2
* ARM NEON vector operations.
4
* Copyright (c) 2007, 2008 CodeSourcery.
5
* Written by Paul Brook
7
* This code is licensed under the GNU GPL v2.
9
#include "qemu/osdep.h"
12
#include "exec/exec-all.h"
13
#include "exec/helper-proto.h"
15
#define SIGNBIT (uint32_t)0x80000000
16
#define SIGNBIT64 ((uint64_t)1 << 63)
18
#define SET_QC() env->vfp.xregs[ARM_VFP_FPSCR] |= CPSR_Q
20
#define NEON_TYPE1(name, type) \
25
#ifdef HOST_WORDS_BIGENDIAN
26
#define NEON_TYPE2(name, type) \
32
#define NEON_TYPE4(name, type) \
41
#define NEON_TYPE2(name, type) \
47
#define NEON_TYPE4(name, type) \
57
NEON_TYPE4(s8, int8_t)
58
NEON_TYPE4(u8, uint8_t)
59
NEON_TYPE2(s16, int16_t)
60
NEON_TYPE2(u16, uint16_t)
61
NEON_TYPE1(s32, int32_t)
62
NEON_TYPE1(u32, uint32_t)
67
/* Copy from a uint32_t to a vector structure type. */
68
#define NEON_UNPACK(vtype, dest, val) do { \
77
/* Copy from a vector structure type to a uint32_t. */
78
#define NEON_PACK(vtype, dest, val) do { \
88
NEON_FN(vdest.v1, vsrc1.v1, vsrc2.v1);
90
NEON_FN(vdest.v1, vsrc1.v1, vsrc2.v1); \
91
NEON_FN(vdest.v2, vsrc1.v2, vsrc2.v2);
93
NEON_FN(vdest.v1, vsrc1.v1, vsrc2.v1); \
94
NEON_FN(vdest.v2, vsrc1.v2, vsrc2.v2); \
95
NEON_FN(vdest.v3, vsrc1.v3, vsrc2.v3); \
96
NEON_FN(vdest.v4, vsrc1.v4, vsrc2.v4);
98
#define NEON_VOP_BODY(vtype, n) \
104
NEON_UNPACK(vtype, vsrc1, arg1); \
105
NEON_UNPACK(vtype, vsrc2, arg2); \
107
NEON_PACK(vtype, res, vdest); \
111
#define NEON_VOP(name, vtype, n) \
112
uint32_t HELPER(glue(neon_,name))(uint32_t arg1, uint32_t arg2) \
113
NEON_VOP_BODY(vtype, n)
115
#define NEON_VOP_ENV(name, vtype, n) \
116
uint32_t HELPER(glue(neon_,name))(CPUARMState *env, uint32_t arg1, uint32_t arg2) \
117
NEON_VOP_BODY(vtype, n)
119
/* Pairwise operations. */
120
/* For 32-bit elements each segment only contains a single element, so
121
the elementwise and pairwise operations are the same. */
123
NEON_FN(vdest.v1, vsrc1.v1, vsrc1.v2); \
124
NEON_FN(vdest.v2, vsrc2.v1, vsrc2.v2);
126
NEON_FN(vdest.v1, vsrc1.v1, vsrc1.v2); \
127
NEON_FN(vdest.v2, vsrc1.v3, vsrc1.v4); \
128
NEON_FN(vdest.v3, vsrc2.v1, vsrc2.v2); \
129
NEON_FN(vdest.v4, vsrc2.v3, vsrc2.v4); \
131
#define NEON_POP(name, vtype, n) \
132
uint32_t HELPER(glue(neon_,name))(uint32_t arg1, uint32_t arg2) \
138
NEON_UNPACK(vtype, vsrc1, arg1); \
139
NEON_UNPACK(vtype, vsrc2, arg2); \
141
NEON_PACK(vtype, res, vdest); \
145
/* Unary operators. */
146
#define NEON_VOP1(name, vtype, n) \
147
uint32_t HELPER(glue(neon_,name))(uint32_t arg) \
151
NEON_UNPACK(vtype, vsrc1, arg); \
153
NEON_PACK(vtype, arg, vdest); \
158
#define NEON_USAT(dest, src1, src2, type) do { \
159
uint32_t tmp = (uint32_t)src1 + (uint32_t)src2; \
160
if (tmp != (type)tmp) { \
166
#define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint8_t)
167
NEON_VOP_ENV(qadd_u8, neon_u8, 4)
169
#define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint16_t)
170
NEON_VOP_ENV(qadd_u16, neon_u16, 2)
174
uint32_t HELPER(neon_qadd_u32)(CPUARMState *env, uint32_t a, uint32_t b)
176
uint32_t res = a + b;
184
uint64_t HELPER(neon_qadd_u64)(CPUARMState *env, uint64_t src1, uint64_t src2)
196
#define NEON_SSAT(dest, src1, src2, type) do { \
197
int32_t tmp = (uint32_t)src1 + (uint32_t)src2; \
198
if (tmp != (type)tmp) { \
201
tmp = (1 << (sizeof(type) * 8 - 1)) - 1; \
203
tmp = 1 << (sizeof(type) * 8 - 1); \
208
#define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int8_t)
209
NEON_VOP_ENV(qadd_s8, neon_s8, 4)
211
#define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int16_t)
212
NEON_VOP_ENV(qadd_s16, neon_s16, 2)
216
uint32_t HELPER(neon_qadd_s32)(CPUARMState *env, uint32_t a, uint32_t b)
218
uint32_t res = a + b;
219
if (((res ^ a) & SIGNBIT) && !((a ^ b) & SIGNBIT)) {
221
res = ~(((int32_t)a >> 31) ^ SIGNBIT);
226
uint64_t HELPER(neon_qadd_s64)(CPUARMState *env, uint64_t src1, uint64_t src2)
231
if (((res ^ src1) & SIGNBIT64) && !((src1 ^ src2) & SIGNBIT64)) {
233
res = ((int64_t)src1 >> 63) ^ ~SIGNBIT64;
238
/* Unsigned saturating accumulate of signed value
240
* Op1/Rn is treated as signed
241
* Op2/Rd is treated as unsigned
243
* Explicit casting is used to ensure the correct sign extension of
244
* inputs. The result is treated as a unsigned value and saturated as such.
246
* We use a macro for the 8/16 bit cases which expects signed integers of va,
247
* vb, and vr for interim calculation and an unsigned 32 bit result value r.
250
#define USATACC(bits, shift) \
252
va = sextract32(a, shift, bits); \
253
vb = extract32(b, shift, bits); \
255
if (vr > UINT##bits##_MAX) { \
257
vr = UINT##bits##_MAX; \
258
} else if (vr < 0) { \
262
r = deposit32(r, shift, bits, vr); \
265
uint32_t HELPER(neon_uqadd_s8)(CPUARMState *env, uint32_t a, uint32_t b)
277
uint32_t HELPER(neon_uqadd_s16)(CPUARMState *env, uint32_t a, uint32_t b)
289
uint32_t HELPER(neon_uqadd_s32)(CPUARMState *env, uint32_t a, uint32_t b)
291
int64_t va = (int32_t)a;
292
int64_t vb = (uint32_t)b;
293
int64_t vr = va + vb;
294
if (vr > UINT32_MAX) {
304
uint64_t HELPER(neon_uqadd_s64)(CPUARMState *env, uint64_t a, uint64_t b)
308
/* We only need to look at the pattern of SIGN bits to detect
311
if (~a & b & ~res & SIGNBIT64) {
314
} else if (a & ~b & res & SIGNBIT64) {
321
/* Signed saturating accumulate of unsigned value
323
* Op1/Rn is treated as unsigned
324
* Op2/Rd is treated as signed
326
* The result is treated as a signed value and saturated as such
328
* We use a macro for the 8/16 bit cases which expects signed integers of va,
329
* vb, and vr for interim calculation and an unsigned 32 bit result value r.
332
#define SSATACC(bits, shift) \
334
va = extract32(a, shift, bits); \
335
vb = sextract32(b, shift, bits); \
337
if (vr > INT##bits##_MAX) { \
339
vr = INT##bits##_MAX; \
340
} else if (vr < INT##bits##_MIN) { \
342
vr = INT##bits##_MIN; \
344
r = deposit32(r, shift, bits, vr); \
347
uint32_t HELPER(neon_sqadd_u8)(CPUARMState *env, uint32_t a, uint32_t b)
359
uint32_t HELPER(neon_sqadd_u16)(CPUARMState *env, uint32_t a, uint32_t b)
372
uint32_t HELPER(neon_sqadd_u32)(CPUARMState *env, uint32_t a, uint32_t b)
375
int64_t op1 = (uint32_t)a;
376
int64_t op2 = (int32_t)b;
378
if (res > INT32_MAX) {
381
} else if (res < INT32_MIN) {
388
uint64_t HELPER(neon_sqadd_u64)(CPUARMState *env, uint64_t a, uint64_t b)
392
/* We only need to look at the pattern of SIGN bits to detect an overflow */
395
| (a & ~b)) & SIGNBIT64) {
403
#define NEON_USAT(dest, src1, src2, type) do { \
404
uint32_t tmp = (uint32_t)src1 - (uint32_t)src2; \
405
if (tmp != (type)tmp) { \
411
#define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint8_t)
412
NEON_VOP_ENV(qsub_u8, neon_u8, 4)
414
#define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint16_t)
415
NEON_VOP_ENV(qsub_u16, neon_u16, 2)
419
uint32_t HELPER(neon_qsub_u32)(CPUARMState *env, uint32_t a, uint32_t b)
421
uint32_t res = a - b;
429
uint64_t HELPER(neon_qsub_u64)(CPUARMState *env, uint64_t src1, uint64_t src2)
442
#define NEON_SSAT(dest, src1, src2, type) do { \
443
int32_t tmp = (uint32_t)src1 - (uint32_t)src2; \
444
if (tmp != (type)tmp) { \
447
tmp = (1 << (sizeof(type) * 8 - 1)) - 1; \
449
tmp = 1 << (sizeof(type) * 8 - 1); \
454
#define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int8_t)
455
NEON_VOP_ENV(qsub_s8, neon_s8, 4)
457
#define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int16_t)
458
NEON_VOP_ENV(qsub_s16, neon_s16, 2)
462
uint32_t HELPER(neon_qsub_s32)(CPUARMState *env, uint32_t a, uint32_t b)
464
uint32_t res = a - b;
465
if (((res ^ a) & SIGNBIT) && ((a ^ b) & SIGNBIT)) {
467
res = ~(((int32_t)a >> 31) ^ SIGNBIT);
472
uint64_t HELPER(neon_qsub_s64)(CPUARMState *env, uint64_t src1, uint64_t src2)
477
if (((res ^ src1) & SIGNBIT64) && ((src1 ^ src2) & SIGNBIT64)) {
479
res = ((int64_t)src1 >> 63) ^ ~SIGNBIT64;
484
#define NEON_FN(dest, src1, src2) dest = (src1 + src2) >> 1
485
NEON_VOP(hadd_s8, neon_s8, 4)
486
NEON_VOP(hadd_u8, neon_u8, 4)
487
NEON_VOP(hadd_s16, neon_s16, 2)
488
NEON_VOP(hadd_u16, neon_u16, 2)
491
int32_t HELPER(neon_hadd_s32)(int32_t src1, int32_t src2)
495
dest = (src1 >> 1) + (src2 >> 1);
501
uint32_t HELPER(neon_hadd_u32)(uint32_t src1, uint32_t src2)
505
dest = (src1 >> 1) + (src2 >> 1);
511
#define NEON_FN(dest, src1, src2) dest = (src1 + src2 + 1) >> 1
512
NEON_VOP(rhadd_s8, neon_s8, 4)
513
NEON_VOP(rhadd_u8, neon_u8, 4)
514
NEON_VOP(rhadd_s16, neon_s16, 2)
515
NEON_VOP(rhadd_u16, neon_u16, 2)
518
int32_t HELPER(neon_rhadd_s32)(int32_t src1, int32_t src2)
522
dest = (src1 >> 1) + (src2 >> 1);
523
if ((src1 | src2) & 1)
528
uint32_t HELPER(neon_rhadd_u32)(uint32_t src1, uint32_t src2)
532
dest = (src1 >> 1) + (src2 >> 1);
533
if ((src1 | src2) & 1)
538
#define NEON_FN(dest, src1, src2) dest = (src1 - src2) >> 1
539
NEON_VOP(hsub_s8, neon_s8, 4)
540
NEON_VOP(hsub_u8, neon_u8, 4)
541
NEON_VOP(hsub_s16, neon_s16, 2)
542
NEON_VOP(hsub_u16, neon_u16, 2)
545
int32_t HELPER(neon_hsub_s32)(int32_t src1, int32_t src2)
549
dest = (src1 >> 1) - (src2 >> 1);
550
if ((~src1) & src2 & 1)
555
uint32_t HELPER(neon_hsub_u32)(uint32_t src1, uint32_t src2)
559
dest = (src1 >> 1) - (src2 >> 1);
560
if ((~src1) & src2 & 1)
565
#define NEON_FN(dest, src1, src2) dest = (src1 > src2) ? ~0 : 0
566
NEON_VOP(cgt_s8, neon_s8, 4)
567
NEON_VOP(cgt_u8, neon_u8, 4)
568
NEON_VOP(cgt_s16, neon_s16, 2)
569
NEON_VOP(cgt_u16, neon_u16, 2)
570
NEON_VOP(cgt_s32, neon_s32, 1)
571
NEON_VOP(cgt_u32, neon_u32, 1)
574
#define NEON_FN(dest, src1, src2) dest = (src1 >= src2) ? ~0 : 0
575
NEON_VOP(cge_s8, neon_s8, 4)
576
NEON_VOP(cge_u8, neon_u8, 4)
577
NEON_VOP(cge_s16, neon_s16, 2)
578
NEON_VOP(cge_u16, neon_u16, 2)
579
NEON_VOP(cge_s32, neon_s32, 1)
580
NEON_VOP(cge_u32, neon_u32, 1)
583
#define NEON_FN(dest, src1, src2) dest = (src1 < src2) ? src1 : src2
584
NEON_VOP(min_s8, neon_s8, 4)
585
NEON_VOP(min_u8, neon_u8, 4)
586
NEON_VOP(min_s16, neon_s16, 2)
587
NEON_VOP(min_u16, neon_u16, 2)
588
NEON_VOP(min_s32, neon_s32, 1)
589
NEON_VOP(min_u32, neon_u32, 1)
590
NEON_POP(pmin_s8, neon_s8, 4)
591
NEON_POP(pmin_u8, neon_u8, 4)
592
NEON_POP(pmin_s16, neon_s16, 2)
593
NEON_POP(pmin_u16, neon_u16, 2)
596
#define NEON_FN(dest, src1, src2) dest = (src1 > src2) ? src1 : src2
597
NEON_VOP(max_s8, neon_s8, 4)
598
NEON_VOP(max_u8, neon_u8, 4)
599
NEON_VOP(max_s16, neon_s16, 2)
600
NEON_VOP(max_u16, neon_u16, 2)
601
NEON_VOP(max_s32, neon_s32, 1)
602
NEON_VOP(max_u32, neon_u32, 1)
603
NEON_POP(pmax_s8, neon_s8, 4)
604
NEON_POP(pmax_u8, neon_u8, 4)
605
NEON_POP(pmax_s16, neon_s16, 2)
606
NEON_POP(pmax_u16, neon_u16, 2)
609
#define NEON_FN(dest, src1, src2) \
610
dest = (src1 > src2) ? (src1 - src2) : (src2 - src1)
611
NEON_VOP(abd_s8, neon_s8, 4)
612
NEON_VOP(abd_u8, neon_u8, 4)
613
NEON_VOP(abd_s16, neon_s16, 2)
614
NEON_VOP(abd_u16, neon_u16, 2)
615
NEON_VOP(abd_s32, neon_s32, 1)
616
NEON_VOP(abd_u32, neon_u32, 1)
619
#define NEON_FN(dest, src1, src2) do { \
621
tmp = (int8_t)src2; \
622
if (tmp >= (ssize_t)sizeof(src1) * 8 || \
623
tmp <= -(ssize_t)sizeof(src1) * 8) { \
625
} else if (tmp < 0) { \
626
dest = src1 >> -tmp; \
628
dest = src1 << tmp; \
630
NEON_VOP(shl_u8, neon_u8, 4)
631
NEON_VOP(shl_u16, neon_u16, 2)
632
NEON_VOP(shl_u32, neon_u32, 1)
635
uint64_t HELPER(neon_shl_u64)(uint64_t val, uint64_t shiftop)
637
int8_t shift = (int8_t)shiftop;
638
if (shift >= 64 || shift <= -64) {
640
} else if (shift < 0) {
648
#define NEON_FN(dest, src1, src2) do { \
650
tmp = (int8_t)src2; \
651
if (tmp >= (ssize_t)sizeof(src1) * 8) { \
653
} else if (tmp <= -(ssize_t)sizeof(src1) * 8) { \
654
dest = src1 >> (sizeof(src1) * 8 - 1); \
655
} else if (tmp < 0) { \
656
dest = src1 >> -tmp; \
658
dest = src1 << tmp; \
660
NEON_VOP(shl_s8, neon_s8, 4)
661
NEON_VOP(shl_s16, neon_s16, 2)
662
NEON_VOP(shl_s32, neon_s32, 1)
665
uint64_t HELPER(neon_shl_s64)(uint64_t valop, uint64_t shiftop)
667
int8_t shift = (int8_t)shiftop;
671
} else if (shift <= -64) {
673
} else if (shift < 0) {
681
#define NEON_FN(dest, src1, src2) do { \
683
tmp = (int8_t)src2; \
684
if ((tmp >= (ssize_t)sizeof(src1) * 8) \
685
|| (tmp <= -(ssize_t)sizeof(src1) * 8)) { \
687
} else if (tmp < 0) { \
688
dest = (src1 + (1 << (-1 - tmp))) >> -tmp; \
690
dest = src1 << tmp; \
692
NEON_VOP(rshl_s8, neon_s8, 4)
693
NEON_VOP(rshl_s16, neon_s16, 2)
696
/* The addition of the rounding constant may overflow, so we use an
697
* intermediate 64 bit accumulator. */
698
uint32_t HELPER(neon_rshl_s32)(uint32_t valop, uint32_t shiftop)
701
int32_t val = (int32_t)valop;
702
int8_t shift = (int8_t)shiftop;
703
if ((shift >= 32) || (shift <= -32)) {
705
} else if (shift < 0) {
706
int64_t big_dest = ((int64_t)val + (1 << (-1 - shift)));
707
dest = big_dest >> -shift;
714
/* Handling addition overflow with 64 bit input values is more
715
* tricky than with 32 bit values. */
716
uint64_t HELPER(neon_rshl_s64)(uint64_t valop, uint64_t shiftop)
718
int8_t shift = (int8_t)shiftop;
720
if ((shift >= 64) || (shift <= -64)) {
722
} else if (shift < 0) {
723
val >>= (-shift - 1);
724
if (val == INT64_MAX) {
725
/* In this case, it means that the rounding constant is 1,
726
* and the addition would overflow. Return the actual
727
* result directly. */
728
val = 0x4000000000000000LL;
739
#define NEON_FN(dest, src1, src2) do { \
741
tmp = (int8_t)src2; \
742
if (tmp >= (ssize_t)sizeof(src1) * 8 || \
743
tmp < -(ssize_t)sizeof(src1) * 8) { \
745
} else if (tmp == -(ssize_t)sizeof(src1) * 8) { \
746
dest = src1 >> (-tmp - 1); \
747
} else if (tmp < 0) { \
748
dest = (src1 + (1 << (-1 - tmp))) >> -tmp; \
750
dest = src1 << tmp; \
752
NEON_VOP(rshl_u8, neon_u8, 4)
753
NEON_VOP(rshl_u16, neon_u16, 2)
756
/* The addition of the rounding constant may overflow, so we use an
757
* intermediate 64 bit accumulator. */
758
uint32_t HELPER(neon_rshl_u32)(uint32_t val, uint32_t shiftop)
761
int8_t shift = (int8_t)shiftop;
762
if (shift >= 32 || shift < -32) {
764
} else if (shift == -32) {
766
} else if (shift < 0) {
767
uint64_t big_dest = ((uint64_t)val + (1 << (-1 - shift)));
768
dest = big_dest >> -shift;
775
/* Handling addition overflow with 64 bit input values is more
776
* tricky than with 32 bit values. */
777
uint64_t HELPER(neon_rshl_u64)(uint64_t val, uint64_t shiftop)
779
int8_t shift = (uint8_t)shiftop;
780
if (shift >= 64 || shift < -64) {
782
} else if (shift == -64) {
783
/* Rounding a 1-bit result just preserves that bit. */
785
} else if (shift < 0) {
786
val >>= (-shift - 1);
787
if (val == UINT64_MAX) {
788
/* In this case, it means that the rounding constant is 1,
789
* and the addition would overflow. Return the actual
790
* result directly. */
791
val = 0x8000000000000000ULL;
802
#define NEON_FN(dest, src1, src2) do { \
804
tmp = (int8_t)src2; \
805
if (tmp >= (ssize_t)sizeof(src1) * 8) { \
812
} else if (tmp <= -(ssize_t)sizeof(src1) * 8) { \
814
} else if (tmp < 0) { \
815
dest = src1 >> -tmp; \
817
dest = src1 << tmp; \
818
if ((dest >> tmp) != src1) { \
823
NEON_VOP_ENV(qshl_u8, neon_u8, 4)
824
NEON_VOP_ENV(qshl_u16, neon_u16, 2)
825
NEON_VOP_ENV(qshl_u32, neon_u32, 1)
828
uint64_t HELPER(neon_qshl_u64)(CPUARMState *env, uint64_t val, uint64_t shiftop)
830
int8_t shift = (int8_t)shiftop;
836
} else if (shift <= -64) {
838
} else if (shift < 0) {
843
if ((val >> shift) != tmp) {
851
#define NEON_FN(dest, src1, src2) do { \
853
tmp = (int8_t)src2; \
854
if (tmp >= (ssize_t)sizeof(src1) * 8) { \
857
dest = (uint32_t)(1 << (sizeof(src1) * 8 - 1)); \
864
} else if (tmp <= -(ssize_t)sizeof(src1) * 8) { \
866
} else if (tmp < 0) { \
867
dest = src1 >> -tmp; \
869
dest = src1 << tmp; \
870
if ((dest >> tmp) != src1) { \
872
dest = (uint32_t)(1 << (sizeof(src1) * 8 - 1)); \
878
NEON_VOP_ENV(qshl_s8, neon_s8, 4)
879
NEON_VOP_ENV(qshl_s16, neon_s16, 2)
880
NEON_VOP_ENV(qshl_s32, neon_s32, 1)
883
uint64_t HELPER(neon_qshl_s64)(CPUARMState *env, uint64_t valop, uint64_t shiftop)
885
int8_t shift = (uint8_t)shiftop;
890
val = (val >> 63) ^ ~SIGNBIT64;
892
} else if (shift <= -64) {
894
} else if (shift < 0) {
899
if ((val >> shift) != tmp) {
901
val = (tmp >> 63) ^ ~SIGNBIT64;
907
#define NEON_FN(dest, src1, src2) do { \
908
if (src1 & (1 << (sizeof(src1) * 8 - 1))) { \
913
tmp = (int8_t)src2; \
914
if (tmp >= (ssize_t)sizeof(src1) * 8) { \
921
} else if (tmp <= -(ssize_t)sizeof(src1) * 8) { \
923
} else if (tmp < 0) { \
924
dest = src1 >> -tmp; \
926
dest = src1 << tmp; \
927
if ((dest >> tmp) != src1) { \
933
NEON_VOP_ENV(qshlu_s8, neon_u8, 4)
934
NEON_VOP_ENV(qshlu_s16, neon_u16, 2)
937
uint32_t HELPER(neon_qshlu_s32)(CPUARMState *env, uint32_t valop, uint32_t shiftop)
939
if ((int32_t)valop < 0) {
943
return helper_neon_qshl_u32(env, valop, shiftop);
946
uint64_t HELPER(neon_qshlu_s64)(CPUARMState *env, uint64_t valop, uint64_t shiftop)
948
if ((int64_t)valop < 0) {
952
return helper_neon_qshl_u64(env, valop, shiftop);
955
#define NEON_FN(dest, src1, src2) do { \
957
tmp = (int8_t)src2; \
958
if (tmp >= (ssize_t)sizeof(src1) * 8) { \
965
} else if (tmp < -(ssize_t)sizeof(src1) * 8) { \
967
} else if (tmp == -(ssize_t)sizeof(src1) * 8) { \
968
dest = src1 >> (sizeof(src1) * 8 - 1); \
969
} else if (tmp < 0) { \
970
dest = (src1 + (1 << (-1 - tmp))) >> -tmp; \
972
dest = src1 << tmp; \
973
if ((dest >> tmp) != src1) { \
978
NEON_VOP_ENV(qrshl_u8, neon_u8, 4)
979
NEON_VOP_ENV(qrshl_u16, neon_u16, 2)
982
/* The addition of the rounding constant may overflow, so we use an
983
* intermediate 64 bit accumulator. */
984
uint32_t HELPER(neon_qrshl_u32)(CPUARMState *env, uint32_t val, uint32_t shiftop)
987
int8_t shift = (int8_t)shiftop;
995
} else if (shift < -32) {
997
} else if (shift == -32) {
999
} else if (shift < 0) {
1000
uint64_t big_dest = ((uint64_t)val + (1 << (-1 - shift)));
1001
dest = big_dest >> -shift;
1003
dest = val << shift;
1004
if ((dest >> shift) != val) {
1012
/* Handling addition overflow with 64 bit input values is more
1013
* tricky than with 32 bit values. */
1014
uint64_t HELPER(neon_qrshl_u64)(CPUARMState *env, uint64_t val, uint64_t shiftop)
1016
int8_t shift = (int8_t)shiftop;
1022
} else if (shift < -64) {
1024
} else if (shift == -64) {
1026
} else if (shift < 0) {
1027
val >>= (-shift - 1);
1028
if (val == UINT64_MAX) {
1029
/* In this case, it means that the rounding constant is 1,
1030
* and the addition would overflow. Return the actual
1031
* result directly. */
1032
val = 0x8000000000000000ULL;
1040
if ((val >> shift) != tmp) {
1048
#define NEON_FN(dest, src1, src2) do { \
1050
tmp = (int8_t)src2; \
1051
if (tmp >= (ssize_t)sizeof(src1) * 8) { \
1054
dest = (typeof(dest))(1 << (sizeof(src1) * 8 - 1)); \
1061
} else if (tmp <= -(ssize_t)sizeof(src1) * 8) { \
1063
} else if (tmp < 0) { \
1064
dest = (src1 + (1 << (-1 - tmp))) >> -tmp; \
1066
dest = src1 << tmp; \
1067
if ((dest >> tmp) != src1) { \
1069
dest = (uint32_t)(1 << (sizeof(src1) * 8 - 1)); \
1075
NEON_VOP_ENV(qrshl_s8, neon_s8, 4)
1076
NEON_VOP_ENV(qrshl_s16, neon_s16, 2)
1079
/* The addition of the rounding constant may overflow, so we use an
1080
* intermediate 64 bit accumulator. */
1081
uint32_t HELPER(neon_qrshl_s32)(CPUARMState *env, uint32_t valop, uint32_t shiftop)
1084
int32_t val = (int32_t)valop;
1085
int8_t shift = (int8_t)shiftop;
1089
dest = (val >> 31) ^ ~SIGNBIT;
1093
} else if (shift <= -32) {
1095
} else if (shift < 0) {
1096
int64_t big_dest = ((int64_t)val + (1 << (-1 - shift)));
1097
dest = big_dest >> -shift;
1099
dest = val << shift;
1100
if ((dest >> shift) != val) {
1102
dest = (val >> 31) ^ ~SIGNBIT;
1108
/* Handling addition overflow with 64 bit input values is more
1109
* tricky than with 32 bit values. */
1110
uint64_t HELPER(neon_qrshl_s64)(CPUARMState *env, uint64_t valop, uint64_t shiftop)
1112
int8_t shift = (uint8_t)shiftop;
1113
int64_t val = valop;
1118
val = (val >> 63) ^ ~SIGNBIT64;
1120
} else if (shift <= -64) {
1122
} else if (shift < 0) {
1123
val >>= (-shift - 1);
1124
if (val == INT64_MAX) {
1125
/* In this case, it means that the rounding constant is 1,
1126
* and the addition would overflow. Return the actual
1127
* result directly. */
1128
val = 0x4000000000000000ULL;
1136
if ((val >> shift) != tmp) {
1138
val = (tmp >> 63) ^ ~SIGNBIT64;
1144
uint32_t HELPER(neon_add_u8)(uint32_t a, uint32_t b)
1147
mask = (a ^ b) & 0x80808080u;
1150
return (a + b) ^ mask;
1153
uint32_t HELPER(neon_add_u16)(uint32_t a, uint32_t b)
1156
mask = (a ^ b) & 0x80008000u;
1159
return (a + b) ^ mask;
1162
#define NEON_FN(dest, src1, src2) dest = src1 + src2
1163
NEON_POP(padd_u8, neon_u8, 4)
1164
NEON_POP(padd_u16, neon_u16, 2)
1167
#define NEON_FN(dest, src1, src2) dest = src1 - src2
1168
NEON_VOP(sub_u8, neon_u8, 4)
1169
NEON_VOP(sub_u16, neon_u16, 2)
1172
#define NEON_FN(dest, src1, src2) dest = src1 * src2
1173
NEON_VOP(mul_u8, neon_u8, 4)
1174
NEON_VOP(mul_u16, neon_u16, 2)
1177
/* Polynomial multiplication is like integer multiplication except the
1178
partial products are XORed, not added. */
1179
uint32_t HELPER(neon_mul_p8)(uint32_t op1, uint32_t op2)
1189
mask |= (0xff << 8);
1190
if (op1 & (1 << 16))
1191
mask |= (0xff << 16);
1192
if (op1 & (1 << 24))
1193
mask |= (0xff << 24);
1194
result ^= op2 & mask;
1195
op1 = (op1 >> 1) & 0x7f7f7f7f;
1196
op2 = (op2 << 1) & 0xfefefefe;
1201
uint64_t HELPER(neon_mull_p8)(uint32_t op1, uint32_t op2)
1203
uint64_t result = 0;
1205
uint64_t op2ex = op2;
1206
op2ex = (op2ex & 0xff) |
1207
((op2ex & 0xff00) << 8) |
1208
((op2ex & 0xff0000) << 16) |
1209
((op2ex & 0xff000000) << 24);
1215
if (op1 & (1 << 8)) {
1216
mask |= (0xffffU << 16);
1218
if (op1 & (1 << 16)) {
1219
mask |= (0xffffULL << 32);
1221
if (op1 & (1 << 24)) {
1222
mask |= (0xffffULL << 48);
1224
result ^= op2ex & mask;
1225
op1 = (op1 >> 1) & 0x7f7f7f7f;
1231
#define NEON_FN(dest, src1, src2) dest = (src1 & src2) ? -1 : 0
1232
NEON_VOP(tst_u8, neon_u8, 4)
1233
NEON_VOP(tst_u16, neon_u16, 2)
1234
NEON_VOP(tst_u32, neon_u32, 1)
1237
#define NEON_FN(dest, src1, src2) dest = (src1 == src2) ? -1 : 0
1238
NEON_VOP(ceq_u8, neon_u8, 4)
1239
NEON_VOP(ceq_u16, neon_u16, 2)
1240
NEON_VOP(ceq_u32, neon_u32, 1)
1243
#define NEON_FN(dest, src, dummy) dest = (src < 0) ? -src : src
1244
NEON_VOP1(abs_s8, neon_s8, 4)
1245
NEON_VOP1(abs_s16, neon_s16, 2)
1248
/* Count Leading Sign/Zero Bits. */
1249
static inline int do_clz8(uint8_t x)
1257
static inline int do_clz16(uint16_t x)
1260
for (n = 16; x; n--)
1265
#define NEON_FN(dest, src, dummy) dest = do_clz8(src)
1266
NEON_VOP1(clz_u8, neon_u8, 4)
1269
#define NEON_FN(dest, src, dummy) dest = do_clz16(src)
1270
NEON_VOP1(clz_u16, neon_u16, 2)
1273
#define NEON_FN(dest, src, dummy) dest = do_clz8((src < 0) ? ~src : src) - 1
1274
NEON_VOP1(cls_s8, neon_s8, 4)
1277
#define NEON_FN(dest, src, dummy) dest = do_clz16((src < 0) ? ~src : src) - 1
1278
NEON_VOP1(cls_s16, neon_s16, 2)
1281
uint32_t HELPER(neon_cls_s32)(uint32_t x)
1286
for (count = 32; x; count--)
1292
uint32_t HELPER(neon_cnt_u8)(uint32_t x)
1294
x = (x & 0x55555555) + ((x >> 1) & 0x55555555);
1295
x = (x & 0x33333333) + ((x >> 2) & 0x33333333);
1296
x = (x & 0x0f0f0f0f) + ((x >> 4) & 0x0f0f0f0f);
1300
/* Reverse bits in each 8 bit word */
1301
uint32_t HELPER(neon_rbit_u8)(uint32_t x)
1303
x = ((x & 0xf0f0f0f0) >> 4)
1304
| ((x & 0x0f0f0f0f) << 4);
1305
x = ((x & 0x88888888) >> 3)
1306
| ((x & 0x44444444) >> 1)
1307
| ((x & 0x22222222) << 1)
1308
| ((x & 0x11111111) << 3);
1312
#define NEON_QDMULH16(dest, src1, src2, round) do { \
1313
uint32_t tmp = (int32_t)(int16_t) src1 * (int16_t) src2; \
1314
if ((tmp ^ (tmp << 1)) & SIGNBIT) { \
1316
tmp = (tmp >> 31) ^ ~SIGNBIT; \
1321
int32_t old = tmp; \
1323
if ((int32_t)tmp < old) { \
1325
tmp = SIGNBIT - 1; \
1330
#define NEON_FN(dest, src1, src2) NEON_QDMULH16(dest, src1, src2, 0)
1331
NEON_VOP_ENV(qdmulh_s16, neon_s16, 2)
1333
#define NEON_FN(dest, src1, src2) NEON_QDMULH16(dest, src1, src2, 1)
1334
NEON_VOP_ENV(qrdmulh_s16, neon_s16, 2)
1336
#undef NEON_QDMULH16
1338
#define NEON_QDMULH32(dest, src1, src2, round) do { \
1339
uint64_t tmp = (int64_t)(int32_t) src1 * (int32_t) src2; \
1340
if ((tmp ^ (tmp << 1)) & SIGNBIT64) { \
1342
tmp = (tmp >> 63) ^ ~SIGNBIT64; \
1347
int64_t old = tmp; \
1348
tmp += (int64_t)1 << 31; \
1349
if ((int64_t)tmp < old) { \
1351
tmp = SIGNBIT64 - 1; \
1356
#define NEON_FN(dest, src1, src2) NEON_QDMULH32(dest, src1, src2, 0)
1357
NEON_VOP_ENV(qdmulh_s32, neon_s32, 1)
1359
#define NEON_FN(dest, src1, src2) NEON_QDMULH32(dest, src1, src2, 1)
1360
NEON_VOP_ENV(qrdmulh_s32, neon_s32, 1)
1362
#undef NEON_QDMULH32
1364
uint32_t HELPER(neon_narrow_u8)(uint64_t x)
1366
return (x & 0xffu) | ((x >> 8) & 0xff00u) | ((x >> 16) & 0xff0000u)
1367
| ((x >> 24) & 0xff000000u);
1370
uint32_t HELPER(neon_narrow_u16)(uint64_t x)
1372
return (x & 0xffffu) | ((x >> 16) & 0xffff0000u);
1375
uint32_t HELPER(neon_narrow_high_u8)(uint64_t x)
1377
return ((x >> 8) & 0xff) | ((x >> 16) & 0xff00)
1378
| ((x >> 24) & 0xff0000) | ((x >> 32) & 0xff000000);
1381
uint32_t HELPER(neon_narrow_high_u16)(uint64_t x)
1383
return ((x >> 16) & 0xffff) | ((x >> 32) & 0xffff0000);
1386
uint32_t HELPER(neon_narrow_round_high_u8)(uint64_t x)
1388
x &= 0xff80ff80ff80ff80ull;
1389
x += 0x0080008000800080ull;
1390
return ((x >> 8) & 0xff) | ((x >> 16) & 0xff00)
1391
| ((x >> 24) & 0xff0000) | ((x >> 32) & 0xff000000);
1394
uint32_t HELPER(neon_narrow_round_high_u16)(uint64_t x)
1396
x &= 0xffff8000ffff8000ull;
1397
x += 0x0000800000008000ull;
1398
return ((x >> 16) & 0xffff) | ((x >> 32) & 0xffff0000);
1401
uint32_t HELPER(neon_unarrow_sat8)(CPUARMState *env, uint64_t x)
1417
res |= (uint32_t)d << (n / 2); \
1428
uint32_t HELPER(neon_narrow_sat_u8)(CPUARMState *env, uint64_t x)
1441
res |= (uint32_t)d << (n / 2);
1451
uint32_t HELPER(neon_narrow_sat_s8)(CPUARMState *env, uint64_t x)
1458
if (s != (int8_t)s) { \
1459
d = (s >> 15) ^ 0x7f; \
1464
res |= (uint32_t)d << (n / 2);
1474
uint32_t HELPER(neon_unarrow_sat16)(CPUARMState *env, uint64_t x)
1479
if (low & 0x80000000) {
1482
} else if (low > 0xffff) {
1487
if (high & 0x80000000) {
1490
} else if (high > 0xffff) {
1494
return low | (high << 16);
1497
uint32_t HELPER(neon_narrow_sat_u16)(CPUARMState *env, uint64_t x)
1507
if (high > 0xffff) {
1511
return low | (high << 16);
1514
uint32_t HELPER(neon_narrow_sat_s16)(CPUARMState *env, uint64_t x)
1519
if (low != (int16_t)low) {
1520
low = (low >> 31) ^ 0x7fff;
1524
if (high != (int16_t)high) {
1525
high = (high >> 31) ^ 0x7fff;
1528
return (uint16_t)low | (high << 16);
1531
uint32_t HELPER(neon_unarrow_sat32)(CPUARMState *env, uint64_t x)
1533
if (x & 0x8000000000000000ull) {
1537
if (x > 0xffffffffu) {
1544
uint32_t HELPER(neon_narrow_sat_u32)(CPUARMState *env, uint64_t x)
1546
if (x > 0xffffffffu) {
1553
uint32_t HELPER(neon_narrow_sat_s32)(CPUARMState *env, uint64_t x)
1555
if ((int64_t)x != (int32_t)x) {
1557
return ((int64_t)x >> 63) ^ 0x7fffffff;
1562
uint64_t HELPER(neon_widen_u8)(uint32_t x)
1567
tmp = (uint8_t)(x >> 8);
1569
tmp = (uint8_t)(x >> 16);
1571
tmp = (uint8_t)(x >> 24);
1576
uint64_t HELPER(neon_widen_s8)(uint32_t x)
1580
ret = (uint16_t)(int8_t)x;
1581
tmp = (uint16_t)(int8_t)(x >> 8);
1583
tmp = (uint16_t)(int8_t)(x >> 16);
1585
tmp = (uint16_t)(int8_t)(x >> 24);
1590
uint64_t HELPER(neon_widen_u16)(uint32_t x)
1592
uint64_t high = (uint16_t)(x >> 16);
1593
return ((uint16_t)x) | (high << 32);
1596
uint64_t HELPER(neon_widen_s16)(uint32_t x)
1598
uint64_t high = (int16_t)(x >> 16);
1599
return ((uint32_t)(int16_t)x) | (high << 32);
1602
uint64_t HELPER(neon_addl_u16)(uint64_t a, uint64_t b)
1605
mask = (a ^ b) & 0x8000800080008000ull;
1606
a &= ~0x8000800080008000ull;
1607
b &= ~0x8000800080008000ull;
1608
return (a + b) ^ mask;
1611
uint64_t HELPER(neon_addl_u32)(uint64_t a, uint64_t b)
1614
mask = (a ^ b) & 0x8000000080000000ull;
1615
a &= ~0x8000000080000000ull;
1616
b &= ~0x8000000080000000ull;
1617
return (a + b) ^ mask;
1620
uint64_t HELPER(neon_paddl_u16)(uint64_t a, uint64_t b)
1625
tmp = a & 0x0000ffff0000ffffull;
1626
tmp += (a >> 16) & 0x0000ffff0000ffffull;
1627
tmp2 = b & 0xffff0000ffff0000ull;
1628
tmp2 += (b << 16) & 0xffff0000ffff0000ull;
1629
return ( tmp & 0xffff)
1630
| ((tmp >> 16) & 0xffff0000ull)
1631
| ((tmp2 << 16) & 0xffff00000000ull)
1632
| ( tmp2 & 0xffff000000000000ull);
1635
uint64_t HELPER(neon_paddl_u32)(uint64_t a, uint64_t b)
1637
uint32_t low = a + (a >> 32);
1638
uint32_t high = b + (b >> 32);
1639
return low + ((uint64_t)high << 32);
1642
uint64_t HELPER(neon_subl_u16)(uint64_t a, uint64_t b)
1645
mask = (a ^ ~b) & 0x8000800080008000ull;
1646
a |= 0x8000800080008000ull;
1647
b &= ~0x8000800080008000ull;
1648
return (a - b) ^ mask;
1651
uint64_t HELPER(neon_subl_u32)(uint64_t a, uint64_t b)
1654
mask = (a ^ ~b) & 0x8000000080000000ull;
1655
a |= 0x8000000080000000ull;
1656
b &= ~0x8000000080000000ull;
1657
return (a - b) ^ mask;
1660
uint64_t HELPER(neon_addl_saturate_s32)(CPUARMState *env, uint64_t a, uint64_t b)
1668
if (((low ^ x) & SIGNBIT) && !((x ^ y) & SIGNBIT)) {
1670
low = ((int32_t)x >> 31) ^ ~SIGNBIT;
1675
if (((high ^ x) & SIGNBIT) && !((x ^ y) & SIGNBIT)) {
1677
high = ((int32_t)x >> 31) ^ ~SIGNBIT;
1679
return low | ((uint64_t)high << 32);
1682
uint64_t HELPER(neon_addl_saturate_s64)(CPUARMState *env, uint64_t a, uint64_t b)
1687
if (((result ^ a) & SIGNBIT64) && !((a ^ b) & SIGNBIT64)) {
1689
result = ((int64_t)a >> 63) ^ ~SIGNBIT64;
1694
/* We have to do the arithmetic in a larger type than
1695
* the input type, because for example with a signed 32 bit
1696
* op the absolute difference can overflow a signed 32 bit value.
1698
#define DO_ABD(dest, x, y, intype, arithtype) do { \
1699
arithtype tmp_x = (intype)(x); \
1700
arithtype tmp_y = (intype)(y); \
1701
dest = ((tmp_x > tmp_y) ? tmp_x - tmp_y : tmp_y - tmp_x); \
1704
uint64_t HELPER(neon_abdl_u16)(uint32_t a, uint32_t b)
1708
DO_ABD(result, a, b, uint8_t, uint32_t);
1709
DO_ABD(tmp, a >> 8, b >> 8, uint8_t, uint32_t);
1710
result |= tmp << 16;
1711
DO_ABD(tmp, a >> 16, b >> 16, uint8_t, uint32_t);
1712
result |= tmp << 32;
1713
DO_ABD(tmp, a >> 24, b >> 24, uint8_t, uint32_t);
1714
result |= tmp << 48;
1718
uint64_t HELPER(neon_abdl_s16)(uint32_t a, uint32_t b)
1722
DO_ABD(result, a, b, int8_t, int32_t);
1723
DO_ABD(tmp, a >> 8, b >> 8, int8_t, int32_t);
1724
result |= tmp << 16;
1725
DO_ABD(tmp, a >> 16, b >> 16, int8_t, int32_t);
1726
result |= tmp << 32;
1727
DO_ABD(tmp, a >> 24, b >> 24, int8_t, int32_t);
1728
result |= tmp << 48;
1732
uint64_t HELPER(neon_abdl_u32)(uint32_t a, uint32_t b)
1736
DO_ABD(result, a, b, uint16_t, uint32_t);
1737
DO_ABD(tmp, a >> 16, b >> 16, uint16_t, uint32_t);
1738
return result | (tmp << 32);
1741
uint64_t HELPER(neon_abdl_s32)(uint32_t a, uint32_t b)
1745
DO_ABD(result, a, b, int16_t, int32_t);
1746
DO_ABD(tmp, a >> 16, b >> 16, int16_t, int32_t);
1747
return result | (tmp << 32);
1750
uint64_t HELPER(neon_abdl_u64)(uint32_t a, uint32_t b)
1753
DO_ABD(result, a, b, uint32_t, uint64_t);
1757
uint64_t HELPER(neon_abdl_s64)(uint32_t a, uint32_t b)
1760
DO_ABD(result, a, b, int32_t, int64_t);
1765
/* Widening multiply. Named type is the source type. */
1766
#define DO_MULL(dest, x, y, type1, type2) do { \
1769
dest = (type2)((type2)tmp_x * (type2)tmp_y); \
1772
uint64_t HELPER(neon_mull_u8)(uint32_t a, uint32_t b)
1777
DO_MULL(result, a, b, uint8_t, uint16_t);
1778
DO_MULL(tmp, a >> 8, b >> 8, uint8_t, uint16_t);
1779
result |= tmp << 16;
1780
DO_MULL(tmp, a >> 16, b >> 16, uint8_t, uint16_t);
1781
result |= tmp << 32;
1782
DO_MULL(tmp, a >> 24, b >> 24, uint8_t, uint16_t);
1783
result |= tmp << 48;
1787
uint64_t HELPER(neon_mull_s8)(uint32_t a, uint32_t b)
1792
DO_MULL(result, a, b, int8_t, uint16_t);
1793
DO_MULL(tmp, a >> 8, b >> 8, int8_t, uint16_t);
1794
result |= tmp << 16;
1795
DO_MULL(tmp, a >> 16, b >> 16, int8_t, uint16_t);
1796
result |= tmp << 32;
1797
DO_MULL(tmp, a >> 24, b >> 24, int8_t, uint16_t);
1798
result |= tmp << 48;
1802
uint64_t HELPER(neon_mull_u16)(uint32_t a, uint32_t b)
1807
DO_MULL(result, a, b, uint16_t, uint32_t);
1808
DO_MULL(tmp, a >> 16, b >> 16, uint16_t, uint32_t);
1809
return result | (tmp << 32);
1812
uint64_t HELPER(neon_mull_s16)(uint32_t a, uint32_t b)
1817
DO_MULL(result, a, b, int16_t, uint32_t);
1818
DO_MULL(tmp, a >> 16, b >> 16, int16_t, uint32_t);
1819
return result | (tmp << 32);
1822
uint64_t HELPER(neon_negl_u16)(uint64_t x)
1826
result = (uint16_t)-x;
1828
result |= (uint64_t)tmp << 16;
1830
result |= (uint64_t)tmp << 32;
1832
result |= (uint64_t)tmp << 48;
1836
uint64_t HELPER(neon_negl_u32)(uint64_t x)
1839
uint32_t high = -(x >> 32);
1840
return low | ((uint64_t)high << 32);
1843
/* Saturating sign manipulation. */
1844
/* ??? Make these use NEON_VOP1 */
1845
#define DO_QABS8(x) do { \
1846
if (x == (int8_t)0x80) { \
1849
} else if (x < 0) { \
1852
uint32_t HELPER(neon_qabs_s8)(CPUARMState *env, uint32_t x)
1855
NEON_UNPACK(neon_s8, vec, x);
1860
NEON_PACK(neon_s8, x, vec);
1865
#define DO_QNEG8(x) do { \
1866
if (x == (int8_t)0x80) { \
1872
uint32_t HELPER(neon_qneg_s8)(CPUARMState *env, uint32_t x)
1875
NEON_UNPACK(neon_s8, vec, x);
1880
NEON_PACK(neon_s8, x, vec);
1885
#define DO_QABS16(x) do { \
1886
if (x == (int16_t)0x8000) { \
1889
} else if (x < 0) { \
1892
uint32_t HELPER(neon_qabs_s16)(CPUARMState *env, uint32_t x)
1895
NEON_UNPACK(neon_s16, vec, x);
1898
NEON_PACK(neon_s16, x, vec);
1903
#define DO_QNEG16(x) do { \
1904
if (x == (int16_t)0x8000) { \
1910
uint32_t HELPER(neon_qneg_s16)(CPUARMState *env, uint32_t x)
1913
NEON_UNPACK(neon_s16, vec, x);
1916
NEON_PACK(neon_s16, x, vec);
1921
uint32_t HELPER(neon_qabs_s32)(CPUARMState *env, uint32_t x)
1926
} else if ((int32_t)x < 0) {
1932
uint32_t HELPER(neon_qneg_s32)(CPUARMState *env, uint32_t x)
1943
uint64_t HELPER(neon_qabs_s64)(CPUARMState *env, uint64_t x)
1945
if (x == SIGNBIT64) {
1948
} else if ((int64_t)x < 0) {
1954
uint64_t HELPER(neon_qneg_s64)(CPUARMState *env, uint64_t x)
1956
if (x == SIGNBIT64) {
1965
/* NEON Float helpers. */
1966
uint32_t HELPER(neon_abd_f32)(uint32_t a, uint32_t b, void *fpstp)
1968
float_status *fpst = fpstp;
1969
float32 f0 = make_float32(a);
1970
float32 f1 = make_float32(b);
1971
return float32_val(float32_abs(float32_sub(f0, f1, fpst)));
1974
/* Floating point comparisons produce an integer result.
1975
* Note that EQ doesn't signal InvalidOp for QNaNs but GE and GT do.
1976
* Softfloat routines return 0/1, which we convert to the 0/-1 Neon requires.
1978
uint32_t HELPER(neon_ceq_f32)(uint32_t a, uint32_t b, void *fpstp)
1980
float_status *fpst = fpstp;
1981
return -float32_eq_quiet(make_float32(a), make_float32(b), fpst);
1984
uint32_t HELPER(neon_cge_f32)(uint32_t a, uint32_t b, void *fpstp)
1986
float_status *fpst = fpstp;
1987
return -float32_le(make_float32(b), make_float32(a), fpst);
1990
uint32_t HELPER(neon_cgt_f32)(uint32_t a, uint32_t b, void *fpstp)
1992
float_status *fpst = fpstp;
1993
return -float32_lt(make_float32(b), make_float32(a), fpst);
1996
uint32_t HELPER(neon_acge_f32)(uint32_t a, uint32_t b, void *fpstp)
1998
float_status *fpst = fpstp;
1999
float32 f0 = float32_abs(make_float32(a));
2000
float32 f1 = float32_abs(make_float32(b));
2001
return -float32_le(f1, f0, fpst);
2004
uint32_t HELPER(neon_acgt_f32)(uint32_t a, uint32_t b, void *fpstp)
2006
float_status *fpst = fpstp;
2007
float32 f0 = float32_abs(make_float32(a));
2008
float32 f1 = float32_abs(make_float32(b));
2009
return -float32_lt(f1, f0, fpst);
2012
uint64_t HELPER(neon_acge_f64)(uint64_t a, uint64_t b, void *fpstp)
2014
float_status *fpst = fpstp;
2015
float64 f0 = float64_abs(make_float64(a));
2016
float64 f1 = float64_abs(make_float64(b));
2017
return -float64_le(f1, f0, fpst);
2020
uint64_t HELPER(neon_acgt_f64)(uint64_t a, uint64_t b, void *fpstp)
2022
float_status *fpst = fpstp;
2023
float64 f0 = float64_abs(make_float64(a));
2024
float64 f1 = float64_abs(make_float64(b));
2025
return -float64_lt(f1, f0, fpst);
2028
#define ELEM(V, N, SIZE) (((V) >> ((N) * (SIZE))) & ((1ull << (SIZE)) - 1))
2030
void HELPER(neon_qunzip8)(CPUARMState *env, uint32_t rd, uint32_t rm)
2032
uint64_t zm0 = float64_val(env->vfp.regs[rm]);
2033
uint64_t zm1 = float64_val(env->vfp.regs[rm + 1]);
2034
uint64_t zd0 = float64_val(env->vfp.regs[rd]);
2035
uint64_t zd1 = float64_val(env->vfp.regs[rd + 1]);
2036
uint64_t d0 = ELEM(zd0, 0, 8) | (ELEM(zd0, 2, 8) << 8)
2037
| (ELEM(zd0, 4, 8) << 16) | (ELEM(zd0, 6, 8) << 24)
2038
| (ELEM(zd1, 0, 8) << 32) | (ELEM(zd1, 2, 8) << 40)
2039
| (ELEM(zd1, 4, 8) << 48) | (ELEM(zd1, 6, 8) << 56);
2040
uint64_t d1 = ELEM(zm0, 0, 8) | (ELEM(zm0, 2, 8) << 8)
2041
| (ELEM(zm0, 4, 8) << 16) | (ELEM(zm0, 6, 8) << 24)
2042
| (ELEM(zm1, 0, 8) << 32) | (ELEM(zm1, 2, 8) << 40)
2043
| (ELEM(zm1, 4, 8) << 48) | (ELEM(zm1, 6, 8) << 56);
2044
uint64_t m0 = ELEM(zd0, 1, 8) | (ELEM(zd0, 3, 8) << 8)
2045
| (ELEM(zd0, 5, 8) << 16) | (ELEM(zd0, 7, 8) << 24)
2046
| (ELEM(zd1, 1, 8) << 32) | (ELEM(zd1, 3, 8) << 40)
2047
| (ELEM(zd1, 5, 8) << 48) | (ELEM(zd1, 7, 8) << 56);
2048
uint64_t m1 = ELEM(zm0, 1, 8) | (ELEM(zm0, 3, 8) << 8)
2049
| (ELEM(zm0, 5, 8) << 16) | (ELEM(zm0, 7, 8) << 24)
2050
| (ELEM(zm1, 1, 8) << 32) | (ELEM(zm1, 3, 8) << 40)
2051
| (ELEM(zm1, 5, 8) << 48) | (ELEM(zm1, 7, 8) << 56);
2052
env->vfp.regs[rm] = make_float64(m0);
2053
env->vfp.regs[rm + 1] = make_float64(m1);
2054
env->vfp.regs[rd] = make_float64(d0);
2055
env->vfp.regs[rd + 1] = make_float64(d1);
2058
void HELPER(neon_qunzip16)(CPUARMState *env, uint32_t rd, uint32_t rm)
2060
uint64_t zm0 = float64_val(env->vfp.regs[rm]);
2061
uint64_t zm1 = float64_val(env->vfp.regs[rm + 1]);
2062
uint64_t zd0 = float64_val(env->vfp.regs[rd]);
2063
uint64_t zd1 = float64_val(env->vfp.regs[rd + 1]);
2064
uint64_t d0 = ELEM(zd0, 0, 16) | (ELEM(zd0, 2, 16) << 16)
2065
| (ELEM(zd1, 0, 16) << 32) | (ELEM(zd1, 2, 16) << 48);
2066
uint64_t d1 = ELEM(zm0, 0, 16) | (ELEM(zm0, 2, 16) << 16)
2067
| (ELEM(zm1, 0, 16) << 32) | (ELEM(zm1, 2, 16) << 48);
2068
uint64_t m0 = ELEM(zd0, 1, 16) | (ELEM(zd0, 3, 16) << 16)
2069
| (ELEM(zd1, 1, 16) << 32) | (ELEM(zd1, 3, 16) << 48);
2070
uint64_t m1 = ELEM(zm0, 1, 16) | (ELEM(zm0, 3, 16) << 16)
2071
| (ELEM(zm1, 1, 16) << 32) | (ELEM(zm1, 3, 16) << 48);
2072
env->vfp.regs[rm] = make_float64(m0);
2073
env->vfp.regs[rm + 1] = make_float64(m1);
2074
env->vfp.regs[rd] = make_float64(d0);
2075
env->vfp.regs[rd + 1] = make_float64(d1);
2078
void HELPER(neon_qunzip32)(CPUARMState *env, uint32_t rd, uint32_t rm)
2080
uint64_t zm0 = float64_val(env->vfp.regs[rm]);
2081
uint64_t zm1 = float64_val(env->vfp.regs[rm + 1]);
2082
uint64_t zd0 = float64_val(env->vfp.regs[rd]);
2083
uint64_t zd1 = float64_val(env->vfp.regs[rd + 1]);
2084
uint64_t d0 = ELEM(zd0, 0, 32) | (ELEM(zd1, 0, 32) << 32);
2085
uint64_t d1 = ELEM(zm0, 0, 32) | (ELEM(zm1, 0, 32) << 32);
2086
uint64_t m0 = ELEM(zd0, 1, 32) | (ELEM(zd1, 1, 32) << 32);
2087
uint64_t m1 = ELEM(zm0, 1, 32) | (ELEM(zm1, 1, 32) << 32);
2088
env->vfp.regs[rm] = make_float64(m0);
2089
env->vfp.regs[rm + 1] = make_float64(m1);
2090
env->vfp.regs[rd] = make_float64(d0);
2091
env->vfp.regs[rd + 1] = make_float64(d1);
2094
void HELPER(neon_unzip8)(CPUARMState *env, uint32_t rd, uint32_t rm)
2096
uint64_t zm = float64_val(env->vfp.regs[rm]);
2097
uint64_t zd = float64_val(env->vfp.regs[rd]);
2098
uint64_t d0 = ELEM(zd, 0, 8) | (ELEM(zd, 2, 8) << 8)
2099
| (ELEM(zd, 4, 8) << 16) | (ELEM(zd, 6, 8) << 24)
2100
| (ELEM(zm, 0, 8) << 32) | (ELEM(zm, 2, 8) << 40)
2101
| (ELEM(zm, 4, 8) << 48) | (ELEM(zm, 6, 8) << 56);
2102
uint64_t m0 = ELEM(zd, 1, 8) | (ELEM(zd, 3, 8) << 8)
2103
| (ELEM(zd, 5, 8) << 16) | (ELEM(zd, 7, 8) << 24)
2104
| (ELEM(zm, 1, 8) << 32) | (ELEM(zm, 3, 8) << 40)
2105
| (ELEM(zm, 5, 8) << 48) | (ELEM(zm, 7, 8) << 56);
2106
env->vfp.regs[rm] = make_float64(m0);
2107
env->vfp.regs[rd] = make_float64(d0);
2110
void HELPER(neon_unzip16)(CPUARMState *env, uint32_t rd, uint32_t rm)
2112
uint64_t zm = float64_val(env->vfp.regs[rm]);
2113
uint64_t zd = float64_val(env->vfp.regs[rd]);
2114
uint64_t d0 = ELEM(zd, 0, 16) | (ELEM(zd, 2, 16) << 16)
2115
| (ELEM(zm, 0, 16) << 32) | (ELEM(zm, 2, 16) << 48);
2116
uint64_t m0 = ELEM(zd, 1, 16) | (ELEM(zd, 3, 16) << 16)
2117
| (ELEM(zm, 1, 16) << 32) | (ELEM(zm, 3, 16) << 48);
2118
env->vfp.regs[rm] = make_float64(m0);
2119
env->vfp.regs[rd] = make_float64(d0);
2122
void HELPER(neon_qzip8)(CPUARMState *env, uint32_t rd, uint32_t rm)
2124
uint64_t zm0 = float64_val(env->vfp.regs[rm]);
2125
uint64_t zm1 = float64_val(env->vfp.regs[rm + 1]);
2126
uint64_t zd0 = float64_val(env->vfp.regs[rd]);
2127
uint64_t zd1 = float64_val(env->vfp.regs[rd + 1]);
2128
uint64_t d0 = ELEM(zd0, 0, 8) | (ELEM(zm0, 0, 8) << 8)
2129
| (ELEM(zd0, 1, 8) << 16) | (ELEM(zm0, 1, 8) << 24)
2130
| (ELEM(zd0, 2, 8) << 32) | (ELEM(zm0, 2, 8) << 40)
2131
| (ELEM(zd0, 3, 8) << 48) | (ELEM(zm0, 3, 8) << 56);
2132
uint64_t d1 = ELEM(zd0, 4, 8) | (ELEM(zm0, 4, 8) << 8)
2133
| (ELEM(zd0, 5, 8) << 16) | (ELEM(zm0, 5, 8) << 24)
2134
| (ELEM(zd0, 6, 8) << 32) | (ELEM(zm0, 6, 8) << 40)
2135
| (ELEM(zd0, 7, 8) << 48) | (ELEM(zm0, 7, 8) << 56);
2136
uint64_t m0 = ELEM(zd1, 0, 8) | (ELEM(zm1, 0, 8) << 8)
2137
| (ELEM(zd1, 1, 8) << 16) | (ELEM(zm1, 1, 8) << 24)
2138
| (ELEM(zd1, 2, 8) << 32) | (ELEM(zm1, 2, 8) << 40)
2139
| (ELEM(zd1, 3, 8) << 48) | (ELEM(zm1, 3, 8) << 56);
2140
uint64_t m1 = ELEM(zd1, 4, 8) | (ELEM(zm1, 4, 8) << 8)
2141
| (ELEM(zd1, 5, 8) << 16) | (ELEM(zm1, 5, 8) << 24)
2142
| (ELEM(zd1, 6, 8) << 32) | (ELEM(zm1, 6, 8) << 40)
2143
| (ELEM(zd1, 7, 8) << 48) | (ELEM(zm1, 7, 8) << 56);
2144
env->vfp.regs[rm] = make_float64(m0);
2145
env->vfp.regs[rm + 1] = make_float64(m1);
2146
env->vfp.regs[rd] = make_float64(d0);
2147
env->vfp.regs[rd + 1] = make_float64(d1);
2150
void HELPER(neon_qzip16)(CPUARMState *env, uint32_t rd, uint32_t rm)
2152
uint64_t zm0 = float64_val(env->vfp.regs[rm]);
2153
uint64_t zm1 = float64_val(env->vfp.regs[rm + 1]);
2154
uint64_t zd0 = float64_val(env->vfp.regs[rd]);
2155
uint64_t zd1 = float64_val(env->vfp.regs[rd + 1]);
2156
uint64_t d0 = ELEM(zd0, 0, 16) | (ELEM(zm0, 0, 16) << 16)
2157
| (ELEM(zd0, 1, 16) << 32) | (ELEM(zm0, 1, 16) << 48);
2158
uint64_t d1 = ELEM(zd0, 2, 16) | (ELEM(zm0, 2, 16) << 16)
2159
| (ELEM(zd0, 3, 16) << 32) | (ELEM(zm0, 3, 16) << 48);
2160
uint64_t m0 = ELEM(zd1, 0, 16) | (ELEM(zm1, 0, 16) << 16)
2161
| (ELEM(zd1, 1, 16) << 32) | (ELEM(zm1, 1, 16) << 48);
2162
uint64_t m1 = ELEM(zd1, 2, 16) | (ELEM(zm1, 2, 16) << 16)
2163
| (ELEM(zd1, 3, 16) << 32) | (ELEM(zm1, 3, 16) << 48);
2164
env->vfp.regs[rm] = make_float64(m0);
2165
env->vfp.regs[rm + 1] = make_float64(m1);
2166
env->vfp.regs[rd] = make_float64(d0);
2167
env->vfp.regs[rd + 1] = make_float64(d1);
2170
void HELPER(neon_qzip32)(CPUARMState *env, uint32_t rd, uint32_t rm)
2172
uint64_t zm0 = float64_val(env->vfp.regs[rm]);
2173
uint64_t zm1 = float64_val(env->vfp.regs[rm + 1]);
2174
uint64_t zd0 = float64_val(env->vfp.regs[rd]);
2175
uint64_t zd1 = float64_val(env->vfp.regs[rd + 1]);
2176
uint64_t d0 = ELEM(zd0, 0, 32) | (ELEM(zm0, 0, 32) << 32);
2177
uint64_t d1 = ELEM(zd0, 1, 32) | (ELEM(zm0, 1, 32) << 32);
2178
uint64_t m0 = ELEM(zd1, 0, 32) | (ELEM(zm1, 0, 32) << 32);
2179
uint64_t m1 = ELEM(zd1, 1, 32) | (ELEM(zm1, 1, 32) << 32);
2180
env->vfp.regs[rm] = make_float64(m0);
2181
env->vfp.regs[rm + 1] = make_float64(m1);
2182
env->vfp.regs[rd] = make_float64(d0);
2183
env->vfp.regs[rd + 1] = make_float64(d1);
2186
void HELPER(neon_zip8)(CPUARMState *env, uint32_t rd, uint32_t rm)
2188
uint64_t zm = float64_val(env->vfp.regs[rm]);
2189
uint64_t zd = float64_val(env->vfp.regs[rd]);
2190
uint64_t d0 = ELEM(zd, 0, 8) | (ELEM(zm, 0, 8) << 8)
2191
| (ELEM(zd, 1, 8) << 16) | (ELEM(zm, 1, 8) << 24)
2192
| (ELEM(zd, 2, 8) << 32) | (ELEM(zm, 2, 8) << 40)
2193
| (ELEM(zd, 3, 8) << 48) | (ELEM(zm, 3, 8) << 56);
2194
uint64_t m0 = ELEM(zd, 4, 8) | (ELEM(zm, 4, 8) << 8)
2195
| (ELEM(zd, 5, 8) << 16) | (ELEM(zm, 5, 8) << 24)
2196
| (ELEM(zd, 6, 8) << 32) | (ELEM(zm, 6, 8) << 40)
2197
| (ELEM(zd, 7, 8) << 48) | (ELEM(zm, 7, 8) << 56);
2198
env->vfp.regs[rm] = make_float64(m0);
2199
env->vfp.regs[rd] = make_float64(d0);
2202
void HELPER(neon_zip16)(CPUARMState *env, uint32_t rd, uint32_t rm)
2204
uint64_t zm = float64_val(env->vfp.regs[rm]);
2205
uint64_t zd = float64_val(env->vfp.regs[rd]);
2206
uint64_t d0 = ELEM(zd, 0, 16) | (ELEM(zm, 0, 16) << 16)
2207
| (ELEM(zd, 1, 16) << 32) | (ELEM(zm, 1, 16) << 48);
2208
uint64_t m0 = ELEM(zd, 2, 16) | (ELEM(zm, 2, 16) << 16)
2209
| (ELEM(zd, 3, 16) << 32) | (ELEM(zm, 3, 16) << 48);
2210
env->vfp.regs[rm] = make_float64(m0);
2211
env->vfp.regs[rd] = make_float64(d0);
2214
/* Helper function for 64 bit polynomial multiply case:
2215
* perform PolynomialMult(op1, op2) and return either the top or
2216
* bottom half of the 128 bit result.
2218
uint64_t HELPER(neon_pmull_64_lo)(uint64_t op1, uint64_t op2)
2223
for (bitnum = 0; bitnum < 64; bitnum++) {
2224
if (op1 & (1ULL << bitnum)) {
2225
res ^= op2 << bitnum;
2230
uint64_t HELPER(neon_pmull_64_hi)(uint64_t op1, uint64_t op2)
2235
/* bit 0 of op1 can't influence the high 64 bits at all */
2236
for (bitnum = 1; bitnum < 64; bitnum++) {
2237
if (op1 & (1ULL << bitnum)) {
2238
res ^= op2 >> (64 - bitnum);