summaryrefslogtreecommitdiffstats
path: root/src/corelib/tools/qsimd_p.h
blob: 1a795a670dfc2854cece13e16f59780c7d987aa1 (plain)
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
/****************************************************************************
**
** Copyright (C) 2016 The Qt Company Ltd.
** Copyright (C) 2016 Intel Corporation.
** Contact: https://www.qt.io/licensing/
**
** This file is part of the QtCore module of the Qt Toolkit.
**
** $QT_BEGIN_LICENSE:LGPL$
** Commercial License Usage
** Licensees holding valid commercial Qt licenses may use this file in
** accordance with the commercial license agreement provided with the
** Software or, alternatively, in accordance with the terms contained in
** a written agreement between you and The Qt Company. For licensing terms
** and conditions see https://www.qt.io/terms-conditions. For further
** information use the contact form at https://www.qt.io/contact-us.
**
** GNU Lesser General Public License Usage
** Alternatively, this file may be used under the terms of the GNU Lesser
** General Public License version 3 as published by the Free Software
** Foundation and appearing in the file LICENSE.LGPL3 included in the
** packaging of this file. Please review the following information to
** ensure the GNU Lesser General Public License version 3 requirements
** will be met: https://www.gnu.org/licenses/lgpl-3.0.html.
**
** GNU General Public License Usage
** Alternatively, this file may be used under the terms of the GNU
** General Public License version 2.0 or (at your option) the GNU General
** Public license version 3 or any later version approved by the KDE Free
** Qt Foundation. The licenses are as published by the Free Software
** Foundation and appearing in the file LICENSE.GPL2 and LICENSE.GPL3
** included in the packaging of this file. Please review the following
** information to ensure the GNU General Public License requirements will
** be met: https://www.gnu.org/licenses/gpl-2.0.html and
** https://www.gnu.org/licenses/gpl-3.0.html.
**
** $QT_END_LICENSE$
**
****************************************************************************/

#ifndef QSIMD_P_H
#define QSIMD_P_H

//
//  W A R N I N G
//  -------------
//
// This file is not part of the Qt API.  It exists purely as an
// implementation detail.  This header file may change from version to
// version without notice, or even be removed.
//
// We mean it.
//

#include <qglobal.h>
#include <qatomic.h>

/*
 * qt_module_config.prf defines the QT_COMPILER_SUPPORTS_XXX macros.
 * They mean the compiler supports the necessary flags and the headers
 * for the x86 and ARM intrinsics:
 *  - GCC: the -mXXX or march=YYY flag is necessary before #include
 *    up to 4.8; GCC >= 4.9 can include unconditionally
 *  - Intel CC: #include can happen unconditionally
 *  - MSVC: #include can happen unconditionally
 *  - RVCT: ???
 *
 * We will try to include all headers possible under this configuration.
 *
 * MSVC does not define __SSE2__ & family, so we will define them. MSVC 2013 &
 * up do define __AVX__ if the -arch:AVX option is passed on the command-line.
 *
 * Supported XXX are:
 *   Flag    | Arch |  GCC  | Intel CC |  MSVC  |
 *  ARM_NEON | ARM  | I & C | None     |   ?    |
 *  SSE2     | x86  | I & C | I & C    | I & C  |
 *  SSE3     | x86  | I & C | I & C    | I only |
 *  SSSE3    | x86  | I & C | I & C    | I only |
 *  SSE4_1   | x86  | I & C | I & C    | I only |
 *  SSE4_2   | x86  | I & C | I & C    | I only |
 *  AVX      | x86  | I & C | I & C    | I & C  |
 *  AVX2     | x86  | I & C | I & C    | I only |
 *  AVX512xx | x86  | I & C | I & C    | I only |
 * I = intrinsics; C = code generation
 *
 * Code can use the following constructs to determine compiler support & status:
 * - #ifdef __XXX__      (e.g: #ifdef __AVX__  or #ifdef __ARM_NEON__)
 *   If this test passes, then the compiler is already generating code for that
 *   given sub-architecture. The intrinsics for that sub-architecture are
 *   #included and can be used without restriction or runtime check.
 *
 * - #if QT_COMPILER_SUPPORTS(XXX)
 *   If this test passes, then the compiler is able to generate code for that
 *   given sub-architecture in another translation unit, given the right set of
 *   flags. Use of the intrinsics is not guaranteed. This is useful with
 *   runtime detection (see below).
 *
 * - #if QT_COMPILER_SUPPORTS_HERE(XXX)
 *   If this test passes, then the compiler is able to generate code for that
 *   given sub-architecture in this translation unit, even if it is not doing
 *   that now (it might be). Individual functions may be tagged with
 *   QT_FUNCTION_TARGET(XXX) to cause the compiler to generate code for that
 *   sub-arch. Only inside such functions is the use of the intrisics
 *   guaranteed to work. This is useful with runtime detection (see below).
 *
 * Runtime detection of a CPU sub-architecture can be done with the
 * qCpuHasFeature(XXX) function. There are two strategies for generating
 * optimized code like that:
 *
 * 1) place the optimized code in a different translation unit (C or assembly
 * sources) and pass the correct flags to the compiler to enable support. Those
 * sources must not include qglobal.h, which means they cannot include this
 * file either. The dispatcher function would look like this:
 *
 *      void foo()
 *      {
 *      #if QT_COMPILER_SUPPORTS(XXX)
 *          if (qCpuHasFeature(XXX)) {
 *              foo_optimized_xxx();
 *              return;
 *          }
 *      #endif
 *          foo_plain();
 *      }
 *
 * 2) place the optimized code in a function tagged with QT_FUNCTION_TARGET and
 * surrounded by #if QT_COMPILER_SUPPORTS_HERE(XXX). That code can freely use
 * other Qt code. The dispatcher function would look like this:
 *
 *      void foo()
 *      {
 *      #if QT_COMPILER_SUPPORTS_HERE(XXX)
 *          if (qCpuHasFeature(XXX)) {
 *              foo_optimized_xxx();
 *              return;
 *          }
 *      #endif
 *          foo_plain();
 *      }
 */

#if defined(__MINGW64_VERSION_MAJOR) || (defined(Q_CC_MSVC) && !defined(Q_OS_WINCE))
#include <intrin.h>
#endif

#define QT_COMPILER_SUPPORTS(x)     (QT_COMPILER_SUPPORTS_ ## x - 0)

#if defined(Q_PROCESSOR_ARM)
#  define QT_COMPILER_SUPPORTS_HERE(x)    (__ARM_FEATURE_ ## x)
#  if defined(Q_CC_GNU) && !defined(Q_CC_INTEL) && Q_CC_GNU >= 600
     /* GCC requires attributes for a function */
#    define QT_FUNCTION_TARGET(x)  __attribute__((__target__(QT_FUNCTION_TARGET_STRING_ ## x)))
#  else
#    define QT_FUNCTION_TARGET(x)
#  endif
#  if !defined(__ARM_FEATURE_NEON) && defined(__ARM_NEON__)
#    define __ARM_FEATURE_NEON           // also support QT_COMPILER_SUPPORTS_HERE(NEON)
#  endif
#elif (defined(Q_CC_INTEL) || defined(Q_CC_MSVC) \
    || (defined(Q_CC_GNU) && !defined(Q_CC_CLANG) && (__GNUC__-0) * 100 + (__GNUC_MINOR__-0) >= 409)) \
    && !defined(QT_BOOTSTRAPPED)
#  define QT_COMPILER_SUPPORTS_SIMD_ALWAYS
#  define QT_COMPILER_SUPPORTS_HERE(x)    QT_COMPILER_SUPPORTS(x)
#  if defined(Q_CC_GNU) && !defined(Q_CC_INTEL)
     /* GCC requires attributes for a function */
#    define QT_FUNCTION_TARGET(x)  __attribute__((__target__(QT_FUNCTION_TARGET_STRING_ ## x)))
#  else
#    define QT_FUNCTION_TARGET(x)
#  endif
#else
#  define QT_COMPILER_SUPPORTS_HERE(x)    (__ ## x ## __)
#  define QT_FUNCTION_TARGET(x)
#endif

// SSE intrinsics
#define QT_FUNCTION_TARGET_STRING_SSE2      "sse2"
#if defined(__SSE2__) || (defined(QT_COMPILER_SUPPORTS_SSE2) && defined(QT_COMPILER_SUPPORTS_SIMD_ALWAYS))
#if defined(QT_LINUXBASE) || defined(Q_OS_ANDROID_NO_SDK)
/// this is an evil hack - the posix_memalign declaration in LSB
/// is wrong - see http://bugs.linuxbase.org/show_bug.cgi?id=2431
#  define posix_memalign _lsb_hack_posix_memalign
#  include <emmintrin.h>
#  undef posix_memalign
#else
#  include <emmintrin.h>
#endif
#if defined(Q_CC_MSVC) && (defined(_M_X64) || _M_IX86_FP >= 2)
#  define __SSE__ 1
#  define __SSE2__ 1
#endif
#endif

// SSE3 intrinsics
#define QT_FUNCTION_TARGET_STRING_SSE3      "sse3"
#if defined(__SSE3__) || (defined(QT_COMPILER_SUPPORTS_SSE3) && defined(QT_COMPILER_SUPPORTS_SIMD_ALWAYS))
#include <pmmintrin.h>
#endif

// SSSE3 intrinsics
#define QT_FUNCTION_TARGET_STRING_SSSE3     "ssse3"
#if defined(__SSSE3__) || (defined(QT_COMPILER_SUPPORTS_SSSE3) && defined(QT_COMPILER_SUPPORTS_SIMD_ALWAYS))
#include <tmmintrin.h>
#endif

// SSE4.1 intrinsics
#define QT_FUNCTION_TARGET_STRING_SSE4_1    "sse4.1"
#if defined(__SSE4_1__) || (defined(QT_COMPILER_SUPPORTS_SSE4_1) && defined(QT_COMPILER_SUPPORTS_SIMD_ALWAYS))
#include <smmintrin.h>
#endif

// SSE4.2 intrinsics
#define QT_FUNCTION_TARGET_STRING_SSE4_2    "sse4.2"
#if defined(__SSE4_2__) || (defined(QT_COMPILER_SUPPORTS_SSE4_2) && defined(QT_COMPILER_SUPPORTS_SIMD_ALWAYS))
#include <nmmintrin.h>
#endif

// AVX intrinsics
#define QT_FUNCTION_TARGET_STRING_AVX       "avx"
#define QT_FUNCTION_TARGET_STRING_AVX2      "avx2"
#if defined(__AVX__) || (defined(QT_COMPILER_SUPPORTS_AVX) && defined(QT_COMPILER_SUPPORTS_SIMD_ALWAYS))
// immintrin.h is the ultimate header, we don't need anything else after this
#include <immintrin.h>

#  if defined(Q_CC_MSVC) && (defined(_M_AVX) || defined(__AVX__))
// MS Visual Studio 2010 has no macro pre-defined to identify the use of /arch:AVX
// MS Visual Studio 2013 adds it: __AVX__
// See: http://connect.microsoft.com/VisualStudio/feedback/details/605858/arch-avx-should-define-a-predefined-macro-in-x64-and-set-a-unique-value-for-m-ix86-fp-in-win32
#    define __SSE3__ 1
#    define __SSSE3__ 1
// no Intel CPU supports SSE4a, so don't define it
#    define __SSE4_1__ 1
#    define __SSE4_2__ 1
#    ifndef __AVX__
#      define __AVX__ 1
#    endif
#  endif
#endif

#define QT_FUNCTION_TARGET_STRING_AVX512F       "avx512f"
#define QT_FUNCTION_TARGET_STRING_AVX512CD      "avx512cd"
#define QT_FUNCTION_TARGET_STRING_AVX512ER      "avx512er"
#define QT_FUNCTION_TARGET_STRING_AVX512PF      "avx512pf"
#define QT_FUNCTION_TARGET_STRING_AVX512BW      "avx512bw"
#define QT_FUNCTION_TARGET_STRING_AVX512DQ      "avx512dq"
#define QT_FUNCTION_TARGET_STRING_AVX512VL      "avx512vl"
#define QT_FUNCTION_TARGET_STRING_AVX512IFMA    "avx512ifma"
#define QT_FUNCTION_TARGET_STRING_AVX512VBMI    "avx512vbmi"

#define QT_FUNCTION_TARGET_STRING_F16C          "f16c"
#define QT_FUNCTION_TARGET_STRING_RDRAND        "rdrnd"
#define QT_FUNCTION_TARGET_STRING_BMI           "bmi"
#define QT_FUNCTION_TARGET_STRING_BMI2          "bmi2"
#define QT_FUNCTION_TARGET_STRING_RDSEED        "rdseed"
#define QT_FUNCTION_TARGET_STRING_SHA           "sha"

// other x86 intrinsics
#if defined(Q_PROCESSOR_X86) && ((defined(Q_CC_GNU) && (Q_CC_GNU >= 404)) \
    || (defined(Q_CC_CLANG) && (Q_CC_CLANG >= 208)) \
    || defined(Q_CC_INTEL))
#  define QT_COMPILER_SUPPORTS_X86INTRIN
#  ifdef Q_CC_INTEL
// The Intel compiler has no <x86intrin.h> -- all intrinsics are in <immintrin.h>;
#    include <immintrin.h>
#  else
// GCC 4.4 and Clang 2.8 added a few more intrinsics there
#    include <x86intrin.h>
#  endif
#endif

// Clang compiler fix, see http://lists.llvm.org/pipermail/cfe-commits/Week-of-Mon-20160222/151168.html
// This should be tweaked with an "upper version" of clang once we know which release fixes the
// issue. At that point we can rely on __ARM_FEATURE_CRC32 again.
#if defined(Q_CC_CLANG) && defined(Q_OS_DARWIN) && defined (__ARM_FEATURE_CRC32)
#  undef __ARM_FEATURE_CRC32
#endif

// NEON intrinsics
// note: as of GCC 4.9, does not support function targets for ARM
#if defined(__ARM_NEON) || defined(__ARM_NEON__)
#include <arm_neon.h>
#define QT_FUNCTION_TARGET_STRING_NEON      "+neon" // unused: gcc doesn't support function targets on non-aarch64, and on Aarch64 NEON is always available.
#ifndef __ARM_NEON__
// __ARM_NEON__ is not defined on AArch64, but we need it in our NEON detection.
#define __ARM_NEON__
#endif
#endif
// AArch64/ARM64
#if defined(Q_PROCESSOR_ARM_V8)
#define QT_FUNCTION_TARGET_STRING_CRC32      "+crc"
#  include <arm_acle.h>
#endif

#undef QT_COMPILER_SUPPORTS_SIMD_ALWAYS

QT_BEGIN_NAMESPACE


enum CPUFeatures {
#if defined(Q_PROCESSOR_ARM)
    CpuFeatureNEON          = 0,
    CpuFeatureARM_NEON      = CpuFeatureNEON,
    CpuFeatureCRC32         = 1,
#elif defined(Q_PROCESSOR_MIPS)
    CpuFeatureDSP           = 0,
    CpuFeatureDSPR2         = 1,
#elif defined(Q_PROCESSOR_X86)
    // The order of the flags is jumbled so it matches most closely the bits in CPUID
    // Out of order:
    CpuFeatureSSE2          = 1,                       // uses the bit for PCLMULQDQ
    // in level 1, ECX
    CpuFeatureSSE3          = (0 + 0),
    CpuFeatureSSSE3         = (0 + 9),
    CpuFeatureSSE4_1        = (0 + 19),
    CpuFeatureSSE4_2        = (0 + 20),
    CpuFeatureMOVBE         = (0 + 22),
    CpuFeaturePOPCNT        = (0 + 23),
    CpuFeatureAES           = (0 + 25),
    CpuFeatureAVX           = (0 + 28),
    CpuFeatureF16C          = (0 + 29),
    CpuFeatureRDRAND        = (0 + 30),
    // 31 is always zero and we've used it for the QSimdInitialized

    // in level 7, leaf 0, EBX
    CpuFeatureBMI           = (32 + 3),
    CpuFeatureHLE           = (32 + 4),
    CpuFeatureAVX2          = (32 + 5),
    CpuFeatureBMI2          = (32 + 8),
    CpuFeatureRTM           = (32 + 11),
    CpuFeatureAVX512F       = (32 + 16),
    CpuFeatureAVX512DQ      = (32 + 17),
    CpuFeatureRDSEED        = (32 + 18),
    CpuFeatureAVX512IFMA    = (32 + 21),
    CpuFeatureAVX512PF      = (32 + 26),
    CpuFeatureAVX512ER      = (32 + 27),
    CpuFeatureAVX512CD      = (32 + 28),
    CpuFeatureSHA           = (32 + 29),
    CpuFeatureAVX512BW      = (32 + 30),
    CpuFeatureAVX512VL      = (32 + 31),

    // in level 7, leaf 0, ECX (out of order, for now)
    CpuFeatureAVX512VBMI    = 2,                       // uses the bit for DTES64
#endif

    // used only to indicate that the CPU detection was initialised
    QSimdInitialized = 0x80000000
};

static const quint64 qCompilerCpuFeatures = 0
#if defined __SHA__
        | (Q_UINT64_C(1) << CpuFeatureSHA)
#endif
#if defined __AES__
        | (Q_UINT64_C(1) << CpuFeatureAES)
#endif
#if defined __RTM__
        | (Q_UINT64_C(1) << CpuFeatureRTM)
#endif
#ifdef __RDRND__
        | (Q_UINT64_C(1) << CpuFeatureRDRAND)
#endif
#ifdef __RDSEED__
        | (Q_UINT64_C(1) << CpuFeatureRDSEED)
#endif
#if defined __BMI__
        | (Q_UINT64_C(1) << CpuFeatureBMI)
#endif
#if defined __BMI2__
        | (Q_UINT64_C(1) << CpuFeatureBMI2)
#endif
#if defined __F16C__
        | (Q_UINT64_C(1) << CpuFeatureF16C)
#endif
#if defined __POPCNT__
        | (Q_UINT64_C(1) << CpuFeaturePOPCNT)
#endif
#if defined __MOVBE__           // GCC and Clang don't seem to define this
        | (Q_UINT64_C(1) << CpuFeatureMOVBE)
#endif
#if defined __AVX512F__
        | (Q_UINT64_C(1) << CpuFeatureAVX512F)
#endif
#if defined __AVX512CD__
        | (Q_UINT64_C(1) << CpuFeatureAVX512CD)
#endif
#if defined __AVX512ER__
        | (Q_UINT64_C(1) << CpuFeatureAVX512ER)
#endif
#if defined __AVX512PF__
        | (Q_UINT64_C(1) << CpuFeatureAVX512PF)
#endif
#if defined __AVX512BW__
        | (Q_UINT64_C(1) << CpuFeatureAVX512BW)
#endif
#if defined __AVX512DQ__
        | (Q_UINT64_C(1) << CpuFeatureAVX512DQ)
#endif
#if defined __AVX512VL__
        | (Q_UINT64_C(1) << CpuFeatureAVX512VL)
#endif
#if defined __AVX512IFMA__
        | (Q_UINT64_C(1) << CpuFeatureAVX512IFMA)
#endif
#if defined __AVX512VBMI__
        | (Q_UINT64_C(1) << CpuFeatureAVX512VBMI)
#endif
#if defined __AVX2__
        | (Q_UINT64_C(1) << CpuFeatureAVX2)
#endif
#if defined __AVX__
        | (Q_UINT64_C(1) << CpuFeatureAVX)
#endif
#if defined __SSE4_2__
        | (Q_UINT64_C(1) << CpuFeatureSSE4_2)
#endif
#if defined __SSE4_1__
        | (Q_UINT64_C(1) << CpuFeatureSSE4_1)
#endif
#if defined __SSSE3__
        | (Q_UINT64_C(1) << CpuFeatureSSSE3)
#endif
#if defined __SSE3__
        | (Q_UINT64_C(1) << CpuFeatureSSE3)
#endif
#if defined __SSE2__
        | (Q_UINT64_C(1) << CpuFeatureSSE2)
#endif
#if defined __ARM_NEON__
        | (Q_UINT64_C(1) << CpuFeatureNEON)
#endif
#if defined __ARM_FEATURE_CRC32
        | (Q_UINT64_C(1) << CpuFeatureCRC32)
#endif
#if defined __mips_dsp
        | (Q_UINT64_C(1) << CpuFeatureDSP)
#endif
#if defined __mips_dspr2
        | (Q_UINT64_C(1) << CpuFeatureDSPR2)
#endif
        ;

#ifdef Q_ATOMIC_INT64_IS_SUPPORTED
extern Q_CORE_EXPORT QBasicAtomicInteger<quint64> qt_cpu_features[1];
#else
extern Q_CORE_EXPORT QBasicAtomicInteger<unsigned> qt_cpu_features[2];
#endif
Q_CORE_EXPORT void qDetectCpuFeatures();

static inline quint64 qCpuFeatures()
{
    quint64 features = qt_cpu_features[0].load();
#ifndef Q_ATOMIC_INT64_IS_SUPPORTED
    features |= quint64(qt_cpu_features[1].load()) << 32;
#endif
    if (Q_UNLIKELY(features == 0)) {
        qDetectCpuFeatures();
        features = qt_cpu_features[0].load();
#ifndef Q_ATOMIC_INT64_IS_SUPPORTED
        features |= quint64(qt_cpu_features[1].load()) << 32;
#endif
        Q_ASSUME(features != 0);
    }
    return features;
}

#define qCpuHasFeature(feature)     ((qCompilerCpuFeatures & (Q_UINT64_C(1) << CpuFeature ## feature)) \
                                     || (qCpuFeatures() & (Q_UINT64_C(1) << CpuFeature ## feature)))

#if QT_HAS_BUILTIN(__builtin_clz) && QT_HAS_BUILTIN(__builtin_ctz) && defined(Q_CC_CLANG) && !defined(Q_CC_INTEL)
static Q_ALWAYS_INLINE unsigned _bit_scan_reverse(unsigned val)
{
    Q_ASSERT(val != 0); // if val==0, the result is undefined.
    unsigned result = static_cast<unsigned>(__builtin_clz(val)); // Count Leading Zeros
    // Now Invert the result: clz will count *down* from the msb to the lsb, so the msb index is 31
    // and the lsb inde is 0. The result for _bit_scan_reverse is expected to be the index when
    // counting up: msb index is 0 (because it starts there), and the lsb index is 31.
    result ^= sizeof(unsigned) * 8 - 1;
    return result;
}
static Q_ALWAYS_INLINE unsigned _bit_scan_forward(unsigned val)
{
    Q_ASSERT(val != 0); // if val==0, the result is undefined.
    return static_cast<unsigned>(__builtin_ctz(val)); // Count Trailing Zeros
}
#elif defined(Q_PROCESSOR_X86)
// Bit scan functions for x86
#  if defined(Q_CC_MSVC) && !defined(Q_OS_WINCE)
// MSVC calls it _BitScanReverse and returns the carry flag, which we don't need
static __forceinline unsigned long _bit_scan_reverse(uint val)
{
    unsigned long result;
    _BitScanReverse(&result, val);
    return result;
}
static __forceinline unsigned long _bit_scan_forward(uint val)
{
    unsigned long result;
    _BitScanForward(&result, val);
    return result;
}
#  elif (defined(Q_CC_CLANG) || (defined(Q_CC_GNU) && Q_CC_GNU < 405)) \
    && !defined(Q_CC_INTEL)
// Clang is missing the intrinsic for _bit_scan_reverse
// GCC only added it in version 4.5
static inline __attribute__((always_inline))
unsigned _bit_scan_reverse(unsigned val)
{
    unsigned result;
    asm("bsr %1, %0" : "=r" (result) : "r" (val));
    return result;
}
static inline __attribute__((always_inline))
unsigned _bit_scan_forward(unsigned val)
{
    unsigned result;
    asm("bsf %1, %0" : "=r" (result) : "r" (val));
    return result;
}
#  endif
#endif // Q_PROCESSOR_X86

#define ALIGNMENT_PROLOGUE_16BYTES(ptr, i, length) \
    for (; i < static_cast<int>(qMin(static_cast<quintptr>(length), ((4 - ((reinterpret_cast<quintptr>(ptr) >> 2) & 0x3)) & 0x3))); ++i)

template <typename T>
Q_ALWAYS_INLINE
T qUnalignedLoad(const void *ptr) Q_DECL_NOTHROW
{
    T result;
#if QT_HAS_BUILTIN(__builtin_memcpy)
    __builtin_memcpy
#else
    memcpy
#endif
    /*memcpy*/(&result, ptr, sizeof result);
    return result;
}

template <typename T>
Q_ALWAYS_INLINE
void qUnalignedStore(void *ptr, T t) Q_DECL_NOTHROW
{
#if QT_HAS_BUILTIN(__builtin_memcpy)
    __builtin_memcpy
#else
    memcpy
#endif
    /*memcpy*/(ptr, &t, sizeof t);
}

QT_END_NAMESPACE

#endif // QSIMD_P_H