened to type v4f32.
// 64-bit packed integer vectors (v2i32) are promoted to type v2i64.
std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
// Treat Transpose as 2-op shuffles - there's no difference in lowering.
if (Kind == TTI::SK_Transpose)
Kind = TTI::SK_PermuteTwoSrc;
// For Broadcasts we are splatting the first element from the first input
// register, so only need to reference that input and all the output
// registers are the same.
if (Kind == TTI::SK_Broadcast)
LT.first = 1;
// Subvector extractions are free if they start at beginning of the
// vector.
if (Kind == TTI::SK_ExtractSubvector && LT.second.isVector() &&
((Index % LT.second.getVectorNumElements()) == 0))
return 0;
// We are going to permute multiple sources and the result will be in multiple
// destinations. Providing an accurate cost only for splits where the element
// type remains the same.
if (Kind == TTI::SK_PermuteSingleSrc && LT.first != 1) {
MVT LegalVT = LT.second;
if (LegalVT.isVector() &&
LegalVT.getVectorElementType().getSizeInBits() ==
Tp->getVectorElementType()->getPrimitiveSizeInBits() &&
LegalVT.getVectorNumElements() < Tp->getVectorNumElements()) {
unsigned VecTySize = DL.getTypeStoreSize(Tp);
unsigned LegalVTSize = LegalVT.getStoreSize();
// Number of source vectors after legalization:
unsigned NumOfSrcs = (VecTySize + LegalVTSize - 1) / LegalVTSize;
// Number of destination vectors after legalization:
unsigned NumOfDests = LT.first;
Type *SingleOpTy = VectorType::get(Tp->getVectorElementType(),
LegalVT.getVectorNumElements());
unsigned NumOfShuffles = (NumOfSrcs - 1) * NumOfDests;
return NumOfShuffles *
getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy, 0, nullptr);
}
return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
}
// For 2-input shuffles, we must account for splitting the 2 inputs into many.
if (Kind == TTI::SK_PermuteTwoSrc && LT.first != 1) {
// We assume that source and destination have the same vector type.
int NumOfDests = LT.first;
int NumOfShufflesPerDest = LT.first * 2 - 1;
LT.first = NumOfDests * NumOfShufflesPerDest;
}
static const CostTblEntry AVX512VBMIShuffleTbl[] = {
{TTI::SK_Reverse, MVT::v64i8, 1}, // vpermb
{TTI::SK_Reverse, MVT::v32i8, 1}, // vpermb
{TTI::SK_PermuteSingleSrc, MVT::v64i8, 1}, // vpermb
{TTI::SK_PermuteSingleSrc, MVT::v32i8, 1}, // vpermb
{TTI::SK_PermuteTwoSrc, MVT::v64i8, 1}, // vpermt2b
{TTI::SK_PermuteTwoSrc, MVT::v32i8, 1}, // vpermt2b
{TTI::SK_PermuteTwoSrc, MVT::v16i8, 1} // vpermt2b
};
if (ST->hasVBMI())
if (const auto *Entry =
CostTableLookup(AVX512VBMIShuffleTbl, Kind, LT.second))
return LT.first * Entry->Cost;
static const CostTblEntry AVX512BWShuffleTbl[] = {
{TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw
{TTI::SK_Broadcast, MVT::v64i8, 1}, // vpbroadcastb
{TTI::SK_Reverse, MVT::v32i16, 1}, // vpermw
{TTI::SK_Reverse, MVT::v16i16, 1}, // vpermw
{TTI::SK_Reverse, MVT::v64i8, 2}, // pshufb + vshufi64x2
{TTI::SK_PermuteSingleSrc, MVT::v32i16, 1}, // vpermw
{TTI::SK_PermuteSingleSrc, MVT::v16i16, 1}, // vpermw
{TTI::SK_PermuteSingleSrc, MVT::v8i16, 1}, // vpermw
{TTI::SK_PermuteSingleSrc, MVT::v64i8, 8}, // extend to v32i16
{TTI::SK_PermuteSingleSrc, MVT::v32i8, 3}, // vpermw + zext/trunc
{TTI::SK_PermuteTwoSrc, MVT::v32i16, 1}, // vpermt2w
{TTI::SK_PermuteTwoSrc, MVT::v16i16, 1}, // vpermt2w
{TTI::SK_PermuteTwoSrc, MVT::v8i16, 1}, // vpermt2w
{TTI::SK_PermuteTwoSrc, MVT::v32i8, 3}, // zext + vpermt2w + trunc
{TTI::SK_PermuteTwoSrc, MVT::v64i8, 19}, // 6 * v32i8 + 1
{TTI::SK_PermuteTwoSrc, MVT::v16i8, 3} // zext + vpermt2w + trunc
};
if (ST->hasBWI())
if (const auto *Entry =
CostTableLookup(AVX512BWShuffleTbl, Kind, LT.second))
return LT.first * Entry->Cost;
static const CostTblEntry AVX512ShuffleTbl[] = {
{TTI::SK_Broadcast, MVT::v8f64, 1}, // vbroadcastpd
{TTI::SK_Broadcast, MVT::v16f32, 1}, // vbroadcastps
{TTI::SK_Broadcast, MVT::v8i64, 1}, // vpbroadcastq
{TTI::SK_Broadcast, MVT::v16i32, 1}, // vpbroadcastd
{TTI::SK_Reverse, MVT::v8f64, 1}, // vpermpd
{TTI::SK_Reverse, MVT::v16f32, 1}, // vpermps
{TTI::SK_Reverse, MVT::v8i64, 1}, // vpermq
{TTI::SK_Reverse, MVT::v16i32, 1}, // vpermd
{TTI::SK_PermuteSingleSrc, MVT::v8f64, 1}, // vpermpd
{TTI::SK_PermuteSingleSrc, MVT::v4f64, 1}, // vpermpd
{TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // vpermpd
{TTI::SK_PermuteSingleSrc, MVT::v16f32, 1}, // vpermps
{TTI::SK_PermuteSingleSrc, MVT::v8f32, 1}, // vpermps
{TTI::SK_PermuteSingleSrc, MVT::v4f32, 1}, // vpermps
{TTI::SK_PermuteSingleSrc, MVT::v8i64, 1}, // vpermq
{TTI::SK_PermuteSingleSrc, MVT::v4i64, 1}, // vpermq
{TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // vpermq
{TTI::SK_PermuteSingleSrc, MVT::v16i32, 1}, // vpermd
{TTI::SK_PermuteSingleSrc, MVT::v8i32, 1}, // vpermd
{TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // vpermd
{TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb
{TTI::SK_PermuteTwoSrc, MVT::v8f64, 1}, // vpermt2pd
{TTI::SK_PermuteTwoSrc, MVT::v16f32, 1}, // vpermt2ps
{TTI::SK_PermuteTwoSrc, MVT::v8i64, 1}, // vpermt2q
{TTI::SK_PermuteTwoSrc, MVT::v16i32, 1}, // vpermt2d
{TTI::SK_PermuteTwoSrc, MVT::v4f64, 1}, // vpermt2pd
{TTI::SK_PermuteTwoSrc, MVT::v8f32, 1}, // vpermt2ps
{TTI::SK_PermuteTwoSrc, MVT::v4i64, 1}, // vpermt2q
{TTI::SK_PermuteTwoSrc, MVT::v8i32, 1}, // vpermt2d
{TTI::SK_PermuteTwoSrc, MVT::v2f64, 1}, // vpermt2pd
{TTI::SK_PermuteTwoSrc, MVT::v4f32, 1}, // vpermt2ps
{TTI::SK_PermuteTwoSrc, MVT::v2i64, 1}, // vpermt2q
{TTI::SK_PermuteTwoSrc, MVT::v4i32, 1} // vpermt2d
};
if (ST->hasAVX512())