1 | #ifndef SIMDJSON_ARM64_BITMANIPULATION_H |
2 | #define SIMDJSON_ARM64_BITMANIPULATION_H |
3 | |
4 | namespace simdjson { |
5 | namespace SIMDJSON_IMPLEMENTATION { |
6 | namespace { |
7 | |
8 | // We sometimes call trailing_zero on inputs that are zero, |
9 | // but the algorithms do not end up using the returned value. |
10 | // Sadly, sanitizers are not smart enough to figure it out. |
11 | SIMDJSON_NO_SANITIZE_UNDEFINED |
12 | simdjson_inline int trailing_zeroes(uint64_t input_num) { |
13 | #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO |
14 | unsigned long ret; |
15 | // Search the mask data from least significant bit (LSB) |
16 | // to the most significant bit (MSB) for a set bit (1). |
17 | _BitScanForward64(&ret, input_num); |
18 | return (int)ret; |
19 | #else // SIMDJSON_REGULAR_VISUAL_STUDIO |
20 | return __builtin_ctzll(input_num); |
21 | #endif // SIMDJSON_REGULAR_VISUAL_STUDIO |
22 | } |
23 | |
24 | /* result might be undefined when input_num is zero */ |
25 | simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) { |
26 | return input_num & (input_num-1); |
27 | } |
28 | |
29 | /* result might be undefined when input_num is zero */ |
30 | simdjson_inline int leading_zeroes(uint64_t input_num) { |
31 | #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO |
32 | unsigned long leading_zero = 0; |
33 | // Search the mask data from most significant bit (MSB) |
34 | // to least significant bit (LSB) for a set bit (1). |
35 | if (_BitScanReverse64(&leading_zero, input_num)) |
36 | return (int)(63 - leading_zero); |
37 | else |
38 | return 64; |
39 | #else |
40 | return __builtin_clzll(input_num); |
41 | #endif// SIMDJSON_REGULAR_VISUAL_STUDIO |
42 | } |
43 | |
44 | /* result might be undefined when input_num is zero */ |
45 | simdjson_inline int count_ones(uint64_t input_num) { |
46 | return vaddv_u8(p0: vcnt_u8(vcreate_u8(input_num))); |
47 | } |
48 | |
49 | |
50 | #if defined(__GNUC__) // catches clang and gcc |
51 | /** |
52 | * ARM has a fast 64-bit "bit reversal function" that is handy. However, |
53 | * it is not generally available as an intrinsic function under Visual |
54 | * Studio (though this might be changing). Even under clang/gcc, we |
55 | * apparently need to invoke inline assembly. |
56 | */ |
57 | /* |
58 | * We use SIMDJSON_PREFER_REVERSE_BITS as a hint that algorithms that |
59 | * work well with bit reversal may use it. |
60 | */ |
61 | #define SIMDJSON_PREFER_REVERSE_BITS 1 |
62 | |
63 | /* reverse the bits */ |
64 | simdjson_inline uint64_t reverse_bits(uint64_t input_num) { |
65 | uint64_t rev_bits; |
66 | __asm("rbit %0, %1" : "=r" (rev_bits) : "r" (input_num)); |
67 | return rev_bits; |
68 | } |
69 | |
70 | /** |
71 | * Flips bit at index 63 - lz. Thus if you have 'leading_zeroes' leading zeroes, |
72 | * then this will set to zero the leading bit. It is possible for leading_zeroes to be |
73 | * greating or equal to 63 in which case we trigger undefined behavior, but the output |
74 | * of such undefined behavior is never used. |
75 | **/ |
76 | SIMDJSON_NO_SANITIZE_UNDEFINED |
77 | simdjson_inline uint64_t zero_leading_bit(uint64_t rev_bits, int leading_zeroes) { |
78 | return rev_bits ^ (uint64_t(0x8000000000000000) >> leading_zeroes); |
79 | } |
80 | |
81 | #endif |
82 | |
83 | simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *result) { |
84 | #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO |
85 | *result = value1 + value2; |
86 | return *result < value1; |
87 | #else |
88 | return __builtin_uaddll_overflow(value1, value2, |
89 | reinterpret_cast<unsigned long long *>(result)); |
90 | #endif |
91 | } |
92 | |
93 | } // unnamed namespace |
94 | } // namespace SIMDJSON_IMPLEMENTATION |
95 | } // namespace simdjson |
96 | |
97 | #endif // SIMDJSON_ARM64_BITMANIPULATION_H |
98 | |