forked from dotnet/runtime
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathlowerarmarch.cpp
1918 lines (1655 loc) · 61.7 KB
/
lowerarmarch.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
// Licensed to the .NET Foundation under one or more agreements.
// The .NET Foundation licenses this file to you under the MIT license.
/*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
XX XX
XX Lowering for ARM and ARM64 common code XX
XX XX
XX This encapsulates common logic for lowering trees for the ARM and ARM64 XX
XX architectures. For a more detailed view of what is lowering, please XX
XX take a look at Lower.cpp XX
XX XX
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
*/
#include "jitpch.h"
#ifdef _MSC_VER
#pragma hdrstop
#endif
#ifdef TARGET_ARMARCH // This file is ONLY used for ARM and ARM64 architectures
#include "jit.h"
#include "sideeffects.h"
#include "lower.h"
#include "lsra.h"
#ifdef FEATURE_HW_INTRINSICS
#include "hwintrinsic.h"
#endif
//------------------------------------------------------------------------
// IsCallTargetInRange: Can a call target address be encoded in-place?
//
// Return Value:
// True if the addr fits into the range.
//
bool Lowering::IsCallTargetInRange(void* addr)
{
return comp->codeGen->validImmForBL((ssize_t)addr);
}
//------------------------------------------------------------------------
// IsContainableImmed: Is an immediate encodable in-place?
//
// Return Value:
// True if the immediate can be folded into an instruction,
// for example small enough and non-relocatable.
//
// TODO-CQ: we can contain a floating point 0.0 constant in a compare instruction
// (vcmp on arm, fcmp on arm64).
//
bool Lowering::IsContainableImmed(GenTree* parentNode, GenTree* childNode) const
{
if (!varTypeIsFloating(parentNode->TypeGet()))
{
// Make sure we have an actual immediate
if (!childNode->IsCnsIntOrI())
return false;
if (childNode->AsIntCon()->ImmedValNeedsReloc(comp))
return false;
// TODO-CrossBitness: we wouldn't need the cast below if GenTreeIntCon::gtIconVal had target_ssize_t type.
target_ssize_t immVal = (target_ssize_t)childNode->AsIntCon()->gtIconVal;
emitAttr attr = emitActualTypeSize(childNode->TypeGet());
emitAttr size = EA_SIZE(attr);
#ifdef TARGET_ARM
insFlags flags = parentNode->gtSetFlags() ? INS_FLAGS_SET : INS_FLAGS_DONT_CARE;
#endif
switch (parentNode->OperGet())
{
case GT_ADD:
case GT_SUB:
return emitter::emitIns_valid_imm_for_add(immVal, size);
#ifdef TARGET_ARM64
case GT_CMPXCHG:
case GT_LOCKADD:
case GT_XORR:
case GT_XAND:
case GT_XADD:
return comp->compOpportunisticallyDependsOn(InstructionSet_Atomics)
? false
: emitter::emitIns_valid_imm_for_add(immVal, size);
#elif defined(TARGET_ARM)
return emitter::emitIns_valid_imm_for_add(immVal, flags);
#endif
break;
#ifdef TARGET_ARM64
case GT_EQ:
case GT_NE:
case GT_LT:
case GT_LE:
case GT_GE:
case GT_GT:
case GT_ARR_BOUNDS_CHECK:
#ifdef FEATURE_SIMD
case GT_SIMD_CHK:
#endif
#ifdef FEATURE_HW_INTRINSICS
case GT_HW_INTRINSIC_CHK:
#endif
return emitter::emitIns_valid_imm_for_cmp(immVal, size);
case GT_AND:
case GT_OR:
case GT_XOR:
case GT_TEST_EQ:
case GT_TEST_NE:
return emitter::emitIns_valid_imm_for_alu(immVal, size);
case GT_JCMP:
assert(((parentNode->gtFlags & GTF_JCMP_TST) == 0) ? (immVal == 0) : isPow2(immVal));
return true;
#elif defined(TARGET_ARM)
case GT_EQ:
case GT_NE:
case GT_LT:
case GT_LE:
case GT_GE:
case GT_GT:
case GT_CMP:
case GT_AND:
case GT_OR:
case GT_XOR:
return emitter::emitIns_valid_imm_for_alu(immVal);
#endif // TARGET_ARM
#ifdef TARGET_ARM64
case GT_STORE_LCL_FLD:
case GT_STORE_LCL_VAR:
if (immVal == 0)
return true;
break;
#endif
default:
break;
}
}
return false;
}
//------------------------------------------------------------------------
// LowerStoreLoc: Lower a store of a lclVar
//
// Arguments:
// storeLoc - the local store (GT_STORE_LCL_FLD or GT_STORE_LCL_VAR)
//
// Notes:
// This involves:
// - Widening operations of unsigneds.
//
void Lowering::LowerStoreLoc(GenTreeLclVarCommon* storeLoc)
{
GenTree* op1 = storeLoc->gtGetOp1();
if ((storeLoc->gtOper == GT_STORE_LCL_VAR) && (op1->gtOper == GT_CNS_INT))
{
// Try to widen the ops if they are going into a local var.
GenTreeIntCon* con = op1->AsIntCon();
ssize_t ival = con->gtIconVal;
unsigned varNum = storeLoc->GetLclNum();
LclVarDsc* varDsc = comp->lvaGetDesc(varNum);
if (varDsc->lvIsSIMDType())
{
noway_assert(storeLoc->gtType != TYP_STRUCT);
}
unsigned size = genTypeSize(storeLoc);
// If we are storing a constant into a local variable
// we extend the size of the store here
if ((size < 4) && !varTypeIsStruct(varDsc))
{
if (!varTypeIsUnsigned(varDsc))
{
if (genTypeSize(storeLoc) == 1)
{
if ((ival & 0x7f) != ival)
{
ival = ival | 0xffffff00;
}
}
else
{
assert(genTypeSize(storeLoc) == 2);
if ((ival & 0x7fff) != ival)
{
ival = ival | 0xffff0000;
}
}
}
// A local stack slot is at least 4 bytes in size, regardless of
// what the local var is typed as, so auto-promote it here
// unless it is a field of a promoted struct
// TODO-CQ: if the field is promoted shouldn't we also be able to do this?
if (!varDsc->lvIsStructField)
{
storeLoc->gtType = TYP_INT;
con->SetIconValue(ival);
}
}
}
if (storeLoc->OperIs(GT_STORE_LCL_FLD))
{
// We should only encounter this for lclVars that are lvDoNotEnregister.
verifyLclFldDoNotEnregister(storeLoc->GetLclNum());
}
ContainCheckStoreLoc(storeLoc);
}
//------------------------------------------------------------------------
// LowerStoreIndir: Determine addressing mode for an indirection, and whether operands are contained.
//
// Arguments:
// node - The indirect store node (GT_STORE_IND) of interest
//
// Return Value:
// None.
//
void Lowering::LowerStoreIndir(GenTreeStoreInd* node)
{
ContainCheckStoreIndir(node);
}
//------------------------------------------------------------------------
// LowerMul: Lower a GT_MUL/GT_MULHI/GT_MUL_LONG node.
//
// For ARM64 recognized GT_MULs that can be turned into GT_MUL_LONGs, as
// those are cheaper. Performs contaiment checks.
//
// Arguments:
// mul - The node to lower
//
// Return Value:
// The next node to lower.
//
GenTree* Lowering::LowerMul(GenTreeOp* mul)
{
assert(mul->OperIsMul());
#ifdef TARGET_ARM64
if (comp->opts.OptimizationEnabled() && mul->OperIs(GT_MUL) && mul->IsValidLongMul())
{
GenTreeCast* op1 = mul->gtGetOp1()->AsCast();
GenTree* op2 = mul->gtGetOp2();
mul->ClearOverflow();
mul->ClearUnsigned();
if (op1->IsUnsigned())
{
mul->SetUnsigned();
}
mul->gtOp1 = op1->CastOp();
BlockRange().Remove(op1);
if (op2->OperIs(GT_CAST))
{
mul->gtOp2 = op2->AsCast()->CastOp();
BlockRange().Remove(op2);
}
else
{
assert(op2->IsIntegralConst());
assert(FitsIn<int32_t>(op2->AsIntConCommon()->IntegralValue()));
op2->ChangeType(TYP_INT);
}
mul->ChangeOper(GT_MUL_LONG);
}
#endif // TARGET_ARM64
ContainCheckMul(mul);
return mul->gtNext;
}
//------------------------------------------------------------------------
// LowerBlockStore: Lower a block store node
//
// Arguments:
// blkNode - The block store node to lower
//
void Lowering::LowerBlockStore(GenTreeBlk* blkNode)
{
GenTree* dstAddr = blkNode->Addr();
GenTree* src = blkNode->Data();
unsigned size = blkNode->Size();
if (blkNode->OperIsInitBlkOp())
{
if (src->OperIs(GT_INIT_VAL))
{
src->SetContained();
src = src->AsUnOp()->gtGetOp1();
}
if (blkNode->OperIs(GT_STORE_OBJ))
{
blkNode->SetOper(GT_STORE_BLK);
}
if (!blkNode->OperIs(GT_STORE_DYN_BLK) && (size <= INITBLK_UNROLL_LIMIT) && src->OperIs(GT_CNS_INT))
{
blkNode->gtBlkOpKind = GenTreeBlk::BlkOpKindUnroll;
// The fill value of an initblk is interpreted to hold a
// value of (unsigned int8) however a constant of any size
// may practically reside on the evaluation stack. So extract
// the lower byte out of the initVal constant and replicate
// it to a larger constant whose size is sufficient to support
// the largest width store of the desired inline expansion.
ssize_t fill = src->AsIntCon()->IconValue() & 0xFF;
if (fill == 0)
{
#ifdef TARGET_ARM64
// On ARM64 we can just use REG_ZR instead of having to load
// the constant into a real register like on ARM32.
src->SetContained();
#endif
}
#ifdef TARGET_ARM64
else if (size >= REGSIZE_BYTES)
{
fill *= 0x0101010101010101LL;
src->gtType = TYP_LONG;
}
#endif
else
{
fill *= 0x01010101;
}
src->AsIntCon()->SetIconValue(fill);
ContainBlockStoreAddress(blkNode, size, dstAddr);
}
else
{
blkNode->gtBlkOpKind = GenTreeBlk::BlkOpKindHelper;
}
}
else
{
assert(src->OperIs(GT_IND, GT_LCL_VAR, GT_LCL_FLD));
src->SetContained();
if (src->OperIs(GT_IND))
{
// TODO-Cleanup: Make sure that GT_IND lowering didn't mark the source address as contained.
// Sometimes the GT_IND type is a non-struct type and then GT_IND lowering may contain the
// address, not knowing that GT_IND is part of a block op that has containment restrictions.
src->AsIndir()->Addr()->ClearContained();
}
else if (src->OperIs(GT_LCL_VAR))
{
// TODO-1stClassStructs: for now we can't work with STORE_BLOCK source in register.
const unsigned srcLclNum = src->AsLclVar()->GetLclNum();
comp->lvaSetVarDoNotEnregister(srcLclNum DEBUGARG(DoNotEnregisterReason::BlockOp));
}
if (blkNode->OperIs(GT_STORE_OBJ))
{
if (!blkNode->AsObj()->GetLayout()->HasGCPtr())
{
blkNode->SetOper(GT_STORE_BLK);
}
else if (dstAddr->OperIsLocalAddr() && (size <= CPBLK_UNROLL_LIMIT))
{
// If the size is small enough to unroll then we need to mark the block as non-interruptible
// to actually allow unrolling. The generated code does not report GC references loaded in the
// temporary register(s) used for copying.
blkNode->SetOper(GT_STORE_BLK);
blkNode->gtBlkOpGcUnsafe = true;
}
}
if (blkNode->OperIs(GT_STORE_OBJ))
{
assert((dstAddr->TypeGet() == TYP_BYREF) || (dstAddr->TypeGet() == TYP_I_IMPL));
blkNode->gtBlkOpKind = GenTreeBlk::BlkOpKindUnroll;
}
else if (blkNode->OperIs(GT_STORE_BLK) && (size <= CPBLK_UNROLL_LIMIT))
{
blkNode->gtBlkOpKind = GenTreeBlk::BlkOpKindUnroll;
if (src->OperIs(GT_IND))
{
ContainBlockStoreAddress(blkNode, size, src->AsIndir()->Addr());
}
ContainBlockStoreAddress(blkNode, size, dstAddr);
}
else
{
assert(blkNode->OperIs(GT_STORE_BLK, GT_STORE_DYN_BLK));
blkNode->gtBlkOpKind = GenTreeBlk::BlkOpKindHelper;
}
}
}
//------------------------------------------------------------------------
// ContainBlockStoreAddress: Attempt to contain an address used by an unrolled block store.
//
// Arguments:
// blkNode - the block store node
// size - the block size
// addr - the address node to try to contain
//
void Lowering::ContainBlockStoreAddress(GenTreeBlk* blkNode, unsigned size, GenTree* addr)
{
assert(blkNode->OperIs(GT_STORE_BLK) && (blkNode->gtBlkOpKind == GenTreeBlk::BlkOpKindUnroll));
assert(size < INT32_MAX);
if (addr->OperIsLocalAddr())
{
addr->SetContained();
return;
}
if (!addr->OperIs(GT_ADD) || addr->gtOverflow() || !addr->AsOp()->gtGetOp2()->OperIs(GT_CNS_INT))
{
return;
}
GenTreeIntCon* offsetNode = addr->AsOp()->gtGetOp2()->AsIntCon();
ssize_t offset = offsetNode->IconValue();
// All integer load/store instructions on both ARM32 and ARM64 support
// offsets in range -255..255. Of course, this is a rather conservative
// check. For example, if the offset and size are a multiple of 8 we
// could allow a combined offset of up to 32760 on ARM64.
if ((offset < -255) || (offset > 255) || (offset + static_cast<int>(size) > 256))
{
return;
}
#ifdef TARGET_ARM64
// If we're going to use LDP/STP we need to ensure that the offset is
// a multiple of 8 since these instructions do not have an unscaled
// offset variant.
if ((size >= 2 * REGSIZE_BYTES) && (offset % REGSIZE_BYTES != 0))
{
return;
}
#endif
if (!IsSafeToContainMem(blkNode, addr))
{
return;
}
BlockRange().Remove(offsetNode);
addr->ChangeOper(GT_LEA);
addr->AsAddrMode()->SetIndex(nullptr);
addr->AsAddrMode()->SetScale(0);
addr->AsAddrMode()->SetOffset(static_cast<int>(offset));
addr->SetContained();
}
//------------------------------------------------------------------------
// LowerCast: Lower GT_CAST(srcType, DstType) nodes.
//
// Arguments:
// tree - GT_CAST node to be lowered
//
// Return Value:
// None.
//
// Notes:
// Casts from float/double to a smaller int type are transformed as follows:
// GT_CAST(float/double, byte) = GT_CAST(GT_CAST(float/double, int32), byte)
// GT_CAST(float/double, sbyte) = GT_CAST(GT_CAST(float/double, int32), sbyte)
// GT_CAST(float/double, int16) = GT_CAST(GT_CAST(double/double, int32), int16)
// GT_CAST(float/double, uint16) = GT_CAST(GT_CAST(double/double, int32), uint16)
//
// Note that for the overflow conversions we still depend on helper calls and
// don't expect to see them here.
// i) GT_CAST(float/double, int type with overflow detection)
//
void Lowering::LowerCast(GenTree* tree)
{
assert(tree->OperGet() == GT_CAST);
JITDUMP("LowerCast for: ");
DISPNODE(tree);
JITDUMP("\n");
GenTree* op1 = tree->AsOp()->gtOp1;
var_types dstType = tree->CastToType();
var_types srcType = genActualType(op1->TypeGet());
var_types tmpType = TYP_UNDEF;
if (varTypeIsFloating(srcType))
{
noway_assert(!tree->gtOverflow());
assert(!varTypeIsSmall(dstType)); // fgMorphCast creates intermediate casts when converting from float to small
// int.
}
assert(!varTypeIsSmall(srcType));
if (tmpType != TYP_UNDEF)
{
GenTree* tmp = comp->gtNewCastNode(tmpType, op1, tree->IsUnsigned(), tmpType);
tmp->gtFlags |= (tree->gtFlags & (GTF_OVERFLOW | GTF_EXCEPT));
tree->gtFlags &= ~GTF_UNSIGNED;
tree->AsOp()->gtOp1 = tmp;
BlockRange().InsertAfter(op1, tmp);
}
// Now determine if we have operands that should be contained.
ContainCheckCast(tree->AsCast());
}
//------------------------------------------------------------------------
// LowerRotate: Lower GT_ROL and GT_ROR nodes.
//
// Arguments:
// tree - the node to lower
//
// Return Value:
// None.
//
void Lowering::LowerRotate(GenTree* tree)
{
if (tree->OperGet() == GT_ROL)
{
// There is no ROL instruction on ARM. Convert ROL into ROR.
GenTree* rotatedValue = tree->AsOp()->gtOp1;
unsigned rotatedValueBitSize = genTypeSize(rotatedValue->gtType) * 8;
GenTree* rotateLeftIndexNode = tree->AsOp()->gtOp2;
if (rotateLeftIndexNode->IsCnsIntOrI())
{
ssize_t rotateLeftIndex = rotateLeftIndexNode->AsIntCon()->gtIconVal;
ssize_t rotateRightIndex = rotatedValueBitSize - rotateLeftIndex;
rotateLeftIndexNode->AsIntCon()->gtIconVal = rotateRightIndex;
}
else
{
GenTree* tmp = comp->gtNewOperNode(GT_NEG, genActualType(rotateLeftIndexNode->gtType), rotateLeftIndexNode);
BlockRange().InsertAfter(rotateLeftIndexNode, tmp);
tree->AsOp()->gtOp2 = tmp;
}
tree->ChangeOper(GT_ROR);
}
ContainCheckShiftRotate(tree->AsOp());
}
#ifdef FEATURE_SIMD
//----------------------------------------------------------------------------------------------
// Lowering::LowerSIMD: Perform containment analysis for a SIMD intrinsic node.
//
// Arguments:
// simdNode - The SIMD intrinsic node.
//
void Lowering::LowerSIMD(GenTreeSIMD* simdNode)
{
assert(simdNode->gtType != TYP_SIMD32);
if (simdNode->TypeGet() == TYP_SIMD12)
{
// GT_SIMD node requiring to produce TYP_SIMD12 in fact
// produces a TYP_SIMD16 result
simdNode->gtType = TYP_SIMD16;
}
ContainCheckSIMD(simdNode);
}
#endif // FEATURE_SIMD
#ifdef FEATURE_HW_INTRINSICS
//----------------------------------------------------------------------------------------------
// LowerHWIntrinsicFusedMultiplyAddScalar: Lowers AdvSimd_FusedMultiplyAddScalar intrinsics
// when some of the operands are negated by "containing" such negation.
//
// Arguments:
// node - The original hardware intrinsic node
//
// | op1 | op2 | op3 |
// | + | + | + | AdvSimd_FusedMultiplyAddScalar
// | + | + | - | AdvSimd_FusedMultiplySubtractScalar
// | + | - | + | AdvSimd_FusedMultiplySubtractScalar
// | + | - | - | AdvSimd_FusedMultiplyAddScalar
// | - | + | + | AdvSimd_FusedMultiplySubtractNegatedScalar
// | - | + | - | AdvSimd_FusedMultiplyAddNegatedScalar
// | - | - | + | AdvSimd_FusedMultiplyAddNegatedScalar
// | - | - | - | AdvSimd_FusedMultiplySubtractNegatedScalar
//
void Lowering::LowerHWIntrinsicFusedMultiplyAddScalar(GenTreeHWIntrinsic* node)
{
assert(node->gtHWIntrinsicId == NI_AdvSimd_FusedMultiplyAddScalar);
const HWIntrinsic intrin(node);
GenTree* op1 = intrin.op1;
GenTree* op2 = intrin.op2;
GenTree* op3 = intrin.op3;
auto lowerOperand = [this](GenTree* op) {
bool wasNegated = false;
if (op->OperIsHWIntrinsic() &&
((op->AsHWIntrinsic()->gtHWIntrinsicId == NI_AdvSimd_Arm64_DuplicateToVector64) ||
(op->AsHWIntrinsic()->gtHWIntrinsicId == NI_Vector64_CreateScalarUnsafe)))
{
GenTreeHWIntrinsic* createVector64 = op->AsHWIntrinsic();
GenTree* valueOp = createVector64->gtGetOp1();
if (valueOp->OperIs(GT_NEG))
{
createVector64->gtOp1 = valueOp->gtGetOp1();
BlockRange().Remove(valueOp);
wasNegated = true;
}
}
return wasNegated;
};
const bool op1WasNegated = lowerOperand(op1);
const bool op2WasNegated = lowerOperand(op2);
const bool op3WasNegated = lowerOperand(op3);
if (op1WasNegated)
{
if (op2WasNegated != op3WasNegated)
{
node->gtHWIntrinsicId = NI_AdvSimd_FusedMultiplyAddNegatedScalar;
}
else
{
node->gtHWIntrinsicId = NI_AdvSimd_FusedMultiplySubtractNegatedScalar;
}
}
else if (op2WasNegated != op3WasNegated)
{
node->gtHWIntrinsicId = NI_AdvSimd_FusedMultiplySubtractScalar;
}
}
//----------------------------------------------------------------------------------------------
// Lowering::LowerHWIntrinsic: Perform containment analysis for a hardware intrinsic node.
//
// Arguments:
// node - The hardware intrinsic node.
//
void Lowering::LowerHWIntrinsic(GenTreeHWIntrinsic* node)
{
assert(node->TypeGet() != TYP_SIMD32);
if (node->TypeGet() == TYP_SIMD12)
{
// GT_HWINTRINSIC node requiring to produce TYP_SIMD12 in fact
// produces a TYP_SIMD16 result
node->gtType = TYP_SIMD16;
}
NamedIntrinsic intrinsicId = node->gtHWIntrinsicId;
switch (intrinsicId)
{
case NI_Vector64_Create:
case NI_Vector128_Create:
{
// We don't directly support the Vector64.Create or Vector128.Create methods in codegen
// and instead lower them to other intrinsic nodes in LowerHWIntrinsicCreate so we expect
// that the node is modified to either not be a HWIntrinsic node or that it is no longer
// the same intrinsic as when it came in.
LowerHWIntrinsicCreate(node);
assert(!node->OperIsHWIntrinsic() || (node->gtHWIntrinsicId != intrinsicId));
LowerNode(node);
return;
}
case NI_Vector64_Dot:
case NI_Vector128_Dot:
{
LowerHWIntrinsicDot(node);
return;
}
case NI_Vector64_op_Equality:
case NI_Vector128_op_Equality:
{
LowerHWIntrinsicCmpOp(node, GT_EQ);
return;
}
case NI_Vector64_op_Inequality:
case NI_Vector128_op_Inequality:
{
LowerHWIntrinsicCmpOp(node, GT_NE);
return;
}
case NI_AdvSimd_FusedMultiplyAddScalar:
LowerHWIntrinsicFusedMultiplyAddScalar(node);
break;
default:
break;
}
ContainCheckHWIntrinsic(node);
}
//----------------------------------------------------------------------------------------------
// Lowering::IsValidConstForMovImm: Determines if the given node can be replaced by a mov/fmov immediate instruction
//
// Arguments:
// node - The hardware intrinsic node.
//
// Returns:
// true if the node can be replaced by a mov/fmov immediate instruction; otherwise, false
//
// IMPORTANT:
// This check may end up modifying node->gtOp1 if it is a cast node that can be removed
bool Lowering::IsValidConstForMovImm(GenTreeHWIntrinsic* node)
{
assert((node->gtHWIntrinsicId == NI_Vector64_Create) || (node->gtHWIntrinsicId == NI_Vector128_Create) ||
(node->gtHWIntrinsicId == NI_Vector64_CreateScalarUnsafe) ||
(node->gtHWIntrinsicId == NI_Vector128_CreateScalarUnsafe) ||
(node->gtHWIntrinsicId == NI_AdvSimd_DuplicateToVector64) ||
(node->gtHWIntrinsicId == NI_AdvSimd_DuplicateToVector128) ||
(node->gtHWIntrinsicId == NI_AdvSimd_Arm64_DuplicateToVector64) ||
(node->gtHWIntrinsicId == NI_AdvSimd_Arm64_DuplicateToVector128));
assert(HWIntrinsicInfo::lookupNumArgs(node) == 1);
GenTree* op1 = node->gtOp1;
GenTree* castOp = nullptr;
if (varTypeIsIntegral(node->GetSimdBaseType()) && op1->OperIs(GT_CAST))
{
// We will sometimes get a cast around a constant value (such as for
// certain long constants) which would block the below containment.
// So we will temporarily check what the cast is from instead so we
// can catch those cases as well.
castOp = op1->AsCast()->CastOp();
op1 = castOp;
}
if (op1->IsCnsIntOrI())
{
const ssize_t dataValue = op1->AsIntCon()->gtIconVal;
if (comp->GetEmitter()->emitIns_valid_imm_for_movi(dataValue, emitActualTypeSize(node->GetSimdBaseType())))
{
if (castOp != nullptr)
{
// We found a containable immediate under
// a cast, so remove the cast from the LIR.
BlockRange().Remove(node->gtOp1);
node->gtOp1 = op1;
}
return true;
}
}
else if (op1->IsCnsFltOrDbl())
{
assert(varTypeIsFloating(node->GetSimdBaseType()));
assert(castOp == nullptr);
const double dataValue = op1->AsDblCon()->gtDconVal;
return comp->GetEmitter()->emitIns_valid_imm_for_fmov(dataValue);
}
return false;
}
//----------------------------------------------------------------------------------------------
// Lowering::LowerHWIntrinsicCmpOp: Lowers a Vector128 or Vector256 comparison intrinsic
//
// Arguments:
// node - The hardware intrinsic node.
// cmpOp - The comparison operation, currently must be GT_EQ or GT_NE
//
void Lowering::LowerHWIntrinsicCmpOp(GenTreeHWIntrinsic* node, genTreeOps cmpOp)
{
NamedIntrinsic intrinsicId = node->gtHWIntrinsicId;
CorInfoType simdBaseJitType = node->GetSimdBaseJitType();
var_types simdBaseType = node->GetSimdBaseType();
unsigned simdSize = node->GetSimdSize();
var_types simdType = Compiler::getSIMDTypeForSize(simdSize);
assert((intrinsicId == NI_Vector64_op_Equality) || (intrinsicId == NI_Vector64_op_Inequality) ||
(intrinsicId == NI_Vector128_op_Equality) || (intrinsicId == NI_Vector128_op_Inequality));
assert(varTypeIsSIMD(simdType));
assert(varTypeIsArithmetic(simdBaseType));
assert(simdSize != 0);
assert(node->gtType == TYP_BOOL);
assert((cmpOp == GT_EQ) || (cmpOp == GT_NE));
// We have the following (with the appropriate simd size and where the intrinsic could be op_Inequality):
// /--* op2 simd
// /--* op1 simd
// node = * HWINTRINSIC simd T op_Equality
GenTree* op1 = node->gtGetOp1();
GenTree* op2 = node->gtGetOp2();
NamedIntrinsic cmpIntrinsic;
switch (simdBaseType)
{
case TYP_BYTE:
case TYP_UBYTE:
case TYP_SHORT:
case TYP_USHORT:
case TYP_INT:
case TYP_UINT:
case TYP_FLOAT:
{
cmpIntrinsic = NI_AdvSimd_CompareEqual;
break;
}
case TYP_LONG:
case TYP_ULONG:
case TYP_DOUBLE:
{
cmpIntrinsic = (simdSize == 8) ? NI_AdvSimd_Arm64_CompareEqualScalar : NI_AdvSimd_Arm64_CompareEqual;
break;
}
default:
{
unreached();
}
}
GenTree* cmp = comp->gtNewSimdHWIntrinsicNode(simdType, op1, op2, cmpIntrinsic, simdBaseJitType, simdSize);
BlockRange().InsertBefore(node, cmp);
LowerNode(cmp);
if ((simdBaseType == TYP_FLOAT) && (simdSize == 12))
{
// For TYP_SIMD12 we don't want the upper bits to participate in the comparison. So, we will insert all ones
// into those bits of the result, "as if" the upper bits are equal. Then if all lower bits are equal, we get the
// expected all-ones result, and will get the expected 0's only where there are non-matching bits.
GenTree* idxCns = comp->gtNewIconNode(3, TYP_INT);
BlockRange().InsertAfter(cmp, idxCns);
GenTree* insCns = comp->gtNewIconNode(-1, TYP_INT);
BlockRange().InsertAfter(idxCns, insCns);
GenTree* tmp = comp->gtNewSimdAsHWIntrinsicNode(simdType, cmp, idxCns, insCns, NI_AdvSimd_Insert,
CORINFO_TYPE_INT, simdSize);
BlockRange().InsertAfter(insCns, tmp);
LowerNode(tmp);
cmp = tmp;
}
GenTree* msk =
comp->gtNewSimdHWIntrinsicNode(simdType, cmp, NI_AdvSimd_Arm64_MinAcross, CORINFO_TYPE_UBYTE, simdSize);
BlockRange().InsertAfter(cmp, msk);
LowerNode(msk);
GenTree* zroCns = comp->gtNewIconNode(0, TYP_INT);
BlockRange().InsertAfter(msk, zroCns);
GenTree* val =
comp->gtNewSimdAsHWIntrinsicNode(TYP_UBYTE, msk, zroCns, NI_AdvSimd_Extract, CORINFO_TYPE_UBYTE, simdSize);
BlockRange().InsertAfter(zroCns, val);
LowerNode(val);
zroCns = comp->gtNewIconNode(0, TYP_INT);
BlockRange().InsertAfter(val, zroCns);
node->ChangeOper(cmpOp);
node->gtType = TYP_INT;
node->gtOp1 = val;
node->gtOp2 = zroCns;
// The CompareEqual will set (condition is true) or clear (condition is false) all bits of the respective element
// The MinAcross then ensures we get either all bits set (all conditions are true) or clear (any condition is false)
// So, we need to invert the condition from the operation since we compare against zero
GenCondition cmpCnd = (cmpOp == GT_EQ) ? GenCondition::NE : GenCondition::EQ;
GenTree* cc = LowerNodeCC(node, cmpCnd);
node->gtType = TYP_VOID;
node->ClearUnusedValue();
LowerNode(node);
}
//----------------------------------------------------------------------------------------------
// Lowering::LowerHWIntrinsicCreate: Lowers a Vector64 or Vector128 Create call
//
// Arguments:
// node - The hardware intrinsic node.
//
void Lowering::LowerHWIntrinsicCreate(GenTreeHWIntrinsic* node)
{
NamedIntrinsic intrinsicId = node->gtHWIntrinsicId;
var_types simdType = node->gtType;
CorInfoType simdBaseJitType = node->GetSimdBaseJitType();
var_types simdBaseType = node->GetSimdBaseType();
unsigned simdSize = node->GetSimdSize();
VectorConstant vecCns = {};
if ((simdSize == 8) && (simdType == TYP_DOUBLE))
{
// TODO-Cleanup: Struct retyping means we have the wrong type here. We need to
// manually fix it up so the simdType checks below are correct.
simdType = TYP_SIMD8;
}
assert(varTypeIsSIMD(simdType));
assert(varTypeIsArithmetic(simdBaseType));
assert(simdSize != 0);
GenTreeArgList* argList = nullptr;
GenTree* op1 = node->gtGetOp1();
GenTree* op2 = node->gtGetOp2();
// Spare GenTrees to be used for the lowering logic below
// Defined upfront to avoid naming conflicts, etc...
GenTree* idx = nullptr;
GenTree* tmp1 = nullptr;
GenTree* tmp2 = nullptr;
GenTree* tmp3 = nullptr;
assert(op1 != nullptr);
unsigned argCnt = 0;
unsigned cnsArgCnt = 0;
if (op1->OperIsList())
{
assert(op2 == nullptr);
for (argList = op1->AsArgList(); argList != nullptr; argList = argList->Rest())
{
if (HandleArgForHWIntrinsicCreate(argList->Current(), argCnt, vecCns, simdBaseType))
{
cnsArgCnt += 1;
}
argCnt += 1;
}
}
else
{
if (HandleArgForHWIntrinsicCreate(op1, argCnt, vecCns, simdBaseType))
{
cnsArgCnt += 1;
}
argCnt += 1;
if (op2 != nullptr)
{
if (HandleArgForHWIntrinsicCreate(op2, argCnt, vecCns, simdBaseType))
{
cnsArgCnt += 1;
}
argCnt += 1;
}
else if (cnsArgCnt == 1)
{
// These intrinsics are meant to set the same value to every element
// so we'll just specially handle it here and copy it into the remaining
// indices.
for (unsigned i = 1; i < simdSize / genTypeSize(simdBaseType); i++)
{
HandleArgForHWIntrinsicCreate(op1, i, vecCns, simdBaseType);
}
}
}
assert((argCnt == 1) || (argCnt == (simdSize / genTypeSize(simdBaseType))));
if ((argCnt == cnsArgCnt) && (argCnt == 1))
{
GenTree* castOp = nullptr;
if (varTypeIsIntegral(simdBaseType) && op1->OperIs(GT_CAST))
{
// We will sometimes get a cast around a constant value (such as for
// certain long constants) which would block the below containment.
// So we will temporarily check what the cast is from instead so we
// can catch those cases as well.