forked from llvm/llvm-project
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathSystemZTargetTransformInfo.cpp
1207 lines (1070 loc) · 45 KB
/
SystemZTargetTransformInfo.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
//===-- SystemZTargetTransformInfo.cpp - SystemZ-specific TTI -------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements a TargetTransformInfo analysis pass specific to the
// SystemZ 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.
//
//===----------------------------------------------------------------------===//
#include "SystemZTargetTransformInfo.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 "systemztti"
//===----------------------------------------------------------------------===//
//
// SystemZ cost model.
//
//===----------------------------------------------------------------------===//
int SystemZTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
TTI::TargetCostKind CostKind) {
assert(Ty->isIntegerTy());
unsigned BitSize = Ty->getPrimitiveSizeInBits();
// There is no cost model for constants with a bit size of 0. Return TCC_Free
// here, so that constant hoisting will ignore this constant.
if (BitSize == 0)
return TTI::TCC_Free;
// No cost model for operations on integers larger than 64 bit implemented yet.
if (BitSize > 64)
return TTI::TCC_Free;
if (Imm == 0)
return TTI::TCC_Free;
if (Imm.getBitWidth() <= 64) {
// Constants loaded via lgfi.
if (isInt<32>(Imm.getSExtValue()))
return TTI::TCC_Basic;
// Constants loaded via llilf.
if (isUInt<32>(Imm.getZExtValue()))
return TTI::TCC_Basic;
// Constants loaded via llihf:
if ((Imm.getZExtValue() & 0xffffffff) == 0)
return TTI::TCC_Basic;
return 2 * TTI::TCC_Basic;
}
return 4 * TTI::TCC_Basic;
}
int SystemZTTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
const APInt &Imm, Type *Ty,
TTI::TargetCostKind CostKind,
Instruction *Inst) {
assert(Ty->isIntegerTy());
unsigned BitSize = Ty->getPrimitiveSizeInBits();
// There is no cost model for constants with a bit size of 0. Return TCC_Free
// here, so that constant hoisting will ignore this constant.
if (BitSize == 0)
return TTI::TCC_Free;
// No cost model for operations on integers larger than 64 bit implemented yet.
if (BitSize > 64)
return TTI::TCC_Free;
switch (Opcode) {
default:
return TTI::TCC_Free;
case Instruction::GetElementPtr:
// Always hoist the base address of a GetElementPtr. This prevents the
// creation of new constants for every base constant that gets constant
// folded with the offset.
if (Idx == 0)
return 2 * TTI::TCC_Basic;
return TTI::TCC_Free;
case Instruction::Store:
if (Idx == 0 && Imm.getBitWidth() <= 64) {
// Any 8-bit immediate store can by implemented via mvi.
if (BitSize == 8)
return TTI::TCC_Free;
// 16-bit immediate values can be stored via mvhhi/mvhi/mvghi.
if (isInt<16>(Imm.getSExtValue()))
return TTI::TCC_Free;
}
break;
case Instruction::ICmp:
if (Idx == 1 && Imm.getBitWidth() <= 64) {
// Comparisons against signed 32-bit immediates implemented via cgfi.
if (isInt<32>(Imm.getSExtValue()))
return TTI::TCC_Free;
// Comparisons against unsigned 32-bit immediates implemented via clgfi.
if (isUInt<32>(Imm.getZExtValue()))
return TTI::TCC_Free;
}
break;
case Instruction::Add:
case Instruction::Sub:
if (Idx == 1 && Imm.getBitWidth() <= 64) {
// We use algfi/slgfi to add/subtract 32-bit unsigned immediates.
if (isUInt<32>(Imm.getZExtValue()))
return TTI::TCC_Free;
// Or their negation, by swapping addition vs. subtraction.
if (isUInt<32>(-Imm.getSExtValue()))
return TTI::TCC_Free;
}
break;
case Instruction::Mul:
if (Idx == 1 && Imm.getBitWidth() <= 64) {
// We use msgfi to multiply by 32-bit signed immediates.
if (isInt<32>(Imm.getSExtValue()))
return TTI::TCC_Free;
}
break;
case Instruction::Or:
case Instruction::Xor:
if (Idx == 1 && Imm.getBitWidth() <= 64) {
// Masks supported by oilf/xilf.
if (isUInt<32>(Imm.getZExtValue()))
return TTI::TCC_Free;
// Masks supported by oihf/xihf.
if ((Imm.getZExtValue() & 0xffffffff) == 0)
return TTI::TCC_Free;
}
break;
case Instruction::And:
if (Idx == 1 && Imm.getBitWidth() <= 64) {
// Any 32-bit AND operation can by implemented via nilf.
if (BitSize <= 32)
return TTI::TCC_Free;
// 64-bit masks supported by nilf.
if (isUInt<32>(~Imm.getZExtValue()))
return TTI::TCC_Free;
// 64-bit masks supported by nilh.
if ((Imm.getZExtValue() & 0xffffffff) == 0xffffffff)
return TTI::TCC_Free;
// Some 64-bit AND operations can be implemented via risbg.
const SystemZInstrInfo *TII = ST->getInstrInfo();
unsigned Start, End;
if (TII->isRxSBGMask(Imm.getZExtValue(), BitSize, Start, End))
return TTI::TCC_Free;
}
break;
case Instruction::Shl:
case Instruction::LShr:
case Instruction::AShr:
// Always return TCC_Free for the shift value of a shift instruction.
if (Idx == 1)
return TTI::TCC_Free;
break;
case Instruction::UDiv:
case Instruction::SDiv:
case Instruction::URem:
case Instruction::SRem:
case Instruction::Trunc:
case Instruction::ZExt:
case Instruction::SExt:
case Instruction::IntToPtr:
case Instruction::PtrToInt:
case Instruction::BitCast:
case Instruction::PHI:
case Instruction::Call:
case Instruction::Select:
case Instruction::Ret:
case Instruction::Load:
break;
}
return SystemZTTIImpl::getIntImmCost(Imm, Ty, CostKind);
}
int SystemZTTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
const APInt &Imm, Type *Ty,
TTI::TargetCostKind CostKind) {
assert(Ty->isIntegerTy());
unsigned BitSize = Ty->getPrimitiveSizeInBits();
// There is no cost model for constants with a bit size of 0. Return TCC_Free
// here, so that constant hoisting will ignore this constant.
if (BitSize == 0)
return TTI::TCC_Free;
// No cost model for operations on integers larger than 64 bit implemented yet.
if (BitSize > 64)
return TTI::TCC_Free;
switch (IID) {
default:
return TTI::TCC_Free;
case Intrinsic::sadd_with_overflow:
case Intrinsic::uadd_with_overflow:
case Intrinsic::ssub_with_overflow:
case Intrinsic::usub_with_overflow:
// These get expanded to include a normal addition/subtraction.
if (Idx == 1 && Imm.getBitWidth() <= 64) {
if (isUInt<32>(Imm.getZExtValue()))
return TTI::TCC_Free;
if (isUInt<32>(-Imm.getSExtValue()))
return TTI::TCC_Free;
}
break;
case Intrinsic::smul_with_overflow:
case Intrinsic::umul_with_overflow:
// These get expanded to include a normal multiplication.
if (Idx == 1 && Imm.getBitWidth() <= 64) {
if (isInt<32>(Imm.getSExtValue()))
return TTI::TCC_Free;
}
break;
case Intrinsic::experimental_stackmap:
if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
return TTI::TCC_Free;
break;
case Intrinsic::experimental_patchpoint_void:
case Intrinsic::experimental_patchpoint_i64:
if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
return TTI::TCC_Free;
break;
}
return SystemZTTIImpl::getIntImmCost(Imm, Ty, CostKind);
}
TargetTransformInfo::PopcntSupportKind
SystemZTTIImpl::getPopcntSupport(unsigned TyWidth) {
assert(isPowerOf2_32(TyWidth) && "Type width must be power of 2");
if (ST->hasPopulationCount() && TyWidth <= 64)
return TTI::PSK_FastHardware;
return TTI::PSK_Software;
}
void SystemZTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
TTI::UnrollingPreferences &UP) {
// Find out if L contains a call, what the machine instruction count
// estimate is, and how many stores there are.
bool HasCall = false;
InstructionCost NumStores = 0;
for (auto &BB : L->blocks())
for (auto &I : *BB) {
if (isa<CallInst>(&I) || isa<InvokeInst>(&I)) {
if (const Function *F = cast<CallBase>(I).getCalledFunction()) {
if (isLoweredToCall(F))
HasCall = true;
if (F->getIntrinsicID() == Intrinsic::memcpy ||
F->getIntrinsicID() == Intrinsic::memset)
NumStores++;
} else { // indirect call.
HasCall = true;
}
}
if (isa<StoreInst>(&I)) {
Type *MemAccessTy = I.getOperand(0)->getType();
NumStores += getMemoryOpCost(Instruction::Store, MemAccessTy, None, 0,
TTI::TCK_RecipThroughput);
}
}
// The z13 processor will run out of store tags if too many stores
// are fed into it too quickly. Therefore make sure there are not
// too many stores in the resulting unrolled loop.
unsigned const NumStoresVal = *NumStores.getValue();
unsigned const Max = (NumStoresVal ? (12 / NumStoresVal) : UINT_MAX);
if (HasCall) {
// Only allow full unrolling if loop has any calls.
UP.FullUnrollMaxCount = Max;
UP.MaxCount = 1;
return;
}
UP.MaxCount = Max;
if (UP.MaxCount <= 1)
return;
// Allow partial and runtime trip count unrolling.
UP.Partial = UP.Runtime = true;
UP.PartialThreshold = 75;
UP.DefaultUnrollRuntimeCount = 4;
// Allow expensive instructions in the pre-header of the loop.
UP.AllowExpensiveTripCount = true;
UP.Force = true;
}
void SystemZTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE,
TTI::PeelingPreferences &PP) {
BaseT::getPeelingPreferences(L, SE, PP);
}
bool SystemZTTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1,
TargetTransformInfo::LSRCost &C2) {
// SystemZ specific: check instruction count (first), and don't care about
// ImmCost, since offsets are checked explicitly.
return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost,
C1.NumIVMuls, C1.NumBaseAdds,
C1.ScaleCost, C1.SetupCost) <
std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost,
C2.NumIVMuls, C2.NumBaseAdds,
C2.ScaleCost, C2.SetupCost);
}
unsigned SystemZTTIImpl::getNumberOfRegisters(unsigned ClassID) const {
bool Vector = (ClassID == 1);
if (!Vector)
// Discount the stack pointer. Also leave out %r0, since it can't
// be used in an address.
return 14;
if (ST->hasVector())
return 32;
return 0;
}
TypeSize
SystemZTTIImpl::getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const {
switch (K) {
case TargetTransformInfo::RGK_Scalar:
return TypeSize::getFixed(64);
case TargetTransformInfo::RGK_FixedWidthVector:
return TypeSize::getFixed(ST->hasVector() ? 128 : 0);
case TargetTransformInfo::RGK_ScalableVector:
return TypeSize::getScalable(0);
}
llvm_unreachable("Unsupported register kind");
}
unsigned SystemZTTIImpl::getMinPrefetchStride(unsigned NumMemAccesses,
unsigned NumStridedMemAccesses,
unsigned NumPrefetches,
bool HasCall) const {
// Don't prefetch a loop with many far apart accesses.
if (NumPrefetches > 16)
return UINT_MAX;
// Emit prefetch instructions for smaller strides in cases where we think
// the hardware prefetcher might not be able to keep up.
if (NumStridedMemAccesses > 32 && !HasCall &&
(NumMemAccesses - NumStridedMemAccesses) * 32 <= NumStridedMemAccesses)
return 1;
return ST->hasMiscellaneousExtensions3() ? 8192 : 2048;
}
bool SystemZTTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) {
EVT VT = TLI->getValueType(DL, DataType);
return (VT.isScalarInteger() && TLI->isTypeLegal(VT));
}
// Return the bit size for the scalar type or vector element
// type. getScalarSizeInBits() returns 0 for a pointer type.
static unsigned getScalarSizeInBits(Type *Ty) {
unsigned Size =
(Ty->isPtrOrPtrVectorTy() ? 64U : Ty->getScalarSizeInBits());
assert(Size > 0 && "Element must have non-zero size.");
return Size;
}
// getNumberOfParts() calls getTypeLegalizationCost() which splits the vector
// type until it is legal. This would e.g. return 4 for <6 x i64>, instead of
// 3.
static unsigned getNumVectorRegs(Type *Ty) {
auto *VTy = cast<FixedVectorType>(Ty);
unsigned WideBits = getScalarSizeInBits(Ty) * VTy->getNumElements();
assert(WideBits > 0 && "Could not compute size of vector");
return ((WideBits % 128U) ? ((WideBits / 128U) + 1) : (WideBits / 128U));
}
InstructionCost SystemZTTIImpl::getArithmeticInstrCost(
unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
TTI::OperandValueKind Op1Info, TTI::OperandValueKind Op2Info,
TTI::OperandValueProperties Opd1PropInfo,
TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args,
const Instruction *CxtI) {
// TODO: Handle more cost kinds.
if (CostKind != TTI::TCK_RecipThroughput)
return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
Op2Info, Opd1PropInfo,
Opd2PropInfo, Args, CxtI);
// TODO: return a good value for BB-VECTORIZER that includes the
// immediate loads, which we do not want to count for the loop
// vectorizer, since they are hopefully hoisted out of the loop. This
// would require a new parameter 'InLoop', but not sure if constant
// args are common enough to motivate this.
unsigned ScalarBits = Ty->getScalarSizeInBits();
// There are thre cases of division and remainder: Dividing with a register
// needs a divide instruction. A divisor which is a power of two constant
// can be implemented with a sequence of shifts. Any other constant needs a
// multiply and shifts.
const unsigned DivInstrCost = 20;
const unsigned DivMulSeqCost = 10;
const unsigned SDivPow2Cost = 4;
bool SignedDivRem =
Opcode == Instruction::SDiv || Opcode == Instruction::SRem;
bool UnsignedDivRem =
Opcode == Instruction::UDiv || Opcode == Instruction::URem;
// Check for a constant divisor.
bool DivRemConst = false;
bool DivRemConstPow2 = false;
if ((SignedDivRem || UnsignedDivRem) && Args.size() == 2) {
if (const Constant *C = dyn_cast<Constant>(Args[1])) {
const ConstantInt *CVal =
(C->getType()->isVectorTy()
? dyn_cast_or_null<const ConstantInt>(C->getSplatValue())
: dyn_cast<const ConstantInt>(C));
if (CVal != nullptr &&
(CVal->getValue().isPowerOf2() || (-CVal->getValue()).isPowerOf2()))
DivRemConstPow2 = true;
else
DivRemConst = true;
}
}
if (!Ty->isVectorTy()) {
// These FP operations are supported with a dedicated instruction for
// float, double and fp128 (base implementation assumes float generally
// costs 2).
if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub ||
Opcode == Instruction::FMul || Opcode == Instruction::FDiv)
return 1;
// There is no native support for FRem.
if (Opcode == Instruction::FRem)
return LIBCALL_COST;
// Give discount for some combined logical operations if supported.
if (Args.size() == 2 && ST->hasMiscellaneousExtensions3()) {
if (Opcode == Instruction::Xor) {
for (const Value *A : Args) {
if (const Instruction *I = dyn_cast<Instruction>(A))
if (I->hasOneUse() &&
(I->getOpcode() == Instruction::And ||
I->getOpcode() == Instruction::Or ||
I->getOpcode() == Instruction::Xor))
return 0;
}
}
else if (Opcode == Instruction::Or || Opcode == Instruction::And) {
for (const Value *A : Args) {
if (const Instruction *I = dyn_cast<Instruction>(A))
if (I->hasOneUse() && I->getOpcode() == Instruction::Xor)
return 0;
}
}
}
// Or requires one instruction, although it has custom handling for i64.
if (Opcode == Instruction::Or)
return 1;
if (Opcode == Instruction::Xor && ScalarBits == 1) {
if (ST->hasLoadStoreOnCond2())
return 5; // 2 * (li 0; loc 1); xor
return 7; // 2 * ipm sequences ; xor ; shift ; compare
}
if (DivRemConstPow2)
return (SignedDivRem ? SDivPow2Cost : 1);
if (DivRemConst)
return DivMulSeqCost;
if (SignedDivRem || UnsignedDivRem)
return DivInstrCost;
}
else if (ST->hasVector()) {
auto *VTy = cast<FixedVectorType>(Ty);
unsigned VF = VTy->getNumElements();
unsigned NumVectors = getNumVectorRegs(Ty);
// These vector operations are custom handled, but are still supported
// with one instruction per vector, regardless of element size.
if (Opcode == Instruction::Shl || Opcode == Instruction::LShr ||
Opcode == Instruction::AShr) {
return NumVectors;
}
if (DivRemConstPow2)
return (NumVectors * (SignedDivRem ? SDivPow2Cost : 1));
if (DivRemConst) {
SmallVector<Type *> Tys(Args.size(), Ty);
return VF * DivMulSeqCost + getScalarizationOverhead(VTy, Args, Tys);
}
if ((SignedDivRem || UnsignedDivRem) && VF > 4)
// Temporary hack: disable high vectorization factors with integer
// division/remainder, which will get scalarized and handled with
// GR128 registers. The mischeduler is not clever enough to avoid
// spilling yet.
return 1000;
// These FP operations are supported with a single vector instruction for
// double (base implementation assumes float generally costs 2). For
// FP128, the scalar cost is 1, and there is no overhead since the values
// are already in scalar registers.
if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub ||
Opcode == Instruction::FMul || Opcode == Instruction::FDiv) {
switch (ScalarBits) {
case 32: {
// The vector enhancements facility 1 provides v4f32 instructions.
if (ST->hasVectorEnhancements1())
return NumVectors;
// Return the cost of multiple scalar invocation plus the cost of
// inserting and extracting the values.
InstructionCost ScalarCost =
getArithmeticInstrCost(Opcode, Ty->getScalarType(), CostKind);
SmallVector<Type *> Tys(Args.size(), Ty);
InstructionCost Cost =
(VF * ScalarCost) + getScalarizationOverhead(VTy, Args, Tys);
// FIXME: VF 2 for these FP operations are currently just as
// expensive as for VF 4.
if (VF == 2)
Cost *= 2;
return Cost;
}
case 64:
case 128:
return NumVectors;
default:
break;
}
}
// There is no native support for FRem.
if (Opcode == Instruction::FRem) {
SmallVector<Type *> Tys(Args.size(), Ty);
unsigned Cost =
(VF * LIBCALL_COST) + getScalarizationOverhead(VTy, Args, Tys);
// FIXME: VF 2 for float is currently just as expensive as for VF 4.
if (VF == 2 && ScalarBits == 32)
Cost *= 2;
return Cost;
}
}
// Fallback to the default implementation.
return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info,
Opd1PropInfo, Opd2PropInfo, Args, CxtI);
}
InstructionCost SystemZTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
VectorType *Tp,
ArrayRef<int> Mask, int Index,
VectorType *SubTp) {
if (ST->hasVector()) {
unsigned NumVectors = getNumVectorRegs(Tp);
// TODO: Since fp32 is expanded, the shuffle cost should always be 0.
// FP128 values are always in scalar registers, so there is no work
// involved with a shuffle, except for broadcast. In that case register
// moves are done with a single instruction per element.
if (Tp->getScalarType()->isFP128Ty())
return (Kind == TargetTransformInfo::SK_Broadcast ? NumVectors - 1 : 0);
switch (Kind) {
case TargetTransformInfo::SK_ExtractSubvector:
// ExtractSubvector Index indicates start offset.
// Extracting a subvector from first index is a noop.
return (Index == 0 ? 0 : NumVectors);
case TargetTransformInfo::SK_Broadcast:
// Loop vectorizer calls here to figure out the extra cost of
// broadcasting a loaded value to all elements of a vector. Since vlrep
// loads and replicates with a single instruction, adjust the returned
// value.
return NumVectors - 1;
default:
// SystemZ supports single instruction permutation / replication.
return NumVectors;
}
}
return BaseT::getShuffleCost(Kind, Tp, Mask, Index, SubTp);
}
// Return the log2 difference of the element sizes of the two vector types.
static unsigned getElSizeLog2Diff(Type *Ty0, Type *Ty1) {
unsigned Bits0 = Ty0->getScalarSizeInBits();
unsigned Bits1 = Ty1->getScalarSizeInBits();
if (Bits1 > Bits0)
return (Log2_32(Bits1) - Log2_32(Bits0));
return (Log2_32(Bits0) - Log2_32(Bits1));
}
// Return the number of instructions needed to truncate SrcTy to DstTy.
unsigned SystemZTTIImpl::
getVectorTruncCost(Type *SrcTy, Type *DstTy) {
assert (SrcTy->isVectorTy() && DstTy->isVectorTy());
assert(SrcTy->getPrimitiveSizeInBits().getFixedSize() >
DstTy->getPrimitiveSizeInBits().getFixedSize() &&
"Packing must reduce size of vector type.");
assert(cast<FixedVectorType>(SrcTy)->getNumElements() ==
cast<FixedVectorType>(DstTy)->getNumElements() &&
"Packing should not change number of elements.");
// TODO: Since fp32 is expanded, the extract cost should always be 0.
unsigned NumParts = getNumVectorRegs(SrcTy);
if (NumParts <= 2)
// Up to 2 vector registers can be truncated efficiently with pack or
// permute. The latter requires an immediate mask to be loaded, which
// typically gets hoisted out of a loop. TODO: return a good value for
// BB-VECTORIZER that includes the immediate loads, which we do not want
// to count for the loop vectorizer.
return 1;
unsigned Cost = 0;
unsigned Log2Diff = getElSizeLog2Diff(SrcTy, DstTy);
unsigned VF = cast<FixedVectorType>(SrcTy)->getNumElements();
for (unsigned P = 0; P < Log2Diff; ++P) {
if (NumParts > 1)
NumParts /= 2;
Cost += NumParts;
}
// Currently, a general mix of permutes and pack instructions is output by
// isel, which follow the cost computation above except for this case which
// is one instruction less:
if (VF == 8 && SrcTy->getScalarSizeInBits() == 64 &&
DstTy->getScalarSizeInBits() == 8)
Cost--;
return Cost;
}
// Return the cost of converting a vector bitmask produced by a compare
// (SrcTy), to the type of the select or extend instruction (DstTy).
unsigned SystemZTTIImpl::
getVectorBitmaskConversionCost(Type *SrcTy, Type *DstTy) {
assert (SrcTy->isVectorTy() && DstTy->isVectorTy() &&
"Should only be called with vector types.");
unsigned PackCost = 0;
unsigned SrcScalarBits = SrcTy->getScalarSizeInBits();
unsigned DstScalarBits = DstTy->getScalarSizeInBits();
unsigned Log2Diff = getElSizeLog2Diff(SrcTy, DstTy);
if (SrcScalarBits > DstScalarBits)
// The bitmask will be truncated.
PackCost = getVectorTruncCost(SrcTy, DstTy);
else if (SrcScalarBits < DstScalarBits) {
unsigned DstNumParts = getNumVectorRegs(DstTy);
// Each vector select needs its part of the bitmask unpacked.
PackCost = Log2Diff * DstNumParts;
// Extra cost for moving part of mask before unpacking.
PackCost += DstNumParts - 1;
}
return PackCost;
}
// Return the type of the compared operands. This is needed to compute the
// cost for a Select / ZExt or SExt instruction.
static Type *getCmpOpsType(const Instruction *I, unsigned VF = 1) {
Type *OpTy = nullptr;
if (CmpInst *CI = dyn_cast<CmpInst>(I->getOperand(0)))
OpTy = CI->getOperand(0)->getType();
else if (Instruction *LogicI = dyn_cast<Instruction>(I->getOperand(0)))
if (LogicI->getNumOperands() == 2)
if (CmpInst *CI0 = dyn_cast<CmpInst>(LogicI->getOperand(0)))
if (isa<CmpInst>(LogicI->getOperand(1)))
OpTy = CI0->getOperand(0)->getType();
if (OpTy != nullptr) {
if (VF == 1) {
assert (!OpTy->isVectorTy() && "Expected scalar type");
return OpTy;
}
// Return the potentially vectorized type based on 'I' and 'VF'. 'I' may
// be either scalar or already vectorized with a same or lesser VF.
Type *ElTy = OpTy->getScalarType();
return FixedVectorType::get(ElTy, VF);
}
return nullptr;
}
// Get the cost of converting a boolean vector to a vector with same width
// and element size as Dst, plus the cost of zero extending if needed.
unsigned SystemZTTIImpl::
getBoolVecToIntConversionCost(unsigned Opcode, Type *Dst,
const Instruction *I) {
auto *DstVTy = cast<FixedVectorType>(Dst);
unsigned VF = DstVTy->getNumElements();
unsigned Cost = 0;
// If we know what the widths of the compared operands, get any cost of
// converting it to match Dst. Otherwise assume same widths.
Type *CmpOpTy = ((I != nullptr) ? getCmpOpsType(I, VF) : nullptr);
if (CmpOpTy != nullptr)
Cost = getVectorBitmaskConversionCost(CmpOpTy, Dst);
if (Opcode == Instruction::ZExt || Opcode == Instruction::UIToFP)
// One 'vn' per dst vector with an immediate mask.
Cost += getNumVectorRegs(Dst);
return Cost;
}
InstructionCost SystemZTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
Type *Src,
TTI::CastContextHint CCH,
TTI::TargetCostKind CostKind,
const Instruction *I) {
// FIXME: Can the logic below also be used for these cost kinds?
if (CostKind == TTI::TCK_CodeSize || CostKind == TTI::TCK_SizeAndLatency) {
auto BaseCost = BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
return BaseCost == 0 ? BaseCost : 1;
}
unsigned DstScalarBits = Dst->getScalarSizeInBits();
unsigned SrcScalarBits = Src->getScalarSizeInBits();
if (!Src->isVectorTy()) {
assert (!Dst->isVectorTy());
if (Opcode == Instruction::SIToFP || Opcode == Instruction::UIToFP) {
if (SrcScalarBits >= 32 ||
(I != nullptr && isa<LoadInst>(I->getOperand(0))))
return 1;
return SrcScalarBits > 1 ? 2 /*i8/i16 extend*/ : 5 /*branch seq.*/;
}
if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
Src->isIntegerTy(1)) {
if (ST->hasLoadStoreOnCond2())
return 2; // li 0; loc 1
// This should be extension of a compare i1 result, which is done with
// ipm and a varying sequence of instructions.
unsigned Cost = 0;
if (Opcode == Instruction::SExt)
Cost = (DstScalarBits < 64 ? 3 : 4);
if (Opcode == Instruction::ZExt)
Cost = 3;
Type *CmpOpTy = ((I != nullptr) ? getCmpOpsType(I) : nullptr);
if (CmpOpTy != nullptr && CmpOpTy->isFloatingPointTy())
// If operands of an fp-type was compared, this costs +1.
Cost++;
return Cost;
}
}
else if (ST->hasVector()) {
// Vector to scalar cast.
auto *SrcVecTy = cast<FixedVectorType>(Src);
auto *DstVecTy = dyn_cast<FixedVectorType>(Dst);
if (!DstVecTy) {
// TODO: tune vector-to-scalar cast.
return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
}
unsigned VF = SrcVecTy->getNumElements();
unsigned NumDstVectors = getNumVectorRegs(Dst);
unsigned NumSrcVectors = getNumVectorRegs(Src);
if (Opcode == Instruction::Trunc) {
if (Src->getScalarSizeInBits() == Dst->getScalarSizeInBits())
return 0; // Check for NOOP conversions.
return getVectorTruncCost(Src, Dst);
}
if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
if (SrcScalarBits >= 8) {
// ZExt/SExt will be handled with one unpack per doubling of width.
unsigned NumUnpacks = getElSizeLog2Diff(Src, Dst);
// For types that spans multiple vector registers, some additional
// instructions are used to setup the unpacking.
unsigned NumSrcVectorOps =
(NumUnpacks > 1 ? (NumDstVectors - NumSrcVectors)
: (NumDstVectors / 2));
return (NumUnpacks * NumDstVectors) + NumSrcVectorOps;
}
else if (SrcScalarBits == 1)
return getBoolVecToIntConversionCost(Opcode, Dst, I);
}
if (Opcode == Instruction::SIToFP || Opcode == Instruction::UIToFP ||
Opcode == Instruction::FPToSI || Opcode == Instruction::FPToUI) {
// TODO: Fix base implementation which could simplify things a bit here
// (seems to miss on differentiating on scalar/vector types).
// Only 64 bit vector conversions are natively supported before z15.
if (DstScalarBits == 64 || ST->hasVectorEnhancements2()) {
if (SrcScalarBits == DstScalarBits)
return NumDstVectors;
if (SrcScalarBits == 1)
return getBoolVecToIntConversionCost(Opcode, Dst, I) + NumDstVectors;
}
// Return the cost of multiple scalar invocation plus the cost of
// inserting and extracting the values. Base implementation does not
// realize float->int gets scalarized.
InstructionCost ScalarCost = getCastInstrCost(
Opcode, Dst->getScalarType(), Src->getScalarType(), CCH, CostKind);
InstructionCost TotCost = VF * ScalarCost;
bool NeedsInserts = true, NeedsExtracts = true;
// FP128 registers do not get inserted or extracted.
if (DstScalarBits == 128 &&
(Opcode == Instruction::SIToFP || Opcode == Instruction::UIToFP))
NeedsInserts = false;
if (SrcScalarBits == 128 &&
(Opcode == Instruction::FPToSI || Opcode == Instruction::FPToUI))
NeedsExtracts = false;
TotCost += getScalarizationOverhead(SrcVecTy, false, NeedsExtracts);
TotCost += getScalarizationOverhead(DstVecTy, NeedsInserts, false);
// FIXME: VF 2 for float<->i32 is currently just as expensive as for VF 4.
if (VF == 2 && SrcScalarBits == 32 && DstScalarBits == 32)
TotCost *= 2;
return TotCost;
}
if (Opcode == Instruction::FPTrunc) {
if (SrcScalarBits == 128) // fp128 -> double/float + inserts of elements.
return VF /*ldxbr/lexbr*/ +
getScalarizationOverhead(DstVecTy, true, false);
else // double -> float
return VF / 2 /*vledb*/ + std::max(1U, VF / 4 /*vperm*/);
}
if (Opcode == Instruction::FPExt) {
if (SrcScalarBits == 32 && DstScalarBits == 64) {
// float -> double is very rare and currently unoptimized. Instead of
// using vldeb, which can do two at a time, all conversions are
// scalarized.
return VF * 2;
}
// -> fp128. VF * lxdb/lxeb + extraction of elements.
return VF + getScalarizationOverhead(SrcVecTy, false, true);
}
}
return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
}
// Scalar i8 / i16 operations will typically be made after first extending
// the operands to i32.
static unsigned getOperandsExtensionCost(const Instruction *I) {
unsigned ExtCost = 0;
for (Value *Op : I->operands())
// A load of i8 or i16 sign/zero extends to i32.
if (!isa<LoadInst>(Op) && !isa<ConstantInt>(Op))
ExtCost++;
return ExtCost;
}
InstructionCost SystemZTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
Type *CondTy,
CmpInst::Predicate VecPred,
TTI::TargetCostKind CostKind,
const Instruction *I) {
if (CostKind != TTI::TCK_RecipThroughput)
return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind);
if (!ValTy->isVectorTy()) {
switch (Opcode) {
case Instruction::ICmp: {
// A loaded value compared with 0 with multiple users becomes Load and
// Test. The load is then not foldable, so return 0 cost for the ICmp.
unsigned ScalarBits = ValTy->getScalarSizeInBits();
if (I != nullptr && ScalarBits >= 32)
if (LoadInst *Ld = dyn_cast<LoadInst>(I->getOperand(0)))
if (const ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1)))
if (!Ld->hasOneUse() && Ld->getParent() == I->getParent() &&
C->isZero())
return 0;
unsigned Cost = 1;
if (ValTy->isIntegerTy() && ValTy->getScalarSizeInBits() <= 16)
Cost += (I != nullptr ? getOperandsExtensionCost(I) : 2);
return Cost;
}
case Instruction::Select:
if (ValTy->isFloatingPointTy())
return 4; // No load on condition for FP - costs a conditional jump.
return 1; // Load On Condition / Select Register.
}
}
else if (ST->hasVector()) {
unsigned VF = cast<FixedVectorType>(ValTy)->getNumElements();
// Called with a compare instruction.
if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) {
unsigned PredicateExtraCost = 0;
if (I != nullptr) {
// Some predicates cost one or two extra instructions.
switch (cast<CmpInst>(I)->getPredicate()) {
case CmpInst::Predicate::ICMP_NE:
case CmpInst::Predicate::ICMP_UGE:
case CmpInst::Predicate::ICMP_ULE:
case CmpInst::Predicate::ICMP_SGE:
case CmpInst::Predicate::ICMP_SLE:
PredicateExtraCost = 1;
break;
case CmpInst::Predicate::FCMP_ONE:
case CmpInst::Predicate::FCMP_ORD:
case CmpInst::Predicate::FCMP_UEQ:
case CmpInst::Predicate::FCMP_UNO:
PredicateExtraCost = 2;
break;
default:
break;
}
}
// Float is handled with 2*vmr[lh]f + 2*vldeb + vfchdb for each pair of
// floats. FIXME: <2 x float> generates same code as <4 x float>.
unsigned CmpCostPerVector = (ValTy->getScalarType()->isFloatTy() ? 10 : 1);
unsigned NumVecs_cmp = getNumVectorRegs(ValTy);
unsigned Cost = (NumVecs_cmp * (CmpCostPerVector + PredicateExtraCost));
return Cost;
}
else { // Called with a select instruction.
assert (Opcode == Instruction::Select);
// We can figure out the extra cost of packing / unpacking if the
// instruction was passed and the compare instruction is found.
unsigned PackCost = 0;
Type *CmpOpTy = ((I != nullptr) ? getCmpOpsType(I, VF) : nullptr);
if (CmpOpTy != nullptr)
PackCost =
getVectorBitmaskConversionCost(CmpOpTy, ValTy);
return getNumVectorRegs(ValTy) /*vsel*/ + PackCost;
}
}
return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind);
}
InstructionCost SystemZTTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
unsigned Index) {
// vlvgp will insert two grs into a vector register, so only count half the
// number of instructions.
if (Opcode == Instruction::InsertElement && Val->isIntOrIntVectorTy(64))
return ((Index % 2 == 0) ? 1 : 0);
if (Opcode == Instruction::ExtractElement) {
int Cost = ((getScalarSizeInBits(Val) == 1) ? 2 /*+test-under-mask*/ : 1);
// Give a slight penalty for moving out of vector pipeline to FXU unit.
if (Index == 0 && Val->isIntOrIntVectorTy())
Cost += 1;
return Cost;
}
return BaseT::getVectorInstrCost(Opcode, Val, Index);
}
// Check if a load may be folded as a memory operand in its user.
bool SystemZTTIImpl::
isFoldableLoad(const LoadInst *Ld, const Instruction *&FoldedValue) {
if (!Ld->hasOneUse())
return false;
FoldedValue = Ld;
const Instruction *UserI = cast<Instruction>(*Ld->user_begin());
unsigned LoadedBits = getScalarSizeInBits(Ld->getType());
unsigned TruncBits = 0;
unsigned SExtBits = 0;
unsigned ZExtBits = 0;
if (UserI->hasOneUse()) {
unsigned UserBits = UserI->getType()->getScalarSizeInBits();
if (isa<TruncInst>(UserI))
TruncBits = UserBits;
else if (isa<SExtInst>(UserI))
SExtBits = UserBits;
else if (isa<ZExtInst>(UserI))
ZExtBits = UserBits;
}
if (TruncBits || SExtBits || ZExtBits) {
FoldedValue = UserI;
UserI = cast<Instruction>(*UserI->user_begin());
// Load (single use) -> trunc/extend (single use) -> UserI
}
if ((UserI->getOpcode() == Instruction::Sub ||