-
Notifications
You must be signed in to change notification settings - Fork 2.2k
/
ArmCompVFPUNEON.cpp
1455 lines (1234 loc) · 37.2 KB
/
ArmCompVFPUNEON.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
// Copyright (c) 2013- PPSSPP Project.
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, version 2.0 or later versions.
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License 2.0 for more details.
// A copy of the GPL 2.0 should have been included with the program.
// If not, see http://www.gnu.org/licenses/
// Official git repository and contact information can be found at
// https://github.com/hrydgard/ppsspp and http://www.ppsspp.org/.
// NEON VFPU
// This is where we will create an alternate implementation of the VFPU emulation
// that uses NEON Q registers to cache pairs/tris/quads, and so on.
// Will require major extensions to the reg cache and other things.
// ARM NEON can only do pairs and quads, not tris and scalars.
// We can do scalars, though, for many operations if all the operands
// are below Q8 (D16, S32) using regular VFP instructions but really not sure
// if it's worth it.
#include "ppsspp_config.h"
#if PPSSPP_ARCH(ARM)
#include <cmath>
#include "base/logging.h"
#include "math/math_util.h"
#include "Common/CPUDetect.h"
#include "Core/MemMap.h"
#include "Core/MIPS/MIPS.h"
#include "Core/MIPS/MIPSAnalyst.h"
#include "Core/MIPS/MIPSCodeUtils.h"
#include "Core/MIPS/MIPSVFPUUtils.h"
#include "Core/Config.h"
#include "Core/Reporting.h"
#include "Core/MIPS/ARM/ArmJit.h"
#include "Core/MIPS/ARM/ArmRegCache.h"
#include "Core/MIPS/ARM/ArmRegCacheFPU.h"
#include "Core/MIPS/ARM/ArmCompVFPUNEONUtil.h"
// TODO: Somehow #ifdef away on ARMv5eabi, without breaking the linker.
// All functions should have CONDITIONAL_DISABLE, so we can narrow things down to a file quickly.
// Currently known non working ones should have DISABLE.
// #define CONDITIONAL_DISABLE { fpr.ReleaseSpillLocksAndDiscardTemps(); Comp_Generic(op); return; }
#define CONDITIONAL_DISABLE ;
#define DISABLE { fpr.ReleaseSpillLocksAndDiscardTemps(); Comp_Generic(op); return; }
#define DISABLE_UNKNOWN_PREFIX { WLOG("DISABLE: Unknown Prefix in %s", __FUNCTION__); fpr.ReleaseSpillLocksAndDiscardTemps(); Comp_Generic(op); return; }
#define _RS MIPS_GET_RS(op)
#define _RT MIPS_GET_RT(op)
#define _RD MIPS_GET_RD(op)
#define _FS MIPS_GET_FS(op)
#define _FT MIPS_GET_FT(op)
#define _FD MIPS_GET_FD(op)
#define _SA MIPS_GET_SA(op)
#define _POS ((op>> 6) & 0x1F)
#define _SIZE ((op>>11) & 0x1F)
#define _IMM16 (signed short)(op & 0xFFFF)
#define _IMM26 (op & 0x03FFFFFF)
namespace MIPSComp {
using namespace ArmGen;
using namespace ArmJitConstants;
static const float minus_one = -1.0f;
static const float one = 1.0f;
static const float zero = 0.0f;
void ArmJit::CompNEON_VecDo3(MIPSOpcode op) {
CONDITIONAL_DISABLE;
if (js.HasUnknownPrefix()) {
DISABLE_UNKNOWN_PREFIX;
}
VectorSize sz = GetVecSize(op);
int n = GetNumVectorElements(sz);
MappedRegs r = NEONMapDirtyInIn(op, sz, sz, sz);
ARMReg temp = MatchSize(Q0, r.vs);
// TODO: Special case for scalar
switch (op >> 26) {
case 24: //VFPU0
switch ((op >> 23) & 7) {
case 0: VADD(F_32, r.vd, r.vs, r.vt); break; // vadd
case 1: VSUB(F_32, r.vd, r.vs, r.vt); break; // vsub
case 7: // vdiv // vdiv THERE IS NO NEON SIMD VDIV :( There's a fast reciprocal iterator thing though.
{
// Implement by falling back to VFP
VMOV(D0, D_0(r.vs));
VMOV(D1, D_0(r.vt));
VDIV(S0, S0, S2);
if (sz >= V_Pair)
VDIV(S1, S1, S3);
VMOV(D_0(r.vd), D0);
if (sz >= V_Triple) {
VMOV(D0, D_1(r.vs));
VMOV(D1, D_1(r.vt));
VDIV(S0, S0, S2);
if (sz == V_Quad)
VDIV(S1, S1, S3);
VMOV(D_1(r.vd), D0);
}
}
break;
default:
DISABLE;
}
break;
case 25: //VFPU1
switch ((op >> 23) & 7) {
case 0: VMUL(F_32, r.vd, r.vs, r.vt); break; // vmul
default:
DISABLE;
}
break;
case 27: //VFPU3
switch ((op >> 23) & 7) {
case 2: VMIN(F_32, r.vd, r.vs, r.vt); break; // vmin
case 3: VMAX(F_32, r.vd, r.vs, r.vt); break; // vmax
case 6: // vsge
VMOV_immf(temp, 1.0f);
VCGE(F_32, r.vd, r.vs, r.vt);
VAND(r.vd, r.vd, temp);
break;
case 7: // vslt
VMOV_immf(temp, 1.0f);
VCLT(F_32, r.vd, r.vs, r.vt);
VAND(r.vd, r.vd, temp);
break;
}
break;
default:
DISABLE;
}
NEONApplyPrefixD(r.vd);
fpr.ReleaseSpillLocksAndDiscardTemps();
}
// #define CONDITIONAL_DISABLE { fpr.ReleaseSpillLocksAndDiscardTemps(); Comp_Generic(op); return; }
void ArmJit::CompNEON_SV(MIPSOpcode op) {
CONDITIONAL_DISABLE;
// Remember to use single lane stores here and not VLDR/VSTR - switching usage
// between NEON and VFPU can be expensive on some chips.
// Here's a common idiom we should optimize:
// lv.s S200, 0(s4)
// lv.s S201, 4(s4)
// lv.s S202, 8(s4)
// vone.s S203
// vtfm4.q C000, E600, C200
// Would be great if we could somehow combine the lv.s into one vector instead of mapping three
// separate quads.
s32 offset = (signed short)(op & 0xFFFC);
int vt = ((op >> 16) & 0x1f) | ((op & 3) << 5);
MIPSGPReg rs = _RS;
bool doCheck = false;
switch (op >> 26)
{
case 50: //lv.s // VI(vt) = Memory::Read_U32(addr);
{
if (!gpr.IsImm(rs) && jo.cachePointers && g_Config.bFastMemory && (offset & 3) == 0 && offset < 0x400 && offset > -0x400) {
INFO_LOG(HLE, "LV.S fastmode!");
// TODO: Also look forward and combine multiple loads.
gpr.MapRegAsPointer(rs);
ARMReg ar = fpr.QMapReg(vt, V_Single, MAP_NOINIT | MAP_DIRTY);
if (offset) {
ADDI2R(R0, gpr.RPtr(rs), offset, R1);
VLD1_lane(F_32, ar, R0, 0, true);
} else {
VLD1_lane(F_32, ar, gpr.RPtr(rs), 0, true);
}
break;
}
INFO_LOG(HLE, "LV.S slowmode!");
// CC might be set by slow path below, so load regs first.
ARMReg ar = fpr.QMapReg(vt, V_Single, MAP_DIRTY | MAP_NOINIT);
if (gpr.IsImm(rs)) {
u32 addr = (offset + gpr.GetImm(rs)) & 0x3FFFFFFF;
gpr.SetRegImm(R0, addr + (u32)Memory::base);
} else {
gpr.MapReg(rs);
if (g_Config.bFastMemory) {
SetR0ToEffectiveAddress(rs, offset);
} else {
SetCCAndR0ForSafeAddress(rs, offset, R1);
doCheck = true;
}
ADD(R0, R0, MEMBASEREG);
}
FixupBranch skip;
if (doCheck) {
skip = B_CC(CC_EQ);
}
VLD1_lane(F_32, ar, R0, 0, true);
if (doCheck) {
SetJumpTarget(skip);
SetCC(CC_AL);
}
}
break;
case 58: //sv.s // Memory::Write_U32(VI(vt), addr);
{
if (!gpr.IsImm(rs) && jo.cachePointers && g_Config.bFastMemory && (offset & 3) == 0 && offset < 0x400 && offset > -0x400) {
INFO_LOG(HLE, "SV.S fastmode!");
// TODO: Also look forward and combine multiple stores.
gpr.MapRegAsPointer(rs);
ARMReg ar = fpr.QMapReg(vt, V_Single, 0);
if (offset) {
ADDI2R(R0, gpr.RPtr(rs), offset, R1);
VST1_lane(F_32, ar, R0, 0, true);
} else {
VST1_lane(F_32, ar, gpr.RPtr(rs), 0, true);
}
break;
}
INFO_LOG(HLE, "SV.S slowmode!");
// CC might be set by slow path below, so load regs first.
ARMReg ar = fpr.QMapReg(vt, V_Single, 0);
if (gpr.IsImm(rs)) {
u32 addr = (offset + gpr.GetImm(rs)) & 0x3FFFFFFF;
gpr.SetRegImm(R0, addr + (u32)Memory::base);
} else {
gpr.MapReg(rs);
if (g_Config.bFastMemory) {
SetR0ToEffectiveAddress(rs, offset);
} else {
SetCCAndR0ForSafeAddress(rs, offset, R1);
doCheck = true;
}
ADD(R0, R0, MEMBASEREG);
}
FixupBranch skip;
if (doCheck) {
skip = B_CC(CC_EQ);
}
VST1_lane(F_32, ar, R0, 0, true);
if (doCheck) {
SetJumpTarget(skip);
SetCC(CC_AL);
}
}
break;
}
fpr.ReleaseSpillLocksAndDiscardTemps();
}
inline int MIPS_GET_VQVT(u32 op) {
return (((op >> 16) & 0x1f)) | ((op & 1) << 5);
}
void ArmJit::CompNEON_SVQ(MIPSOpcode op) {
CONDITIONAL_DISABLE;
int offset = (signed short)(op & 0xFFFC);
int vt = MIPS_GET_VQVT(op.encoding);
MIPSGPReg rs = _RS;
bool doCheck = false;
switch (op >> 26)
{
case 54: //lv.q
{
// Check for four-in-a-row
const u32 ops[4] = {
op.encoding,
GetOffsetInstruction(1).encoding,
GetOffsetInstruction(2).encoding,
GetOffsetInstruction(3).encoding,
};
if (g_Config.bFastMemory && (ops[1] >> 26) == 54 && (ops[2] >> 26) == 54 && (ops[3] >> 26) == 54) {
int offsets[4] = {offset, (s16)(ops[1] & 0xFFFC), (s16)(ops[2] & 0xFFFC), (s16)(ops[3] & 0xFFFC)};
int rss[4] = {MIPS_GET_RS(op), MIPS_GET_RS(ops[1]), MIPS_GET_RS(ops[2]), MIPS_GET_RS(ops[3])};
if (offsets[1] == offset + 16 && offsets[2] == offsets[1] + 16 && offsets[3] == offsets[2] + 16 &&
rss[0] == rss[1] && rss[1] == rss[2] && rss[2] == rss[3]) {
int vts[4] = {MIPS_GET_VQVT(op.encoding), MIPS_GET_VQVT(ops[1]), MIPS_GET_VQVT(ops[2]), MIPS_GET_VQVT(ops[3])};
// TODO: Also check the destination registers!
// Detected four consecutive ones!
// gpr.MapRegAsPointer(rs);
// fpr.QLoad4x4(vts[4], rs, offset);
INFO_LOG(JIT, "Matrix load detected! TODO: optimize");
// break;
}
}
if (!gpr.IsImm(rs) && jo.cachePointers && g_Config.bFastMemory && offset < 0x400-16 && offset > -0x400-16) {
gpr.MapRegAsPointer(rs);
ARMReg ar = fpr.QMapReg(vt, V_Quad, MAP_DIRTY | MAP_NOINIT);
if (offset) {
ADDI2R(R0, gpr.RPtr(rs), offset, R1);
VLD1(F_32, ar, R0, 2, ALIGN_128);
} else {
VLD1(F_32, ar, gpr.RPtr(rs), 2, ALIGN_128);
}
break;
}
// CC might be set by slow path below, so load regs first.
ARMReg ar = fpr.QMapReg(vt, V_Quad, MAP_DIRTY | MAP_NOINIT);
if (gpr.IsImm(rs)) {
u32 addr = (offset + gpr.GetImm(rs)) & 0x3FFFFFFF;
gpr.SetRegImm(R0, addr + (u32)Memory::base);
} else {
gpr.MapReg(rs);
if (g_Config.bFastMemory) {
SetR0ToEffectiveAddress(rs, offset);
} else {
SetCCAndR0ForSafeAddress(rs, offset, R1);
doCheck = true;
}
ADD(R0, R0, MEMBASEREG);
}
FixupBranch skip;
if (doCheck) {
skip = B_CC(CC_EQ);
}
VLD1(F_32, ar, R0, 2, ALIGN_128);
if (doCheck) {
SetJumpTarget(skip);
SetCC(CC_AL);
}
}
break;
case 62: //sv.q
{
const u32 ops[4] = {
op.encoding,
GetOffsetInstruction(1).encoding,
GetOffsetInstruction(2).encoding,
GetOffsetInstruction(3).encoding,
};
if (g_Config.bFastMemory && (ops[1] >> 26) == 54 && (ops[2] >> 26) == 54 && (ops[3] >> 26) == 54) {
int offsets[4] = { offset, (s16)(ops[1] & 0xFFFC), (s16)(ops[2] & 0xFFFC), (s16)(ops[3] & 0xFFFC) };
int rss[4] = { MIPS_GET_RS(op), MIPS_GET_RS(ops[1]), MIPS_GET_RS(ops[2]), MIPS_GET_RS(ops[3]) };
if (offsets[1] == offset + 16 && offsets[2] == offsets[1] + 16 && offsets[3] == offsets[2] + 16 &&
rss[0] == rss[1] && rss[1] == rss[2] && rss[2] == rss[3]) {
int vts[4] = { MIPS_GET_VQVT(op.encoding), MIPS_GET_VQVT(ops[1]), MIPS_GET_VQVT(ops[2]), MIPS_GET_VQVT(ops[3]) };
// TODO: Also check the destination registers!
// Detected four consecutive ones!
// gpr.MapRegAsPointer(rs);
// fpr.QLoad4x4(vts[4], rs, offset);
INFO_LOG(JIT, "Matrix store detected! TODO: optimize");
// break;
}
}
if (!gpr.IsImm(rs) && jo.cachePointers && g_Config.bFastMemory && offset < 0x400-16 && offset > -0x400-16) {
gpr.MapRegAsPointer(rs);
ARMReg ar = fpr.QMapReg(vt, V_Quad, 0);
if (offset) {
ADDI2R(R0, gpr.RPtr(rs), offset, R1);
VST1(F_32, ar, R0, 2, ALIGN_128);
} else {
VST1(F_32, ar, gpr.RPtr(rs), 2, ALIGN_128);
}
break;
}
// CC might be set by slow path below, so load regs first.
ARMReg ar = fpr.QMapReg(vt, V_Quad, 0);
if (gpr.IsImm(rs)) {
u32 addr = (offset + gpr.GetImm(rs)) & 0x3FFFFFFF;
gpr.SetRegImm(R0, addr + (u32)Memory::base);
} else {
gpr.MapReg(rs);
if (g_Config.bFastMemory) {
SetR0ToEffectiveAddress(rs, offset);
} else {
SetCCAndR0ForSafeAddress(rs, offset, R1);
doCheck = true;
}
ADD(R0, R0, MEMBASEREG);
}
FixupBranch skip;
if (doCheck) {
skip = B_CC(CC_EQ);
}
VST1(F_32, ar, R0, 2, ALIGN_128);
if (doCheck) {
SetJumpTarget(skip);
SetCC(CC_AL);
}
}
break;
default:
DISABLE;
break;
}
fpr.ReleaseSpillLocksAndDiscardTemps();
}
void ArmJit::CompNEON_VVectorInit(MIPSOpcode op) {
CONDITIONAL_DISABLE;
// WARNING: No prefix support!
if (js.HasUnknownPrefix()) {
DISABLE_UNKNOWN_PREFIX;
}
VectorSize sz = GetVecSize(op);
DestARMReg vd = NEONMapPrefixD(_VD, sz, MAP_NOINIT | MAP_DIRTY);
switch ((op >> 16) & 0xF) {
case 6: // vzero
VEOR(vd.rd, vd.rd, vd.rd);
break;
case 7: // vone
VMOV_immf(vd.rd, 1.0f);
break;
default:
DISABLE;
break;
}
NEONApplyPrefixD(vd);
fpr.ReleaseSpillLocksAndDiscardTemps();
}
void ArmJit::CompNEON_VDot(MIPSOpcode op) {
CONDITIONAL_DISABLE;
if (js.HasUnknownPrefix()) {
DISABLE_UNKNOWN_PREFIX;
}
VectorSize sz = GetVecSize(op);
MappedRegs r = NEONMapDirtyInIn(op, V_Single, sz, sz);
switch (sz) {
case V_Pair:
VMUL(F_32, r.vd, r.vs, r.vt);
VPADD(F_32, r.vd, r.vd, r.vd);
break;
case V_Triple:
VMUL(F_32, Q0, r.vs, r.vt);
VPADD(F_32, D0, D0, D0);
VADD(F_32, r.vd, D0, D1);
break;
case V_Quad:
VMUL(F_32, D0, D_0(r.vs), D_0(r.vt));
VMLA(F_32, D0, D_1(r.vs), D_1(r.vt));
VPADD(F_32, r.vd, D0, D0);
break;
case V_Single:
case V_Invalid:
;
}
NEONApplyPrefixD(r.vd);
fpr.ReleaseSpillLocksAndDiscardTemps();
}
void ArmJit::CompNEON_VHdp(MIPSOpcode op) {
CONDITIONAL_DISABLE;
if (js.HasUnknownPrefix()) {
DISABLE_UNKNOWN_PREFIX;
}
DISABLE;
// Similar to VDot but the last component is only s instead of s * t.
// A bit tricky on NEON...
}
void ArmJit::CompNEON_VScl(MIPSOpcode op) {
CONDITIONAL_DISABLE;
if (js.HasUnknownPrefix()) {
DISABLE_UNKNOWN_PREFIX;
}
VectorSize sz = GetVecSize(op);
MappedRegs r = NEONMapDirtyInIn(op, sz, sz, V_Single);
ARMReg temp = MatchSize(Q0, r.vt);
// TODO: VMUL_scalar directly when possible
VMOV_neon(temp, r.vt);
VMUL_scalar(F_32, r.vd, r.vs, DScalar(Q0, 0));
NEONApplyPrefixD(r.vd);
fpr.ReleaseSpillLocksAndDiscardTemps();
}
void ArmJit::CompNEON_VV2Op(MIPSOpcode op) {
CONDITIONAL_DISABLE;
if (js.HasUnknownPrefix()) {
DISABLE_UNKNOWN_PREFIX;
}
// Pre-processing: Eliminate silly no-op VMOVs, common in Wipeout Pure
if (((op >> 16) & 0x1f) == 0 && _VS == _VD && js.HasNoPrefix()) {
return;
}
// Must bail before we start mapping registers.
switch ((op >> 16) & 0x1f) {
case 0: // d[i] = s[i]; break; //vmov
case 1: // d[i] = fabsf(s[i]); break; //vabs
case 2: // d[i] = -s[i]; break; //vneg
case 17: // d[i] = 1.0f / sqrtf(s[i]); break; //vrsq
break;
default:
DISABLE;
break;
}
VectorSize sz = GetVecSize(op);
int n = GetNumVectorElements(sz);
MappedRegs r = NEONMapDirtyIn(op, sz, sz);
ARMReg temp = MatchSize(Q0, r.vs);
switch ((op >> 16) & 0x1f) {
case 0: // d[i] = s[i]; break; //vmov
// Probably for swizzle.
VMOV_neon(r.vd, r.vs);
break;
case 1: // d[i] = fabsf(s[i]); break; //vabs
VABS(F_32, r.vd, r.vs);
break;
case 2: // d[i] = -s[i]; break; //vneg
VNEG(F_32, r.vd, r.vs);
break;
case 4: // if (s[i] < 0) d[i] = 0; else {if(s[i] > 1.0f) d[i] = 1.0f; else d[i] = s[i];} break; // vsat0
if (IsD(r.vd)) {
VMOV_immf(D0, 0.0f);
VMOV_immf(D1, 1.0f);
VMAX(F_32, r.vd, r.vs, D0);
VMIN(F_32, r.vd, r.vd, D1);
} else {
VMOV_immf(Q0, 1.0f);
VMIN(F_32, r.vd, r.vs, Q0);
VMOV_immf(Q0, 0.0f);
VMAX(F_32, r.vd, r.vd, Q0);
}
break;
case 5: // if (s[i] < -1.0f) d[i] = -1.0f; else {if(s[i] > 1.0f) d[i] = 1.0f; else d[i] = s[i];} break; // vsat1
if (IsD(r.vd)) {
VMOV_immf(D0, -1.0f);
VMOV_immf(D1, 1.0f);
VMAX(F_32, r.vd, r.vs, D0);
VMIN(F_32, r.vd, r.vd, D1);
} else {
VMOV_immf(Q0, 1.0f);
VMIN(F_32, r.vd, r.vs, Q0);
VMOV_immf(Q0, -1.0f);
VMAX(F_32, r.vd, r.vd, Q0);
}
break;
case 16: // d[i] = 1.0f / s[i]; break; //vrcp
// Can just fallback to VFP and use VDIV.
DISABLE;
{
ARMReg temp2 = fpr.QAllocTemp(sz);
// Needs iterations on NEON. And two temps - which is a problem if vs == vd! Argh!
VRECPE(F_32, temp, r.vs);
VRECPS(temp2, r.vs, temp);
VMUL(F_32, temp2, temp2, temp);
VRECPS(temp2, r.vs, temp);
VMUL(F_32, temp2, temp2, temp);
}
// http://stackoverflow.com/questions/6759897/how-to-divide-in-neon-intrinsics-by-a-float-number
// reciprocal = vrecpeq_f32(b);
// reciprocal = vmulq_f32(vrecpsq_f32(b, reciprocal), reciprocal);
// reciprocal = vmulq_f32(vrecpsq_f32(b, reciprocal), reciprocal);
DISABLE;
break;
case 17: // d[i] = 1.0f / sqrtf(s[i]); break; //vrsq
DISABLE;
// Needs iterations on NEON
{
if (true) {
// Not-very-accurate estimate
VRSQRTE(F_32, r.vd, r.vs);
} else {
ARMReg temp2 = fpr.QAllocTemp(sz);
// TODO: It's likely that some games will require one or two Newton-Raphson
// iterations to refine the estimate.
VRSQRTE(F_32, temp, r.vs);
VRSQRTS(temp2, r.vs, temp);
VMUL(F_32, r.vd, temp2, temp);
//VRSQRTS(temp2, r.vs, temp);
// VMUL(F_32, r.vd, temp2, temp);
}
}
break;
case 18: // d[i] = sinf((float)M_PI_2 * s[i]); break; //vsin
DISABLE;
break;
case 19: // d[i] = cosf((float)M_PI_2 * s[i]); break; //vcos
DISABLE;
break;
case 20: // d[i] = powf(2.0f, s[i]); break; //vexp2
DISABLE;
break;
case 21: // d[i] = logf(s[i])/log(2.0f); break; //vlog2
DISABLE;
break;
case 22: // d[i] = sqrtf(s[i]); break; //vsqrt
// Let's just defer to VFP for now. Better than calling the interpreter for sure.
VMOV_neon(MatchSize(Q0, r.vs), r.vs);
for (int i = 0; i < n; i++) {
VSQRT((ARMReg)(S0 + i), (ARMReg)(S0 + i));
}
VMOV_neon(MatchSize(Q0, r.vd), r.vd);
break;
case 23: // d[i] = asinf(s[i] * (float)M_2_PI); break; //vasin
DISABLE;
break;
case 24: // d[i] = -1.0f / s[i]; break; // vnrcp
// Needs iterations on NEON. Just do the same as vrcp and negate.
DISABLE;
break;
case 26: // d[i] = -sinf((float)M_PI_2 * s[i]); break; // vnsin
DISABLE;
break;
case 28: // d[i] = 1.0f / expf(s[i] * (float)M_LOG2E); break; // vrexp2
DISABLE;
break;
default:
DISABLE;
break;
}
NEONApplyPrefixD(r.vd);
fpr.ReleaseSpillLocksAndDiscardTemps();
}
void ArmJit::CompNEON_Mftv(MIPSOpcode op) {
CONDITIONAL_DISABLE;
int imm = op & 0xFF;
MIPSGPReg rt = _RT;
switch ((op >> 21) & 0x1f) {
case 3: //mfv / mfvc
// rt = 0, imm = 255 appears to be used as a CPU interlock by some games.
if (rt != 0) {
if (imm < 128) { //R(rt) = VI(imm);
ARMReg r = fpr.QMapReg(imm, V_Single, MAP_READ);
gpr.MapReg(rt, MAP_NOINIT | MAP_DIRTY);
// TODO: Gotta be a faster way
VMOV_neon(MatchSize(Q0, r), r);
VMOV(gpr.R(rt), S0);
} else if (imm < 128 + VFPU_CTRL_MAX) { //mtvc
// In case we have a saved prefix.
FlushPrefixV();
if (imm - 128 == VFPU_CTRL_CC) {
gpr.MapDirtyIn(rt, MIPS_REG_VFPUCC);
MOV(gpr.R(rt), gpr.R(MIPS_REG_VFPUCC));
} else {
gpr.MapReg(rt, MAP_NOINIT | MAP_DIRTY);
LDR(gpr.R(rt), CTXREG, offsetof(MIPSState, vfpuCtrl) + 4 * (imm - 128));
}
} else {
//ERROR - maybe need to make this value too an "interlock" value?
ERROR_LOG(CPU, "mfv - invalid register %i", imm);
}
}
break;
case 7: // mtv
if (imm < 128) {
// TODO: It's pretty common that this is preceded by mfc1, that is, a value is being
// moved from the regular floating point registers. It would probably be faster to do
// the copy directly in the FPRs instead of going through the GPRs.
ARMReg r = fpr.QMapReg(imm, V_Single, MAP_DIRTY | MAP_NOINIT);
if (gpr.IsMapped(rt)) {
VMOV(S0, gpr.R(rt));
VMOV_neon(r, MatchSize(Q0, r));
} else {
ADDI2R(R0, CTXREG, gpr.GetMipsRegOffset(rt), R1);
VLD1_lane(F_32, r, R0, 0, true);
}
} else if (imm < 128 + VFPU_CTRL_MAX) { //mtvc //currentMIPS->vfpuCtrl[imm - 128] = R(rt);
if (imm - 128 == VFPU_CTRL_CC) {
gpr.MapDirtyIn(MIPS_REG_VFPUCC, rt);
MOV(gpr.R(MIPS_REG_VFPUCC), rt);
} else {
gpr.MapReg(rt);
STR(gpr.R(rt), CTXREG, offsetof(MIPSState, vfpuCtrl) + 4 * (imm - 128));
}
// TODO: Optimization if rt is Imm?
// Set these BEFORE disable!
if (imm - 128 == VFPU_CTRL_SPREFIX) {
js.prefixSFlag = JitState::PREFIX_UNKNOWN;
} else if (imm - 128 == VFPU_CTRL_TPREFIX) {
js.prefixTFlag = JitState::PREFIX_UNKNOWN;
} else if (imm - 128 == VFPU_CTRL_DPREFIX) {
js.prefixDFlag = JitState::PREFIX_UNKNOWN;
}
} else {
//ERROR
_dbg_assert_msg_(CPU,0,"mtv - invalid register");
}
break;
default:
DISABLE;
}
fpr.ReleaseSpillLocksAndDiscardTemps();
}
void ArmJit::CompNEON_Vmfvc(MIPSOpcode op) {
DISABLE;
}
void ArmJit::CompNEON_Vmtvc(MIPSOpcode op) {
CONDITIONAL_DISABLE;
int vs = _VS;
int imm = op & 0xFF;
if (imm >= 128 && imm < 128 + VFPU_CTRL_MAX) {
ARMReg r = fpr.QMapReg(vs, V_Single, 0);
ADDI2R(R0, CTXREG, offsetof(MIPSState, vfpuCtrl[0]) + (imm - 128) * 4, R1);
VST1_lane(F_32, r, R0, 0, true);
fpr.ReleaseSpillLocksAndDiscardTemps();
if (imm - 128 == VFPU_CTRL_SPREFIX) {
js.prefixSFlag = JitState::PREFIX_UNKNOWN;
} else if (imm - 128 == VFPU_CTRL_TPREFIX) {
js.prefixTFlag = JitState::PREFIX_UNKNOWN;
} else if (imm - 128 == VFPU_CTRL_DPREFIX) {
js.prefixDFlag = JitState::PREFIX_UNKNOWN;
}
}
}
void ArmJit::CompNEON_VMatrixInit(MIPSOpcode op) {
CONDITIONAL_DISABLE;
MatrixSize msz = GetMtxSize(op);
int n = GetMatrixSide(msz);
ARMReg cols[4];
fpr.QMapMatrix(cols, _VD, msz, MAP_NOINIT | MAP_DIRTY);
switch ((op >> 16) & 0xF) {
case 3: // vmidt
// There has to be a better way to synthesize: 1.0, 0.0, 0.0, 1.0 in a quad
VEOR(D0, D0, D0);
VMOV_immf(D1, 1.0f);
VTRN(F_32, D0, D1);
VREV64(I_32, D0, D0);
switch (msz) {
case M_2x2:
VMOV_neon(cols[0], D0);
VMOV_neon(cols[1], D1);
break;
case M_3x3:
VMOV_neon(D_0(cols[0]), D0);
VMOV_imm(I_8, D_1(cols[0]), VIMMxxxxxxxx, 0);
VMOV_neon(D_0(cols[1]), D1);
VMOV_imm(I_8, D_1(cols[1]), VIMMxxxxxxxx, 0);
VMOV_imm(I_8, D_0(cols[2]), VIMMxxxxxxxx, 0);
VMOV_neon(D_1(cols[2]), D0);
break;
case M_4x4:
VMOV_neon(D_0(cols[0]), D0);
VMOV_imm(I_8, D_1(cols[0]), VIMMxxxxxxxx, 0);
VMOV_neon(D_0(cols[1]), D1);
VMOV_imm(I_8, D_1(cols[1]), VIMMxxxxxxxx, 0);
VMOV_imm(I_8, D_0(cols[2]), VIMMxxxxxxxx, 0);
VMOV_neon(D_1(cols[2]), D0);
VMOV_imm(I_8, D_0(cols[3]), VIMMxxxxxxxx, 0);
VMOV_neon(D_1(cols[3]), D1);
// NEONTranspose4x4(cols);
break;
default:
_assert_msg_(JIT, 0, "Bad matrix size");
break;
}
break;
case 6: // vmzero
for (int i = 0; i < n; i++) {
VEOR(cols[i], cols[i], cols[i]);
}
break;
case 7: // vmone
for (int i = 0; i < n; i++) {
VMOV_immf(cols[i], 1.0f);
}
break;
}
fpr.ReleaseSpillLocksAndDiscardTemps();
}
void ArmJit::CompNEON_Vmmov(MIPSOpcode op) {
CONDITIONAL_DISABLE;
if (_VS == _VD) {
// A lot of these no-op matrix moves in Wipeout... Just drop the instruction entirely.
return;
}
MatrixSize msz = GetMtxSize(op);
MatrixOverlapType overlap = GetMatrixOverlap(_VD, _VS, msz);
if (overlap != OVERLAP_NONE) {
// Too complicated to bother handling in the JIT.
// TODO: Special case for in-place (and other) transpose, etc.
DISABLE;
}
ARMReg s_cols[4], d_cols[4];
fpr.QMapMatrix(s_cols, _VS, msz, 0);
fpr.QMapMatrix(d_cols, _VD, msz, MAP_DIRTY | MAP_NOINIT);
int n = GetMatrixSide(msz);
for (int i = 0; i < n; i++) {
VMOV_neon(d_cols[i], s_cols[i]);
}
fpr.ReleaseSpillLocksAndDiscardTemps();
}
void ArmJit::CompNEON_Vmmul(MIPSOpcode op) {
CONDITIONAL_DISABLE;
MatrixSize msz = GetMtxSize(op);
int n = GetMatrixSide(msz);
bool overlap = GetMatrixOverlap(_VD, _VS, msz) || GetMatrixOverlap(_VD, _VT, msz);
if (overlap) {
// Later. Fortunately, the VFPU also seems to prohibit overlap for matrix mul.
ILOG("Matrix overlap, ignoring.");
DISABLE;
}
// Having problems with 2x2s for some reason.
if (msz == M_2x2) {
DISABLE;
}
ARMReg s_cols[4], t_cols[4], d_cols[4];
// For some reason, vmmul is encoded with the first matrix (S) transposed from the real meaning.
fpr.QMapMatrix(t_cols, _VT, msz, MAP_FORCE_LOW); // Need to see if we can avoid having to force it low in some sane way. Will need crazy prediction logic for loads otherwise.
fpr.QMapMatrix(s_cols, Xpose(_VS), msz, MAP_PREFER_HIGH);
fpr.QMapMatrix(d_cols, _VD, msz, MAP_PREFER_HIGH | MAP_NOINIT | MAP_DIRTY);
// TODO: Getting there but still getting wrong results.
for (int i = 0; i < n; i++) {
for (int j = 0; j < n; j++) {
if (i == 0) {
VMUL_scalar(F_32, d_cols[j], s_cols[i], XScalar(t_cols[j], i));
} else {
VMLA_scalar(F_32, d_cols[j], s_cols[i], XScalar(t_cols[j], i));
}
}
}
fpr.ReleaseSpillLocksAndDiscardTemps();
}
void ArmJit::CompNEON_Vmscl(MIPSOpcode op) {
CONDITIONAL_DISABLE;
MatrixSize msz = GetMtxSize(op);
bool overlap = GetMatrixOverlap(_VD, _VS, msz) != OVERLAP_NONE;
if (overlap) {
DISABLE;
}
int n = GetMatrixSide(msz);
ARMReg s_cols[4], t, d_cols[4];
fpr.QMapMatrix(s_cols, _VS, msz, 0);
fpr.QMapMatrix(d_cols, _VD, msz, MAP_NOINIT | MAP_DIRTY);
t = fpr.QMapReg(_VT, V_Single, 0);
VMOV_neon(D0, t);
for (int i = 0; i < n; i++) {
VMUL_scalar(F_32, d_cols[i], s_cols[i], DScalar(D0, 0));
}
fpr.ReleaseSpillLocksAndDiscardTemps();
}
void ArmJit::CompNEON_Vtfm(MIPSOpcode op) {
CONDITIONAL_DISABLE;
if (js.HasUnknownPrefix()) {
DISABLE;
}
if (_VT == _VD) {
DISABLE;
}
VectorSize sz = GetVecSize(op);
MatrixSize msz = GetMtxSize(op);
int n = GetNumVectorElements(sz);
int ins = (op >> 23) & 7;
bool homogenous = false;
if (n == ins) {
n++;
sz = (VectorSize)((int)(sz)+1);
msz = (MatrixSize)((int)(msz)+1);
homogenous = true;
}
// Otherwise, n should already be ins + 1.
else if (n != ins + 1) {
DISABLE;
}
ARMReg s_cols[4], t, d;
t = fpr.QMapReg(_VT, sz, MAP_FORCE_LOW);
fpr.QMapMatrix(s_cols, Xpose(_VS), msz, MAP_PREFER_HIGH);
d = fpr.QMapReg(_VD, sz, MAP_DIRTY | MAP_NOINIT | MAP_PREFER_HIGH);
VMUL_scalar(F_32, d, s_cols[0], XScalar(t, 0));
for (int i = 1; i < n; i++) {
if (homogenous && i == n - 1) {
VADD(F_32, d, d, s_cols[i]);
} else {
VMLA_scalar(F_32, d, s_cols[i], XScalar(t, i));
}
}
// VTFM does not have prefix support.
fpr.ReleaseSpillLocksAndDiscardTemps();
}
void ArmJit::CompNEON_VCrs(MIPSOpcode op) {
DISABLE;
}
void ArmJit::CompNEON_VDet(MIPSOpcode op) {
DISABLE;
}
void ArmJit::CompNEON_Vi2x(MIPSOpcode op) {
DISABLE;
}
void ArmJit::CompNEON_Vx2i(MIPSOpcode op) {
DISABLE;
}
void ArmJit::CompNEON_Vf2i(MIPSOpcode op) {
DISABLE;
}
void ArmJit::CompNEON_Vi2f(MIPSOpcode op) {
CONDITIONAL_DISABLE;
if (js.HasUnknownPrefix()) {
DISABLE;
}
DISABLE;
VectorSize sz = GetVecSize(op);
int n = GetNumVectorElements(sz);
int imm = (op >> 16) & 0x1f;
const float mult = 1.0f / (float)(1UL << imm);
MappedRegs regs = NEONMapDirtyIn(op, sz, sz);